Web Information Systems and Technologies: Third International Conference, WEBIST 2007, Barcelona, Spain, March 3-6, 2007, Revised Selected Papers (Lecture Notes in Business Information Processing)
This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
Lecture Notes in Business Information Processing Series Editors Wil van der Aalst Eindhoven Technical University, The Netherlands John Mylopoulos University of Trento, Italy Norman M. Sadeh Carnegie Mellon University, Pittsburgh, PA, USA Michael J. Shaw University of Illinois, Urbana-Champaign, IL, USA Clemens Szyperski Microsoft Research, Redmond, WA, USA
8
Joaquim Filipe José Cordeiro (Eds.)
Web Information Systems and Technologies Third International Conference, WEBIST 2007 Barcelona, Spain, March 3-6, 2007 Revised Selected Papers
13
Volume Editors Joaquim Filipe José Cordeiro Polytechnic Institute of Setúbal – INSTICC 2910-761 Setúbal, Portugal E-mail: {j.filipe,jcordeiro}@est.ips.pt
Library of Congress Control Number: 2008930563 ACM Computing Classification (1998): H.3.5, H.4, J.1, K.4.4 ISSN ISBN-10 ISBN-13
1865-1348 3-540-68257-0 Springer Berlin Heidelberg New York 978-3-540-68257-8 Springer Berlin Heidelberg New York
This book contains the best papers from the International Conference on Web Information Systems and Technologies (WEBIST 2007), organized by the Institute for Systems and Technologies of Information, Control and Communication (INSTICC), endorsed by IW3C2, and held in Barcelona, Spain. The purpose of WEBIST is to bring together researchers, engineers, and practitioners interested in the technological advances and business applications of web-based information systems. It has four main topic areas, covering different aspects of web information systems, namely, internet technology; web interfaces and applications; society, e-business and e-government; and e-learning. WEBIST 2007 received 367 submissions from more than 50 countries across all continents. After a double-blind review process, with the help of more than 200 experts from the international program committee, and also after presentation at the conference, 23 papers were finally selected. Their extended and revised versions are published in this book. This strict selection made the conference appealing to a global audience of engineers, scientists, business practitioners, and policy experts. The papers accepted and presented at the conference demonstrated a number of new and innovative solutions for e-business and web information systems in general, showing that the technical problems in this field are challenging and worth further R&D effort. The program of this conference also included three outstanding keynote lectures presented by internationally renowned distinguished researchers. Their keynote speeches reinforced the overall quality of the event. The organization of these conferences required the dedicated effort of many people. Firstly, we must thank the authors, whose research and development efforts are recorded here. Secondly, we thank the members of the program committee and the additional reviewers for their diligence and expert reviewing. Thirdly, we congratulate the INSTICC organizing team for their excellent work. Last but not least, we thank the invited speakers for their invaluable contribution and for taking the time to synthesize and prepare their talks. We hope that you will find this collection of papers interesting and a helpful source of inspiration and that you will look forward to the book containing the best papers of next year’s conference. March 2008
Joaquim Filipe José Cordeiro
Organization
Conference Co-chairs Joaquim Filipe, Polytechnic Institute of Setúbal/INSTICC, Portugal Julià Minguillón, Open University of Catalonia, Spain
Program Co-chairs José Cordeiro, Polytechnic Institute of Setúbal/INSTICC, Portugal Slimane Hammoudi, E.S.E.O., France
Organizing Committee Paulo Brito, INSTICC, Portugal Marina Carvalho, INSTICC, Portugal Helder Coelhas, INSTICC, Portugal Bruno Encarnação, INSTICC, Portugal Vítor Pedrosa, INSTICC, Portugal Mónica Saramago, INSTICC, Portugal
Program Committee Jacky Akoka, France Abdullah Alghamdi, Saudi Arabia A.Y. Al-Zoubi, Jordan Rachid Anane, UK Margherita Antona, Greece Azita Bahrami, USA Matteo Baldoni, Italy Ken Barker, Canada Cristina Baroglio, Italy Sean Bechhofer, UK Andreas Becks, Germany David Bell, UK Orlando Belo, Portugal Bharat Bhargava, USA Paolo Bouquet, Italy Christos Bouras, Greece Amy Bruckman, USA Jack Brzezinski, USA Maria Claudia Buzzi, Italy
Elena Calude, New Zealand Diego Calvanese, Italy John Carroll, USA Nunzio Casalino, Italy Maiga Chang, Taiwan Weiqin Chen, Norway Adrian David Cheok, Singapore Evangelos Christou, Greece Jae Chung, USA Christophe Claramunt, France Isabelle Comyn-Wattiau, France Michel Crampes, France Alexandra Cristea, UK Daniel Cunliffe, UK John Philip Cuthell, UK Alfredo Cuzzocrea, Italy Georgios Dafoulas, UK Vasilios Darlagiannis, Switzerland Alessandro D'Atri, Italy
VIII
Organization
Valeria de Antonellis, Italy Sergio de Cesare, UK Steven Demurjian, USA Darina Dicheva, USA Y. Ding, Austria Asuman Dogac, Turkey Josep Domingo-Ferrer, Spain Chyi-Ren Dow, Taiwan Schahram Dustdar, Austria Ali El Kateeb, USA A. El Saddik, Canada Atilla Elci, Turkey Luis Ferreira Pires, The Netherlands Filomena Ferrucci, Italy Stefan Fischer, Germany Akira Fukuda, Japan Giovanni Fulantelli, Italy Erich Gams, Austria Daniel Garcia, Spain Franca Garzotto, Italy Dragan Gasevic, Canada José A. Gil Salinas, Spain Karl Goeschka, Austria Nicolás González-Deleito, Belgium Anna Grabowska, Poland Begoña Gros, Spain Vic Grout, UK Francesco Guerra, Italy Aaron Gulliver, Canada Pål Halvorsen, Norway Ioannis Hatzilygeroudis, Greece Stylianos Hatzipanagos, UK Dominik Heckmann, Germany Nicola Henze, Germany Dominic Heutelbeck, Germany David Hicks, Denmark Pascal Hitzler, Germany Kathleen Hornsby, USA Wen Shyong Hsieh, Taiwan Brian Hudson, Sweden Tanko Ishaya, UK Alexander Ivannikov, Russia Kai Jakobs, Germany Anne James, UK Narayana Jayaram, UK Ivan Jelinek, Czech Republic Jingwen Jin, USA
Qun Jin, Japan Carlos Juiz, Spain Michail Kalogiannakis, France Jussi Kangasharju, Germany Vassilis Kapsalis, Greece George Karabatis, USA Frank Kargl, Germany Roland Kaschek, New Zealand Sokratis Katsikas, Greece David Kennedy, China Ralf Klamma, Germany Mamadou Tadiou Kone, Canada Tsvi Kuflik, Israel Axel Küpper, Germany Daniel Lemire, Canada Tayeb Lemlouma, France Weigang Li, Brazil Leszek Lilien, USA Claudia Linnhoff-Popien, Germany Hui Liu, USA Pascal Lorenz, France Heiko Ludwig, USA Vicente Luque-Centeno, Spain Anna Maddalena, Italy George Magoulas, UK Johannes Mayer, Germany Carmel McNaught, China Ingo Melzer, Germany Elisabeth Metais, France Alessandro Micarelli, Italy Debajyoti Mukhopadhyay, India Kia Ng, UK David Nichols, New Zealand Andreas Ninck, Switzerland Dusica Novakovic, UK Vladimir Oleshchuk, Norway Andrea Omicini, Italy Kok-Leong Ong, Australia Jose Onieva, Spain Vasile Palade, UK Jun Pang, Germany Kalpdrum Passi, Canada Viviana Patti, Italy Chang-Shyh Peng, USA Guenther Pernul, Germany Rajesh Pillania, India Carsten Pils, Ireland
Organization
Volkmar Pipek, Germany Pierluigi Plebani, Italy Bhanu Prasad, USA Joshua Pun, Hong Kong Mimi Recker, USA Frank Reichert, Norway Werner Retschitzegger, Austria Norman Revell, UK Yacine Rezgui, UK Thomas Risse, Germany Marco Roccetti, Italy Dumitru Roman, Austria Danguole Rutkauskiene, Lithuania Wasim Sadiq, Australia Maytham Safar, Kuwait Eduardo Sanchez, Spain Abdolhossein Sarrafzadeh, New Zealand Anthony Savidis, Greece Vittorio Scarano, Italy Alexander Schatten, Austria Alexander Schill, Germany Heiko Schuldt, Switzerland Magy Seif El-Nasr, USA Jochen Seitz, Germany Tony Shan, USA Jamshid Shanbehzadeh, Iran Quan Z. Sheng, Australia Charles A. Shoniregun, UK Keng Siau, USA Miguel-Angel Sicilia, Spain Marianna Sigala, Greece Eva Söderström, Sweden
Miguel Soriano, Spain J. Michael Spector, USA Martin Sperka, Slovakia Frank Stowell, UK Carlo Strapparava, Italy Eleni Stroulia, Canada Hussein Suleman, South Africa Aixin Sun, Singapore Junichi Suzuki, USA Ramayah T., Malaysia Taro Tezuka, Japan Dirk Thissen, Germany Ivan Tomek, Canada Bettina Törpel, Denmark Arun-Kumar Tripathi, Germany Andre Trudel, Canada Thrasyvoulos Tsiatsos, Greece Klaus Turowski, Germany Michail Vaitis, Greece Christelle Vangenot, Switzerland Athanasios Vasilakos, Greece Jari Veijalainen, Finland Juan D. Velasquez, Chile Maria Esther Vidal, Venezuela Maurizio Vincini, Italy Viacheslav Wolfengagen, Russia Martin Wolpers, Belgium Huahui Wu, USA Lu Yan, UK Sung-Ming Yen, Taiwan Janette Young, UK Weihua Zhuang, Canada
Auxiliary Reviewers Isaac Agudo, Spain Gunes Aluc, Turkey Simona Barresi, UK Solomon Berhe, USA Simone Braun, Germany Tobias Buerger, Austria Patrícia Dockhorn Costa, The Netherlands Volker Derballa, Germany Wolfgang Dobmeier, Germany Thanh Tran Duc, Germany Stefan Dürbeck, Germany
Seth Freeman, USA Vittorio Fuccella, Italy Carmine Gravino, Italy Fu-Hau Hsu, Taiwan Larissa Ismailova, Russia Yildiray Kabak, Turkey Jan Kolter, Germany Jacek Kopecky, Austria Sergey Kosikov, Russia Damien Magoni, France Zoubir Mammeri, France
IX
X
Organization
Jaime Pavlich Mariscal, USA Sergio di Martino, Italy Jing Mei, China Michele Melchiori, Italy Alexandre Miège, Canada Alper Okcan, Turkey Wladimir Palant, Norway Andreas Petlund, Norway Vassilis Poulopoulos, Greece Sylvie Ranwez, France
Cyril Ray, France Christos Riziotis, Greece Claudio Schifanella, Italy Christian Schläger, Germany Paolo Spagnoletti, Italy Agusti Solanas, Spain Heiko Stoermer, Italy Fulya Tuncer, Turkey Knut-Helge Vik, Norway Christian Werner, Germany
Invited Speakers Azzelarabe Taleb-Bendiab, Liverpool John Moores University, UK Schahram Dustdar, Information Systems Institute, Vienna University of Technology, Austria Ernesto Damiani, University of Milan, Italy
Table of Contents
Invited Papers Programming Support and Governance for Process-Oriented Software Autonomy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A. Taleb-Bendiab, P. Miseldine, M. Randles, and Thar Baker Representing and Validating Digital Business Processes . . . . . . . . . . . . . . . Ernesto Damiani, Paolo Ceravolo, Cristiano Fugazza, and Karl Reed
3
19
Part I: Internet Technology Semi-automated Content Zoning of Spam Emails . . . . . . . . . . . . . . . . . . . . . Claudine Brucks, Michael Hilker, Christoph Schommer, Cynthia Wagner, and Ralph Weires
The Structure of Web-Based Information Systems Satisfaction: An Application of Confirmatory Factor Analysis . . . . . . . . . . . . . . . . . . . . . . . . Christy M.K. Cheung and Matthew K.O. Lee
257
Part IV: e-Learning Introducing Communities to e-Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Gavin McArdle, Teresa Monahan, and Michela Bertolotto Evaluating an Online Module on Copyright Law and Intellectual Property . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Carmel McNaught, Paul Lam, Shirley Leung, and Kin-Fai Cheng Using Games-Based Learning to Teach Software Engineering . . . . . . . . . . Thomas M. Connolly, Mark Stansfield, and Tom Hainey
277
292 304
Table of Contents
XIII
Agility in Serious Games Development with Distributed Teams: A Case Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Manuel Oliveira and Heiko Duin
314
Improving the Flexibility of Learning Environments: Developing Applications for Wired and Wireless Use . . . . . . . . . . . . . . . . . . . . . . . . . . . . David M. Kennedy and Doug Vogel
327
Evaluating an e-Learning Experience Based on the Sakai Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . F´elix Buend´ıa Garc´ıa and Antonio Herv´ as Jorge
Programming Support and Governance for Process-Oriented Software Autonomy A. Taleb-Bendiab, P. Miseldine, M. Randles, and Thar Baker School of Computing and Mathematical Sciences Liverpool John Moores University, Byrom Street, Liverpool, L3 3AF, UK [email protected], [email protected][email protected], [email protected]
Abstract. Business Process models seek to orchestrate business functions through the development of automated task completion, which is becoming increasingly used for Service-Oriented Architectures. This had led to many advances in the methods and tools available for software and language support in process modelling and enactment. Recent development in Business Process Execution languages, such as WS-BPEL 2.0 has widened the scope of process modelling to encompass cross-enterprise and inter-enterprise processes with a wide spread of often heterogonous business processes together with a range of associated modules for enactment, governance and assurance, to name but a few, to address non-functional requirements. Hence, the task of provisioning and managing such systems far outstrips the capabilities of human operatives, with most adaptations to operational circumstances requiring the system to be taken offline reprogrammed, recompiled and redeployed. This work focuses on the application of recent developments in language support for software autonomy whilst guaranteeing autonomic software behaviour. The issues to be addressed are stated with a supporting framework and language, Neptune. This is illustrated through a representative example with a case study evaluation reported upon. Keywords: Self-governance, programming language, process models.
software. Currently there is much interest in web based service orchestration to describe business processes through SOA and the emerging Service Component Architecture model, for example. The primary language for Business process support in this field is the Business Process Execution Language (BPEL), most usually represented by BPEL4WS and its successor WS-BPEL 2.0 [1]; though, other alternatives exist to support process–oriented software for implementing business processes on top of web services. In view of the complexity engendered by the heterogeneity in processes and their functional and non-functional requirements to support enterprise agility and business ecosystems, this work focuses on the support for runtime self-management and its predictable governance [2]. The aim is to contribute to new models of autonomic process-oriented Service Component Architecture (SCA). This approach is seen as vital to support current service intensive systems providing the core infrastructure for commerce and government [3], with its need for constant availability, to be always on and its continuous evolution based on current knowledge and process. Such requirements are evident in a healthcare decision process model which is oftenchanging to reflect new understandings, and the software used to aid the decision processes workflow is required to remain up to date to produce best-practice care [4]. Delays in updating software mean patients are diagnosed using out-of-date medical knowledge and guidelines. Coupled with distribution, where different parts of the application may be executed and stored on different machines, the complexity of deployment and its operating environment drives the need for computational techniques to aid in the analysis and actuation of changes in requirement whilst the software remains operational. Research into autonomic software has yielded many potential solutions to adaptation with respect to complexity [5], often from the viewpoint of providing self-aware behaviour to predict and react to changes in requirements for governance [6]. As identified in [7] however, a systematic approach to the adaptation of software at runtime to new or emergent requirements is still an outstanding issue for research. In addition, providing adaptation of a business process based on assured, bounded governance requires knowledge of not only what should be adapted, but also its impact on other processes and the safety of producing such adaptation within its defined boundaries. In effect, knowledge of the construction and the meanings behind the formation of a business process system is vital for bounded autonomy through adaptation to be accomplished [8]. Issues arise when this knowledge for autonomy is to be inferred at runtime, and adaptation is required to be instigated. Process implementations are most often deployed in a compiled denotation unsuited for reasoning upon to perform coordinated adaptation. Similarly, the compilation process and language used removes much of the semantic direction of the code; code is commonly written as instructions to achieve a goal, without implicitly specifying why those specific instructions were chosen for the task. In modern, distributed systems, it is often the case that source code is not available for adaptation, with the use of software as a service model hiding the implementation from consumers. As a result, much research focuses on behaviour driven adaptation and tuning of an application, rather than focusing on the construction and adaptation of its deployed code base.
Programming Support and Governance for Process-Oriented Software Autonomy
5
It is the contention of this paper that the functionality and knowledge required for autonomic adaptation of process models can be achieved via semantic analysis of the core components and code base of the software itself. This analysis can in turn inform traditional techniques of critique used in Autonomic Software, such as emergence identification and signal-grounding [9]. The paper highlights the limitations in current research into computational runtime adaptation and moves to identify properties of a new language Neptune that attempts to provide the requisite features to meet the goal of runtime computational adaptation of a system in general and process models in particular. Several proof of concept scenarios are given to further illustrate the approach advocated by Neptune and to justify a semantics-based approach. The paper concludes with discussion of limitations introduced by the approach, with work relating to possible future solutions given.
2 Business Process Runtime Requirements for Self-governance The inherent complexity of Business Process Models enactment, where process interactions and cross enterprise applications are involved, contributes towards the increasing interest in emerging autonomic SOA provision. Where autonomic software Table 1. Language Support Issues
Issues 1. Deployment Platform Reliance
2. Development Platform Reliance
3. Expressiveness
4. Application Representation 5. Component Effects 6. Data Distribution
Details There must be no reliance on an extension to, or specialism of, a specific language, operating system, or other system component to function. In effect, it must be deploymentplatform agnostic. The method of code adaptation for autonomic management must be accessible such that it has the ability to be dispatched by, and respond via, standard methods of invocation and communication. In effect, it must be developmentplatform agnostic. The semantic labelling of the processes ultimate purpose must be expressive enough to represent the semantics of base language components – which themselves may be implemented in any programming language that is supported in issue 2 – such that the effect, requirement, and mechanism of actuation of the component can be ascertained at runtime for use by the autonomic manager. A service application must be expressed in terms of the processes it is to represent using an activity-based classification, using the same semantic forms defined in Issue 3. The intended purpose of a component as provided in Issue 3 must be defined by its side effects to the state of the process it is used within. There must be support for runtime distribution of all information expressed within the process model via computation methods and communication standards.
6
A. Taleb-Bendiab et al.
capabilities attempt to predict and react to changes in software requirements for selfpreservation. Using the naturally derived autonomous behaviour exhibited by living organisms, research into autonomic computing is often linked with runtime adaptation and analysis to provide solutions to imbue self-aware behaviour to systems, machines and processes [10]. A wide variety of solutions exist, often targeting different aspects of A wide variety of solutions exist, often targeting different aspects of behaviour in terms of implementation and deployment. As stated in the introduction, Autonomic Computing has traditionally viewed systems from a behavioural perspective via monitoring. There are techniques however, that use code base analysis. Toskana [11] and similar approaches use AOP as a basis for adaptable change. Further holistic paradigms for autonomic software have been made using agent theory [12] treating a complete system as a series of agents, each with their own behaviour models to collectively perform set tasks and behaviour. Such an approach can also be applied to view business processes as sets of agent tasks and roles. Other problems arise from the lack of availability of source code. Often in systems such as service-oriented architectures, only consumption of behaviour is available with the actual implementation and its source code hidden. In these cases, adaptive middleware components that marshal and re-route calls are popular solutions [13]; these, however, often work at an abstraction layer above the code base, via policy documentation, forming of metamodel of behaviour reliant on observation. In line with the highlighted concerns, associated with providing autonomic management of implemented business process applications, there are a number of issues that need to be addressed; these have been collated in Table 1. A more detailed description may be found in the associated technical report [14].
3 Implementation of Autonomic Management As a conceptual basis, the qualities specified previously have been formed to provide the foundation of a scripting language, Neptune, to solve many of the outstanding issues highlighted in current research into runtime analysis and adaptation. Neptune can enable full runtime adaptation for autonomic self-management by addressing the issues previously identified. Issues 1 and 2 both indicate that the solution should sit above and be separate from any single language used to produce language components. In effect, the components are hidden from Neptune by way of a semantic interface: Neptune need only be able to dispatch calls defined within the semantic interface, rather than needing to directly host and manage the component itself. In this way, a bridge between Neptune and a base language is required both to facilitate communication between the base language and Neptune so that the language can instruct Neptune and interpret results from it, but also so Neptune can dispatch a call that is required to be handled by the base language. It is also clear that Neptune requires a form of expression to be defined such that it is capable of representing the ontological model desired to fulfil the remaining qualities. In keeping with established, classical works [15], Neptune will present information using a context-free language NeptuneScript.
Programming Support and Governance for Process-Oriented Software Autonomy
7
Figure 1 shows a schematic view of the method in which Neptune communicates with base language components. The Base Language Connector’s are themselves written within a base language rather than with NeptuneScript, and encapsulate the structure and communication methods needed to interact fully with Neptune and to be able to interpret and write NeptuneScript. Neptune itself provides numerous interfaces using standardised protocols such as COM, SOAP, and RPC forms such as Java RPC or .NET Remoting to enable a base language connector to dispatch a call to Neptune, and receive responses from it, as governed by Issue 2.
Fig. 1. Neptune/Base Language Interface
These calls are marshalled either directly to the higher Neptune Framework for processing, or situated as Neptune Base Language Objects (NBLOs), a semantic form that describes the intentions and conditions needed by a base language component. In addition, the NBLOs contain the requisite information to determine how to interoperate with the appropriate base language (BL) connector to marshal a call to the base language component. This information can be used by the Neptune Framework, to invoke the behaviour described in the NBLO. A semantic linker module within the framework encapsulates the functionality required to match and relate the semantics defined within process models, to those contained within the NBLOs to provide solutions for activities or tasks given in the process. Once a match is found, the calls to the base language contained within the NBLO are actuated. In effect, the axiomatic and operational semantics of a task are related to the denotational semantics given as an actuation within an NBLO’s definition. As discussed in Issue 3, the process of semantic manipulation and inference are well understood in literature and similar techniques are used to inform the linking module. This process is illustrated in Figure 2.
8
A. Taleb-Bendiab et al.
Fig. 2. The Neptune Framework
3.1 A Simple Example In order to illustrate the design basis of Neptune, and show the form NeptuneScripts take, the following example is used: A simple task is portrayed as a process that formats a given string as a title, using different styles that are implemented in distinct languages; a method bold in C# and a method italic in Java as shown below (Figures 3 and 4).
Fig. 3. Code for format.cs
Fig. 4. Code for format.java
Both are described using operational and axiomatic semantics within an NBLO definition to show that the purpose is to provide a feature to a type defined in Neptune, NString to transcend the types used in the base languages. In effect, the NBLO is defined by its effect on the state of any application it is part of – in keeping with
Programming Support and Governance for Process-Oriented Software Autonomy
9
Issue 5 – such that the type NString represents the state of a string through the process being executed. This is shown for the C# bold method in Figure 5. Tasks that require particular features of the type, requirements that are axiomatically defined, can thus be related to the relevant NBLO that can operationally ensure conformance to the requirement.
Fig. 5. NBLO Definitions
Evaluations, if possible, of the operational semantics defined can be included to determine if the goal of the identified process has been successfully actuated. Details of the base language component are provided in the form of an actuation and actual call needed by the base language connector, to provide the denotational semantics that the semantic linker module should produce. Each base language connector defines the meaning of a set of base types provided by Neptune which are related to equivalent types in the base language, such that NString represents a string as understood by Neptune, which is related to System.String in C# and the String type in Java via their respective base language connectors, to maintain interoperability. With this behavioural basis, a task can be written that requires a string to have two features, bold and italic. These are named a and b respectively, so that the requirements can be referenced within the body of the task for adaptation purposes. Neptune can rewrite the task to provide an action element that can satisfy the requirements specified in the task by inferring suitable denotational semantics by linking them to the purposes defined in the NBLOs such that the requirements hold. In this regard, Neptune relates the semantic description given in a task to the denotational semantics provided in the NBLOs given as actual source code invocations, without human intervention. This provides the maximum level of semantic-based assurance to the solution of the process task: each operation to take place via the base language components can be traced and asserted to individual requirements. As the requirements expressed in the task are met, Neptune can then marshal the appropriate calls to the base language connectors to instigate the actual functionality. To highlight the adaptive qualities of Neptune, the task can be rewritten at runtime to only require the code shown in Figure 6:
10
A. Taleb-Bendiab et al.
Fig. 6. Adapted Task Definition
As the actions produced before no longer satisfy the single requirement, such that nblo.Java.Italic provides a feature not required by the task, this can be treated as a trigger for adaptation. This is important as the requirements defined in the task can thus act as a governance mechanism ensuring the behaviour required is actually actuated. In this way, a developer can specify actions, which are asserted against the requirements and refined accordingly. The process of semantic linking can repeat given this trigger, to refine the action element of the task. Furthermore, a requirement can be specified to only be applied if the NBLO evaluates appropriately using the logic outlined in the evaluate element of the NBLO definition. As such, the task can be written a shown in Figure 7.
Fig. 7. Adapted Task Definition with New Inferred Action
In this way, the feature Bold is defined to be evaluated to true if a string starts with “” and ends with “”. If the string does not contain this feature – it equates to false – then the requirement is actuated. This control structure is repeated within the action element of the task to optimise the process; without the control structure, the action element would require adaptation upon each execution of the task with a given string. To show how a Neptune process can be adapted to reflect new behaviour, a new assembly can be added to the runtime environment of the application, and described via a new NBLO that is registered with the controller at runtime with a similar underline class in C#. The task FormatStringAsTitle can be rewritten at runtime due to its exposition, to include the requirement of the functionality provided for by the new behaviour which is semantically linked and adapted by the respective modules. Thus, the semantic definition of NString.Underline implies that of NString.Bold to be present, allowing the task to be adapted as shown in Figure 8.
Programming Support and Governance for Process-Oriented Software Autonomy
11
Fig. 8. Inferred Bolded Behaviour
4 A Case Study Neptune has been used to develop a number of autonomic process models to test and evaluate its design, performance and further analyse its approach. In particular, Neptune has been used to implement an autonomic service-oriented decision-support process, which enables via situation and role-awareness, the reconfiguration of medical decision agents to adapt to the individual preferences of expert clinicians and hospitals local clinical governance rules. Below is a survey of Neptune’s suitability to produce self-adaptive autonomic business process systems bounded within organisational guidelines via examples of architectural and behaviour adaptation required to be implemented at runtime. 4.1 Architectural Autonomy In partnership with several NHS dental trusts, the triage process of dental access services were assessed and implemented in Neptune, using an activity-based BPM. The model defined individual tasks that represented data input, such that the process required knowledge of a patient. The data could feasibly be read from a database, if it was available, or from the operator directly asking the patient. As such, two NBLOs were defined to describe these methods of data input for the patient’s data, given their NHS ID as shown in Figure 9. As such, both NBLOs have the same purpose, however define different base language methods of obtaining the data. In addition, the getPatientFromOperator purpose specifies that it should only be enacted if the data was not available in the database. A task was setup to obtain this data, however, how the data was to be found was left to Neptune to deliberate upon. In this way, the actual behaviour within the system of locating a patient was left to autonomous control, it is not specified in any source code other than via intentions and semantics of its desired effect. During the lifetime of the application, a new Electronic Patient Register was produced. The existing mechanism for retrieving data within the process now had to not use the local database, but rather the new EPR via a web service interface, written in Java. Accordingly, the existing process model still held, however any reference to nblo.getPatientFromDatabase was required to be adapted to point towards a new
12
A. Taleb-Bendiab et al.
Fig. 9. NBLO Definition for Patient Lookup
NBLO getPatientFromEPR, which semantically defined the new mechanism for retrieving data. For brevity, its source is not fully included, however due to its web service basis, it set a requirement that a SOAP endpoint was made available to it before it could execute, giving a different semantic to that of the database. The refactor types are pre-defined blocks of utility, executed by Neptune via its sources. In this way, when called, the NBLO will rewrite any reference to the existing NBLO in the tasks and processes to the new method required. The original NBLO is then removed from Neptune, giving a redirection of behaviour at runtime. These mechanisms are the same as employed by IDE’s such as Eclipse and Visual Studio, and as such follow classical research into adaptation. In Neptune however, such adaptation is machine-led and performed at runtime to the existing source base without a recompilation, as the code is simply re-interpreted rather than being at a pre-compile stage. However, the interpreter handles refactoring tasks such that when called, it modifies the source code directly to instigate the change. Thus, rather than applying the policy in a meta-form, either via additional policy or monitoring, the changes are applied at the source code level, leading to cleaner and more manageable code base. To conclude, the actions are re-interpreted after the refactor commands were interpreted. As the getPatientFromEPR NBLO required a SOAP service endpoint to be discovered, the actions, after re- interpretation, contain the new requirement set forth by the NBLO, to yield line 7 in Figure 10. Thus, any point in which the process required a feature from the database was automatically adapted at runtime to provide the same feature from a different approach that itself required modifications to the source based on its own requirements, to maintain integrity. As all inferences are semantic based, changes to the requirements and properties of the new system could easily be instigated, and via the checks made by Neptune, full adaptation within the
Programming Support and Governance for Process-Oriented Software Autonomy
13
task could be made. In addition Neptune can be used to describe components implemented in different languages; these could be combined to produce a single solution to a set process. This was achieved without requiring the execution sequence of components to be specified. It was, instead, inferred autonomically from the boundaries described semantically via BPMs.
Fig. 10. Task with New Re-factored Actions
4.2 Behavioural Autonomy Using Neptune, the process that occurred after pain levels were ascertained could be modified directly, producing a unified decision process model for each clinician. Axioms set forth in the base decision model were set, allowing behaviour, such as instructing a patient to go to hospital, to be ratified against institutional procedures. In this way, as long as these were met, the requirements of the model could be instigated. Thus, the task, afterPain could be rewritten at runtime to invoke the method p.ToHospital (say). Other tasks and NBLOs could be written to extend the definition of p.ToHospital such that this can only occur if there is space in the hospital, or waiting times are lower than a specified value. As each aspect of information is defined independently, the cause and effect of such policy does not need to be determined: Neptune relates the semantics to an end-goal, rather than requiring human intervention to achieve runtime autonomic refactoring of behaviour. Where a behaviour is deemed unsuitable, a boundary is formed, such that if no hospitals were available to send a patient, the preference could not be safely realised, and as such, control defaults back to the original model which is rigorously defined. In this way concerns can be autonomously introduced to the system with assured behaviour through maintaining the purposes and requirements of the system.
5 Evaluation Given the increasing requirements for future distributed computing systems to be endowed with runtime adaptation capabilities, there are many applications for the Neptune framework and language, in that, autonomic middleware will incorporate a logical reasoning mechanism to deduce responses, throughout runtime, without halting the system. Presently the rules for autonomic response are often portrayed to the
14
A. Taleb-Bendiab et al.
system as a static rule set consisting of norms or modalities, such as permissible or obligatory. In many distributed settings, as the system evolves and moves away from its starting configuration, these rules will quickly become outdated. It is at this point that the logical mechanisms within Neptune become paramount to dynamically evolve the rule set and code base in line with the systems evolution. There have, therefore been many case study evaluations for this described approach in autonomic middleware applications. The following is a further evaluation of Neptune performance against the Sun Microsystem’s Pet Store and Microsoft’s PetShop benchmark for enterprise applications, which is outlined below. Though, the full description of this evaluation is outside the scope of this paper and can be found in [16]. 5.1 Adaptation Scenario Pet Store [17] is an architectural blueprint from Sun Microsystems highlighting many of the advanced capabilities available with Java as well as detailing the design guidelines for the language. The application produced by the Pet Store blueprint builds an imaginary web-based solution for browsing and purchasing animals within a storefront interface. There are facilitates within the application architecture for administrators of the system to add, remove, and modify the animals available for sale, with their data stored on a database. Thus the Pet Store is publicly available, best-practice system design architecture for e-commerce and enterprise systems in general that Neptune may be evaluated against. Furthermore shortly after the release of Pet Store, Microsoft implemented its own interpretation of the system, PetShop [18], such that a comparison could be made against the effectiveness of a Java implemented enterprise system, and a Microsoft .NET implementation of the same behaviour. The design of the Pet Store model is one that advocates the use of 3-tier architecture. In this approach, a logic layer, provided by language support, sits between a data layer, consisting of database functions and a presentation layer displaying dynamic web pages to the user. To assess the benefits of Neptune it is the logic layer that will be primarily of interest here to evaluate the impact on the system of the adaptation of process. In the PetShop model, the use of thin classes, that represent directly the data types used within the software, act as isolators of dependency between layers. The logic layer consists of classes that each encapsulate an entity used within the shop model; a customer class represents all the features and data model required from a customer, whereas the order class represents the processes that pertain to order transactions within the system. Methods are used to encapsulate processes, such that Product.getProductFromID encapsulates the process needed to obtain a product from the database given an ID. As such, to adapt a process, its associated class and method contained within the logic layer needs to be determined. 5.2 Introducing Neptune to the PetShop/Pet Store Other implementations of the original Pet Store model exist; this model is particularly of interest, as Neptune is required to be introduced to an existing model. In this way, Neptune can be considered as a concern, and thus crosscut against the existing classes representing the processes requiring adaptation. The logic layer in PetShop performs
Programming Support and Governance for Process-Oriented Software Autonomy
15
several key processes: The process of ordering a product, where customer product functions are executed to produce an order, is an example of such a process. It can be seen that adapting the behaviour of the order process, such that a modified process is generated in reference to a new behaviour using a Process-Oriented Architecture (POA), may highlight many of the design qualities given in Neptune, and contrast this against current best-practice advocated by the original designs. In the comparison between PetShop and Pet Store, several criteria were considered to compare the designs. In effect, the benefits of the PetShop were determined by the number of lines of code produced with performance and scalability judged by response time per user, and CPU resources required by both approaches. The scenario outlines that a process should be chosen from the PetShop example to be modelled within Neptune, and analysis made using the criteria set by Microsoft and Sun upon the relative impact of introducing Neptune to an existing system. The model should then be adapted to introduce a new behaviour, with a comparison made between the method required in Neptune to achieve this adaptation, and the method used in the traditional system design, with benefits and limitations discussed accordingly. The process of performing an order within the PetShop is split between the logic layer and the presentation layer. Data is retrieved to inform the process via discrete web pages, such that OrderShipping.aspx, for example, contains the trigger code that moves the process towards another task, when the user has completed the page and clicked a relevant button. The execution of the process then passes to the logic layer to interpret the data, and store this within the database. Once stored, the logic moves the execution to produce another page for the user. As such, the classes within the logic layer can be said to change the state of the process by signifying that a task is complete, and that data is added to the data layer structures, as these are the side effects of its actuation. Similarly, pages within the presentation layer can be said to introduce data to the state of process, and signify that the task of producing data is complete. In this way, exposing the logic layer and presentation layer classes with NBLOs to Neptune will allow the operational semantics of the classes to be imparted. 5.3 Results Using the criteria set forth in the original release of PetShop, the following results were obtained for the execution of a complete order process. These results will, therefore, show the impact of using Neptune in terms of complexity and performance. For response times, Microsoft ACT was used to provide dummy information to the pages so that the process could complete via machine instruction. Averages of 10,000 process executions were made with 100 concurrent visitors. Whilst the original PetShop did not include any need for reflection, it can be considered vital, due to the relationship between runtime analysis and adaptation, that for a system to be available for adaptation, reflection is required. As such, the same test was performed using a modified PetShop implementation that called its methods via reflection. In this way, a benchmark for any form of adaptation could be ascertained.
16
A. Taleb-Bendiab et al.
Fig. 11. Performance Relationship
The results, shown in Figure 11 indicate that the overhead of using Neptune is significant upon the first execution of the process. This is due to the semantic linker having to process the available NBLO descriptions and refactor the description to perform the task. Upon further execution however, as the tasks have already been provided the NBLOs that satisfy their requirements, execution time increases dramatically, introducing a slight overhead in terms of computational performance over the adaptation benchmark.
Fig. 12. Neptune Scalability Testing
Programming Support and Governance for Process-Oriented Software Autonomy
17
In terms of CPU usage the processes of Neptune produced more in the way of computational complexity than the traditionally compiled application. This is expected due to the interpretation that occurs at runtime via use of Neptune. The result of using reflection produced a slight increase in CPU consumption. However, the compiled Neptune system – produced after the first semantic linking – performed relatively worse. It should be noted that the implementation of the Neptune framework is such that it was designed as a prototype, rather than being optimised for performance. As such, it could be expected theoretically, that any overhead can be reduced further with subsequent revisions of the implementation of the framework. The scalability of Neptune in terms of performance (blue line) is measured against a baseline execution rate (green line) in Figure 12. This shows that at present the initial overhead of Neptune performance is quite high. This, however, may be offset by benefits of continuous availability later in the systems evolution.
6 Conclusions This work has proposed applying autonomic runtime support for Business Process Models to provide methods to address self-governance in service-oriented architectures. Specifically this has involved adaptation of the actual code base of the system by employing a scripting language and framework, Neptune, to reason over the perceived goals of the deployed software. This means that interacting or complex business processes specified in any manner can benefit from runtime adaptation. Further work is required around the issues of type safety and scalability in the semantic linking process; as the volume of NBLOs and semantic concerns increases there will be a corresponding rise in the complexity associated with maintaining the optimal behaviour required. In this work the method is shown to have practical and theoretical value in providing autonomous management in SOAs. Further work is required to make this approach more scalable. The present position is that a high initial computational overhead is encountered. Optimisation of the approach will aid scalability; offsetting the initial costs against future saving in system downtime due to the added robustness gained from the runtime adaptation facilities.
References 1. Arkin, A., Askary, S., Bloch, B., Curbera, F., Goland, Y., Kartha, N., Liu, C.K., Thatte, S., Yendluri, P., Yiu, A. (eds.): Web Services Business Process Execution Language Version 2.0. WS-BPEL TC OASIS (December 2005), http://docs.oasis-open.org/wsbpel/ 2.0/wsbpel-v2.0.pdf 2. Luciano, B., Di Nitto, E., Ghezzi, C.: Toward Open-World Software: Issue and Challenges. Computer 39(10), 36–43 (2006) 3. Adeshara, P., Juric, R., Kuljis, J., Paul, R.: A Survey of Acceptance of E-government Services in the UK. In: Proceedings of the 26th International Conference on Information Technology Interfaces, pp. 415–420 (2004)
18
A. Taleb-Bendiab et al.
4. Rousseau, N., McColl, E., Newton, J., Grimshaw, J., Eccles, M.: Practice based, longitudinal, qualitative interview study of computerised evidence based guidelines in primary care. British Medical Journal 326(7384), 1–8 (2003) 5. Dharini, B., Morrison, R., Kirby, G., Mickan, K., Warboys, B., Robertson, I., Snowdon, B., Greenwood, R.M., Seet, W.: A software architecture approach for structuring autonomic systems. In: DEAS 2005: Proceedings of the 2005 workshop on Design and evolution of autonomic application software, pp. 1–7. ACM Press, New York (2005) 6. Kephart Jeffrey, O., Chess, D.M.: The vision of autonomic computing. Computer 36(1), 41–50 (2003) 7. Mazeiar, S., Tahvildari, L.: Autonomic computing: emerging trends and open problems. In: DEAS 2005: Proceedings of the 2005 workshop on Design and evolution of autonomic application software, pp. 1–7. ACM Press, New York (2005) 8. Sharon, A., Estrin, D.: On supporting autonomy and interdependence in distributed systems. In: Proceedings of the 3rd workshop on ACM SIGOPS European workshop, pp. 1–4. ACM Press, New York (1988) 9. Martin, R., Taleb-Bendiab, A., Miseldine, P.: Addressing the signal grounding problem for autonomic systems. In: The Proceedings of the International Conference on Autonomic and Autonomous Systems, 2006 (ICAS 2006), July 2006, pp. 21–27 (2006) 10. Manish, P., Hariri, S.: Autonomic computing: An overview. Springer, Heidelberg (2005) 11. Michael, E., Freisleben, B.: Supporting autonomic computing functionality via dynamic operating system kernel aspects. In: AOSD 2005: Proceedings of the 4th international conference on Aspect-oriented software development, pp. 51–62. ACM Press, New York (2005) 12. Cristiano, C.: Guarantees for autonomy in cognitive agent architecture. In: ECAI 1994: Proceedings of the workshop on agent theories, architectures, and languages on Intelligent agents, pp. 56–70. Springer, New York (1995) 13. Yves, C.: Self-adaptive middleware: Supporting business process priorities and service level agreements. Advanced Engineering Informatics 19(3), 199–211 (2005) 14. Philip, M., Taleb-Bendiab, A.: Neptune: Supporting Semantics-Based Runtime Software Refactoring to Achieve Assured System Autonomy, Technical Report, Liverpool John Moores University http://neptune.cms.livjm.ac.uk/dasel 15. Chomsky, N.: Three models for the description of language. Information Theory, IEEE Transactions 2(3), 113–124 (1956) 16. Miseldine, P.: Language Support for Process-Oriented Programming of Autonomic Software Systems. PhD. Thesis, Liverpool John Moores University (2007) 17. Sun Microsystems, Java Pet Store demo. Java Blueprint (Accessed November 2007), https://blueprints.dev.java.net/petstore/ 18. Microsoft, using .NET to Implement Sun Microsystems’ Java Pet Store J2EE Blueprint Application (Accessed November 2007), http://msdn2.microsoft.com/en-us/ library/ms954626.aspx
Representing and Validating Digital Business Processes Ernesto Damiani1 , Paolo Ceravolo1 , Cristiano Fugazza1 , and Karl Reed2 1
Department of Information Technology, University of Milan via Bramante, 65 - 26013, Crema (CR), Italy {damiani,ceravolo,fugazza}@dti.unimi.it 2 Department of Computer Science, LaTrobe University Bundoora, Melbourne, Australia [email protected]
Abstract. Business Process Modeling is increasingly important for the digitalization of both IT and non-IT business processes, as well as for their deployment on service-oriented architectures. A number of methodologies, languages, and software tools have been proposed to support digital business process design; nonetheless, a lot remains to be done for assessing a business process model validity with respect to an existing organizational structure or external constraints like the ones imposed by security compliance regulations. In particular, web-based business coalitions and other inter-organizational transactions pose a number of research problems. The Model Driven Architecture (MDA) provides a framework for representing processes at different levels of abstraction. In this paper, a MDA-driven notion of business process model is introduced, composed of a static domain model including the domain entities and actors, plus a platformindependent workflow model providing a specification of process activities. The paper focuses on semantics-aware representation techniques, introducing logicsbased static domain models and their relationship with Description Logics and current Semantic Web metadata formats. Then, the paper discusses some issues emerging from the literature on business processes representation and presents some research directions on the evaluation of the compatibility of business and IT processes with existing organizational environments and practices. The problem of implicit knowledge and of its capture in a manner which allows it to be included in business process design is also discussed, presenting some open research issues.
1 Introduction Today, the term “extended enterprise” is used designate any collection of organizations sharing a common set of goal. In this broad sense, an enterprise can be a whole corporation, a government organization, or a network of geographically distributed entities. Extended enterprise applications support digitalization of traditional business processes, adding new processes enabled by e-business technology (e.g. large scale Customer Relationship Management). Often, they span company boundaries, supporting a web of relationships between a company and its employees, managers, partners, customers, suppliers, and markets. Digital modeling improves the enterprise ability to dynamically J. Filipe and J. Cordeiro (Eds.): WEBIST 2007, LNBIP 8, pp. 19–32, 2008. c Springer-Verlag Berlin Heidelberg 2008
20
E. Damiani et al.
set up new business processes or to adjust existing ones to changing market conditions, increasing their ability to compete. Companies are therefore developing a growing interest in concepts, technologies and systems that help them to flexibly align their businesses and engineering processes to meet changing needs and to optimize their interactions with customers and business partners. In such a scenario, Business Process Modeling (BPM) techniques are becoming increasingly important. Less than a decade ago, BPM was known as “workflow design” and was aimed at describing human-based processes within a corporate department. Today, BPM is used to design orchestration of complex system interactions, including communicating with the processes of other companies according to well-defined technical contracts. Also, it is used to check compatibility and consistence of the business processes of collaborating business entities. A number of methodologies, languages and software tools have been proposed to support digital BPM; nonetheless, a lot remains to be done for assessing a business process model validity with respect to an existing organizational structure or w.r.t. external constraints like the ones imposed by security compliance regulations. In particular, web-based business coalitions and other inter-organizational transactions pose a number of research problems. Recently, the BPM and Web services communities have proposed many languages and standardization approaches that aim at describing business processes, especially from the perspective of Web services orchestration. Well-known examples include BPEL4WS, BPML, XLANG and WSFL. Most languages focus on the representation of the process message exchanges (the so-called process choreography) and on the control and data flow among the involved Web services (the process orchestration). The OMG’s Model Driven Architecture (MDA) provides a framework for representing processes at different levels of abstraction. In this paper we rely on a MDA-driven notion of business process model, composed of a static domain model including the domain entities and actors, plus a platform-independent workflow model providing a specification of process activities. We shall focus on semantics-aware business process representation techniques, introducing logics-based static domain models and their relationship with Description Logics and current Semantic Web metadata formats. Then, the problem of implicit knowledge and of its capture in a manner which allows it to be included in business process design is also discussed, presenting some open research issues. The paper is structured as follows: in Section 2 we introduce rule based business modeling, while in Section 3 we discuss some semantic models for rules. In Section 4, the relation of these semantic models with the ones proposed for Semantic Web metadata is discussed; Section 5 discussed how these two approaches can be successfully hybridized to meet modeling requirements. Finally, Section 6 shows a worked-out hybrid model example, and Section 8 draws the conclusion and outlines our future work.
2 Rule-Based Business Modeling Business modeling languages are aimed at providing companies with the capability of describing their internal business procedures using a formal syntax. Rule-based formalization, being close to controlled natural language statements, will will ensure that businesses involved in processes fully understand descriptions of themselves and other
Representing and Validating Digital Business Processes
21
participants in their business; also, it will enable organizations to flexibly adjust to new internal and B2B business circumstances. More specifically, rule-based business modeling is aimed at producing machine-understandable business statements in a standard rext-based notation. Business modeling activity usually starts from a shared, domainwide business vocabulary and derives an object-oriented model from it. The Object Management Group’s (OMG) Model Driven Architecture (MDA) decomposes the modeling acvtivity’s outcome into two main work products: – Business Rules (BR): declarative statements describing business domains: Controlled english syntax, easy enough to be understood by humans – Business Process Models (BPM): stepwise specifications of individual business processes Machine-understandable and detailed enough to be translated into computersupported workflows Business Rules determine what states and state transitions are possible or permitted for a given business domain, and may be of alethic or deontic modality. Alethic rules are used to model necessities (e.g. implied by physical laws) which cannot be violated, even in principle. For example, an alethic rule may state that an employee must be born on at most one date. Deontic rules are used to model obligations (e.g., resulting from company policy) which ought to be obeyed, but may be violated in real world scenarios. For example, a deontic rule may state that it is forbidden that any person smokes inside any company building. It is important to remark that widespread domain modeling languages such as the Universal Modeling Language (UML) typically express alethic statements only. When drawing a UML class diagram, for instance, the modeler is stating that domain objects belonging to each UML class MUST have the attribute list reported in the class definition, implicitly taking an alethic approach to domain modeling. In business practice, however, many statements are deontic, and it is often important (e.g., for computing metrics) to know if and how often they are violated. The recognition of the importance of this distinction was a major driver of OMG’s SBVR (Semantics of Business Vocabulary and Rules) proposal, aimed at specifying a business semantics definition layer on top of its software-oriented layers. We will now focus on BR syntax, postponing a discussion on their semantics to Section 3. To fix our ideas, in this Section we shall mention a few simple examples based on the (fictitious) global car rental company EU-Rent used in the Semantics of Business Vocabulary and Business Rules (SBVR) draft standard1. The basic types of business rules are listed below, together with some illustrative examples: – Business Rule (deontic: obligation or prohibition) – Example: It is required that the drop-off date of a rental precedes the expiration date on the driverlicense of the customer reserving the rental – Definition (alethic: something true by definition) – Example: The sales tax rate for a rental is the sales tax rate at the pick-up branch of the rental on the drop-off date of the rental 1
The EU-Rent case study was developed by Model Systems, Ltd., along with several other organizations, and has been used in many papers. The body of descriptions and examples may be freely used, providing its source is clearly acknowledged.
22
E. Damiani et al.
– Authorization (permission) – Example: A customer reserving a rental may provide what credit card guarantees the rental if the name on the credit card is the name of the customer or the credit card is for a corporate account authorized for use by the customer. In business rules, a structured vocabulary is used whose symbols belong to three lexical types, namely: – Terms – Examples: account, assigned car, branch, credit card, driver license, drop-off branch, drop-off date, pick-up branch – Names – Examples: EU-Rent, MilanBranch – Verbs (Forms of Expression) – Examples: credit card guarantees rental, customer has driver license Of course, in business rules the relation between terms and concepts is not one-to-one; a single concept can be represented by many terms. For instance, the concept “car” can be represented by alternative synonyms like “automobile” or by translations like “coche” or “voiture”. The BR language must also include a way to express facts as well as rules, and also metadata (annotations) about documents, facts, and rules. However, the OMG SBVR specification does not specify neither how to parse (controlled) natural language, nor how to phrase business rules. This is a source of flexibility, as the same fact can be represented in many equivalent sentential forms. For example, the assertion “the driver’s license expires on date XX/XX/XX” can be alternatively expressed as “The expiration date xx/xx/xxxxis on driver’s license” or as “The driver’s license has expiration date xx/xx/xxxx ”.
3 Logical Models for BR Languages The semantics of BRs has been traditionally described by providing a mapping from the business rules syntax to some well-known logic formalism. Three (potentially conflicting) basic requirements for the logical models underlying BRs have been identified: 1. High expressive power A basic overall requirement for the underlying logics is a high expressive power in specifying rules, in order to provide most of the description capabilities of controlled English. Business modelers accustomed to using English syntax would not accept any too severe limitation to the modeling language’s expressive power. 2. Tractability The need for performing rule checking and for executability, however, implies a strong desire for computational tractability. In other words, the underlying logics expressive power has to be carefully balanced against tractability. 3. Non-falsifiability BR logical semantics should rely on monotonic reasoning, i.e. on inference whose conclusions cannot be contradicted by simply adding new knowledge.
Representing and Validating Digital Business Processes
23
These three requirements have been often emphasized by business analysts and researchers as guidelines toward finding a correct logical interpretation of business models, but are NOT satisfied by current SBVR logical modeling. In order to understand this paradoxical situation, let us briefly discuss the state or research in this area. Much research work has been done to provide a logics-based model for BR including modalities. Indeed, supporting modalities does not mean that it is mandatory to map the BTs to a modal logic. For instance, work by the BR OMG team and specifically by Terry Halpin (including his open source package NORMA, which supports deontic and alethic rules, see sourceforge.net/projects/orm). The approach behind NORMA, the Object-Role Modeling (ORM) starts from controlled natural language to express available information in terms of simple or elementary facts. The information is then converted into an abstract model in terms of natural concepts, like objects and roles. NORMA addresses logical formalization for SBVR by mapping BR’s deontic modalities Required, Prohibited into modal operators obligatory (O), permitted (P) (used when no modality is specified in the rule) and forbidden (F). Deontic modal operators have the following rules w.r.t. negation: ¬Op ≡ P ¬p
(1)
¬P p ≡ O¬p ≡ F p
(2)
Other modal operators used for mapping BR alethic rules are necessary (true in all possible states of the business domain, ), possible (i.e. true in some state of the business domain, and , impossible (). Alethic operators’ negation rules are as follows: ¬♦p ≡ ¬p
(3)
¬♦p ≡ ♦¬p
(4)
Halpin’s NORMA approach represents BRs as rules where the only modal operator is the main rule operator, thus avoiding the need for a modal logics model. Some allowed SBVR formulations that violate this restriction may be transformed into an equivalent NORMA expression by applying modal negation rules and the Barcan formulae and their converses, i.e. ∀pF p ≡ ∀pF p (5) ∃p♦F p ≡ ♦∃pF p
(6)
For instance, the BR For each RentedCar, it is necessary that RentedCar was rented in at most one Date is transformed into It is necessary that each RentedCar was hired in at most one Date2 However, BR rules emerging from business modeling cannot be always transformed into rules where the only modal operator is the main operator. To support such cases, in principle there is no alternative but to adopt a semantics based on a modal logic; but 2
Another transformation that could be used in this context is the one based on Barcan formulae’s deontic variations, i.e. ∀pOF p ≡ O∀F p. We shall not discuss here in detail the application of these transformations to normalizing modal logic formulas; the interested reader can is referred to [12].
24
E. Damiani et al.
the choice of the “right” modal logic is by no means a trivial exercise, due to tractability and expressive power problems [12]. Modal logics reasoning engines do exist; for instance, MOLOG has been developed by the Applied Logic Group at IRIT from 1985 on, initially supported by the ESPRIT project ALPES. MOLOG is a general inference machine for building a large class of meta-interpreters. It handles Prolog clauses (with conjunctions and implications) qualified by modal operators, The language used by MOLOG can be multi-modal (i.e., contain several modal operators at the same time. Classical resolution is extended with modal resolution rules defining the operations that can be performed on the modal operators. according to the chosen modal logics. However, instead of choosing a modal logics and applying MOLOG-style modal reasoning, most current approaches to BR semantics are based on Horn Logic, the well-studied sublanguage of First-Order Logic (FOL) which is the basis of Logic Programming and counts on thousands of robust implementations. Of course, not every software rule engine is (or should be) able to process the entire Horn Logics; full Horn Logic rules are well-known to be Turing complete, hence undecidable. Again, here we shall not attempt to describe the BR-to-Horn mapping in full. Below we give some intuitive correspondences : – Each, Some: Quantifiers. – and, or, not: Logical Connectives. – The, An : Quantification and variable binding. For example. let’s consider a EU-Rent rule expressed in natural language as follows: “For class C rentals, if pick-up branch does not coincide with the drop-off branch, then rental duration cannot be less than two days”. This rule can be parsed to obtain a semiformal (parsed) rule syntax as follows: For each rental X, if X.class = C and X.pickup < > X.dropoff, then days(X.duration) > 1. The corresponding FOL formula can be written as follows: ∀x ∈ Rental, IsEqual(Class(x), C) ∪ IsEqual(P ickU p(x), DropOf f (x)) → IsGreater(Duration(x), 1) In many practical applications, it is possible to carry out deductions over a knowledge base composed of rules like Eq. 7 by standard Prolog-style forward- and backwardchaining rule reasoning. With respect to the requirements we identified at the beginning of the section, however, we remark that the OMG proposed format for encoding BR as generic Horn rules is undecidable in its full expressivity, so we can expect only partial implementations of SBRL reasoners. Also, this reasoning is clearly not monotonic (see the next Section 4.).
4 BR vs. Semantic Web-Style Knowledge Representation Semantic Web (SW) knowledge representation formats aim to integrate heterogeneous information sources by enabling decidable reasoning on metadata. Since individual Web-based systems are not expected to store an exhaustive “mirror” of external data
Representing and Validating Digital Business Processes
25
sources, Semantic Web languages (unlike FOL-based business rules) rely on monotonic reasoning. Intuitively, deductions that are made on data structures that are not exhaustively known to the system must not be contradicted by mere extension of the knowledge base (KB)3 . For applications, the most important effect of taking the SW modeling approach (as opposed to the SBVR one) is that monotonic reasoning yields a stronger notion of negation (¬): If individual i is not known to be member of class C (i.e., the assertion C(i) cannot be found in the KB), then it cannot be assumed to be member of class ¬C. This requirement, generally referred to as open-world assumption (OWA), has often been indicated as the correct interpretation of digital business models (see Requirement 3 in Section 3) because of the frequent under-specification of local business models and the possibly partial knowledge of remote data sources. In Sec. 5, concept Competitor will be defined as an example of open-world entity that may be included in a business model to adjust its dynamic behavior by considering external data structures. Open-world, SW-oriented knowledge representation languages like OWL [17] are based on a different subset of FOL than business rules. Namely, they rely on Description Logics (DL) [1]. However, mature implementations of DL reasoners exist and are constantly developing new features. OWL reasoners are capable of processing data structures that cannot be expressed in rule-based systems, such as recursive definitions and individuals not explicitly defined in the KB). As an example, consider the following SBVR definition and the corresponding DL axiom4: a product is produced by a manufacturer and is distributed in a country Product ≡ ∃producedBy.Manufacturer
∃distributedIn.Country If provided with the assertion Product(PROD-01), a DL reasoner will derive the following assertions with regard to individual PROD-01: ∃producedBy.Manufacturer(PROD-01)
(7)
∃distributedIn.Country(PROD-01)
(8)
Conversely, if individuals ITALY and COMP-01 are known to be related with PROD-01 as in the following assertions: distributedIn(PROD-01,ITALY) producedBy(PROD-01,COMP-01) 3
4
It is interesting to remark that, besides being different from BR semantics, monotonicity is also not the way mass market data storage applications, such as relational and object-oriented databases, interpret information when queried, e.g. by means of a SQL-like interface. Here we concatenate terms to obtain valid tokens, such as distributedIn and distinguish concept names with a leading uppercase letter.
26
E. Damiani et al.
a DL reasoner will derive, as a consequence of the definition above, the following assertion: Product(PROD-01)
(9)
The FOL counterpart of these derivations requires the biunivocal relationship to be split in two rules, because Horn clauses require the separation of the conditions triggering a rule (the antecedent or body) from the corresponding derivation (the consequent or head). Moreover, variables are assumed to be universally quantified; therefore it is not possible to achieve the exact semantics of the DL axiom above. Product(x) ← producedBy(x, y) ∧ Manufacturer(y) ∧ distributedIn(x, z) ∧ Country(z) Product(x) → producedBy(x, y) ∧ Manufacturer(y) ∧ distributedIn(x, z) ∧ Country(z)
Here only the first rule can be implemented, essentially because in Horn engines all variables in the consequent must also be present in the antecedent5. Consequently, only the implication in (9) can be derived using rule languages. While DL is technically a smaller FOL fragment than Horn logics, the above discussion shows that traditional production rules used by Horn rule engines are able to render only part of the semantics of DL-based formalisms. More importantly, rule systems share a weaker notion of negation (∼) according to which lack of knowledge is considered false (negative) knowledge, thus leading to a non-monotonic reasoning generally referred to as closed-world assumption (CWA): If individual i is not known to be member of class C (i.e., the assertion C(i) cannot be found in the KB), then it is assumed to be member of class ∼ C 6 . When modeling business domains, it should be possible to declare concepts and roles as closed (i.e., characterized by complete knowledge). Although in principle it would be possible to apply rule reasoning in order to evaluate a DL-based knowledge base in a closed-world fashion (e.g., by asserting negative facts on closed concepts), a more straightforward approach is the one of the so-called epistemic operators K and A. The former allows for evaluating DL concepts according to facts that are explicitly known to hold (i.e., have not been inferred by reasoning); as an example, the following concept definition identifies individuals whose filler of role producedBy is an individual that is known to be a Manufacturer instance: K-QUERY ≡ KproducedBy.KManufacturer
5
6
Unlike what happens in DL reasoners, which are capable of handling “hypotetical” individuals as fillers of roles producedBy and distributedIn. Note that we distinguish between classical negation as implemented by DL reasoners (¬) and the weaker notion implemented by rule systems (∼).
Representing and Validating Digital Business Processes
27
The epistemic operator K has been implemented (albeit only in queries, not as concept constructor) in the Pellet reasoner [14]. In turn, the auto-epistemic operator A is related with the notion negation-as-failure. Intuitively, AC corresponds to ¬ ∼ C. The following concept definition identifies individuals having as filler of property producedBy an individual that is assumed to be (i.e., is not explicitly asserted as not being) a Manufacturer instance. A-QUERY ≡ KproducedBy.AManufacturer
However, to the best of our knowledge, the autoepistemic operator A has not been implemented so far in any DL reasoner. While DL reasoners extensions look promising for their applications to business modeling, Horn rule reasoning remains useful to manage application-dependent data structures that need not be shared with the structural (DL-based) component of the KB. Also, when translating business rules into the corresponding logic representation, we may need to model constructs not expressible with DL alone. In fact, let aside the restriction to monadic and binary relations, it very easy to derive from SBVR rules role concatenations that are not comprised in the tree-shaped models that can be expressed with DL. As an example, the following SBVR rule =configures a simple cyclic pattern that nonetheless requires a Horn rule for its definition: a direct product is a product that is produced by a manufacturer that is a reseller of the product DirectProduct ← Product(x) ∧ producedBy(x, y)
∧Manufacturer(y) ∧ resellerOf(y, x)
5 Combining DL and Horn Rules As we have seen in Section 4, Horn rule engines do not entirely render the semantics of DL-based formalisms. On the other hand, when writing business rules we may need to model constructs which are not expressible with DL alone7 . This suggests the idea of representing complex business systems using separate structural (i.e., DL-based) and relational (i.e., rule-based) components . However, it is difficult to obtain all the expected logic entailments by mere composition of structural and relational reasoning on the corresponding portions of the KB. To clarify this, consider a barebones structural component constituted by concept definitions Manufacturer and Distributor, together with the following assertion, expressing that individual K-MART is known to be either a Manufacturer or a Distributor (although the system cannot tell which one): (Manufacturer Distributor)(K-MART) 7
Actually, the OWL 1.1 proposal extends the relational expressivity of the language to encompass role composition and additional features. We do not address the OWL 1.1 proposal in this paper as no mature software tools based on it are likely to be available in the short term.
28
E. Damiani et al.
Now we define concept Reseller in the relational component of the KB, as follows: a reseller is a manufacturer or a distributor Reseller(x) ← Manufacturer(x) Reseller(x) ← Distributor(x)
It is clear that is not possible to derive the assertion Reseller(K-MART) by applying the rules to the structural component of the KB or the other way around. The above example shows that even simple implications are not necessarily preserved when dividing the KB into separate components. Furthermore, exploiting full Horn clauses expressive power easily leads to undecidability. In order to achieve sound and complete reasoning (see Requirement 2 of Section 3) on hybrid KBs, it is essential to monitor the flow of assertions between the two components by constraining the possible clauses that populate the relational component of the KB. This can be done by requiring variables referred to by clauses to be associated with at least one non-DL atom in the rule body. The approach proposed in CARIN [11] requires variables from role predicates to satisfy this constrain. Another formalism, DL+log [16], generalizes existing results (namely, AL-log and DL-safe rules [4]) and proposes a weaker safety condition by applying this requirement only to predicates in the rule consequent. Finally, MKNF [5] is a recent proposal integrating all the above formalisms and relies on CARIN’s stronger notion of safety.
6 An Example of Hybrid Reasoning Let us now consider a hybrid knowledge base (HKB) composed of a structural component S, that is a theory in a subset of FOL (e.g. DL), and a rule component R (e.g. Datalog¬∨). In this example we follow the approach proposed in[16], where the rule component constrained in its interaction with the structural component according to a safeness condition. A = AS ∪ AR , with AS ∩ AR = 0 p ∈ AS is a structural predicate r ∈ AR is a rule component A HKB H is a pair (S, R), where the following safeness condition holds. Each rule R has the following form: p1 (X1 ) ∨ ... ∨ pn (Xn ) ← r1 (Y1 ), ...rn (Yn ), sk (Zk ), ∼ u1 (W1 ), ... ∼ un (Wn )
Such that: – – – –
each pi is a predicate of A; each ri , ui is a predicate of AR ; each si is a predicate of AS ; each variable occurring in R must occur in one of the ri ’s.
Representing and Validating Digital Business Processes
29
6.1 Hybrid Reasoning Let H be a KB where S defines the following predicates: ∀x P RODUCT (x) → ∃y.P RODUCED BY (x, y) ∧ MAN UF ACT URER(y) [A1] ∀x P RODUCT (x) → ∃y.DIST RIBUT ED IN (x, y) ∧ COUN T RY (y) [A2] ∀x COMP ET IT OR(x) → ∃y.SELL IN (x, y) ∧ T ARGET COUN T RY (y) [A3] ∀x T ARGET COUN T RY (x) → COUN T RY (x) [A4] ∀x MAN UF ACT URER(x) → RESELLER(x) [A5] ∀x DIST RIBUT OR(x) → RESELLER(x) [A6] ∀x IRRELEV AN TC OUN T RY (x) → COUN T RY (x) [A7] ∀x IRRELEV AN TC OUN T RY (x) → ¬T ARGET COUN T RY (x)) [A8] T ARGET COUN T RY (IT ALY ) [F1] T ARGET COUN T RY (F RAN CE) [F2] T ARGET COUN T RY (SP AIN ) [F3] COUN T RY (HUN GARY ) [F4] MAN UF ACT URER(COMP 01) [F5] DIST RIBUT OR(COMP 02) [F6] DIST RIBUT OR(COMP 03) [F7] SELL IN (COMP 01, IT ALY ) [F8] SELL IN (COMP 01, F RAN CE) [F9] SELL IN (COMP 03, HUN GARY ) [F10] P RODUCED BY (P ROD01, COMP 01) [F11] P RODUCED BY (P ROD03, COMP 03) [F12]
Also, in H, R defines the following predicates: DirectP roduct(x) ← resellerOf (y, x), P RODUCT (x), MAN UF ACT URER(y), P RODUCED BY (x, y) [R1] DirectCompetitor(x) ← resellerOf (y, x), DirectP roduct(x), COMP ET IT OR(y) [R2] resellerOf (COMP 01, P ROD01) [F13] resellerOf (COMP 03, P ROD03) [F14]
It can be easily verified that H satisfies the following ground assertions: – COM P ET IT OR(COM P 01). Based on axiom A3, selling in a target country it is sufficient to be a Competitor. – DirectP roduct(P ROD01) Based on axiom A1 together with facts F5 and F11, PROD01 is a Product; so rule R1 in conjunction with F13 satisfy this assertion. – DirectP roduct(P ROD03) Based on A1 and F11,; PROD03 is a Product so rule R1 in conjunction with F13 satisfy this assertion. Note that it is not relevant that COMP03 is not a Manufacturer because no assertion contradicts this possibility and according to the open world semantics a fact cannot be negated only on the basis of the lack of an assertion stating it. – DirectCompetitor(COM P 01) Since A3, COMP01 is a Competitor, then R2 in conjunction with F13 and R1 satisfy this assertion.
30
E. Damiani et al.
On the contrary, the following assertion cannot be derived in H: – ¬COM P ET IT OR(COM P 03). Due to the open world semantics a company selling in irrelevan t countries (i.e., in countries that are out of our companies’ target ) cannot be inferred not to be a Competitor, because the partial knowledge assumption do not justify us to consider asserted facts as the only possible ones. Note that this semantics prevents deriving false information without justification of asserted facts, and can be important in a strategical context.
7 Related Work Both the business and the software engineering communities have produced much work on Business Process Management. Providing a complete overview of the business management approach to process representation is clearly outside the scope of this paper. The reader interested in a business vision of BPM, seen as a tool for innovative management of virtual organizations, is referred to [8]. Software engineering research has been trying since long to add business process models to the palette of the system designer. For instance, the UML Profile for Automated Business Processes allows BPEL4WS processes to be modeled using ordinary UML tools. The BPEL4WS representation can be used to automatically generate web services artifacts (BPEL, WSDL, XSD) from a UML model meeting the profile, as described for instance in [7]. More recent work by Christian Huemer and Birgit Hofreiter [9] represents the semantics of UN/CEFACTs Modeling Methodology (UMM) business processes using BPEL, showing that BPEL is appropriate to capture business collaborations. An important aspect of inter-organizational business processes is governing different evolution paces due to distributed administration. Work by Stefanie Rinderle, Andreas Wombacher and Manfred Reichert [15] addresses one important challenge which has not been adequately addressed by other approaches, i.e. independent evolution of business process choreographies. If modifications are applied to processes in an uncontrolled manner, inconsistencies or errors might ensue. In particular, modifications of part of a shared process may affect the implementation of the other part of the process by other partners. The authors sketch a framework that allows process engineers to detect how changes to process fragments affect the whole process, relying on change semantics. Other works from the software engineering research community, focus on the alignment between business modeling and software development or execution. As discussed in [10] the general idea is that requirements models are built from organizational goals. Some of these works, such as for instance [13] or [6], are strongly related to UML and do not clearly specify some organizational aspects such as technology that implements business processes or the relationships among the different organizational views. Our own Knowledge Management group at the Software Engineering and Software Architectures Research (SESAR) lab, at the University of Milan’s Department of Information technology (http://ra.crema.unimi.it), has been carrying out work in business process representation and monitoring in the framework of the TEKNE and SecureSCM projects, as in [3] or [2].
Representing and Validating Digital Business Processes
31
8 Conclusions Over the past 30 years, many techniques have grown up around the development, management, and execution of business rules. In parallel, DL-based formalisms have been developed as a part of the Semantic Web proposal. Both technologies seem to have their own strengths and weaknesses. In this paper we we discussed the state of the art of business modeling formalisms, outlining the potential role of hybrid representations in business domain descriptions. We provided preliminary but hopefully useful answers to the following questions: – Are there ways to combine each technology to make them more beneficial than they are being used alone? – Are there enhancements that can be done to further leverage the rules technologies? These questions are being addressed by a collective interdisciplinary effort of the computer science and business management communities. At SESAR lab, we are developing answers to these questions as a part of our current research work. Acknowledgements. This work was partly funded by the Italian Ministry of Research under FIRB contract RBNE05FKZ2_004 (project TEKNE), and by the European Commission under project SecureSCM. The authors wish to thank Claudio Pizzi for his valuable comments on modal logic reasoning. Also, many thanks are due to Schahram Dustdar and Azzelarabe Taleb-Bendiab, Ernesto Damiani’s fellow keynote speakers at WEBIST, and to the conference chair Joaquim Filipe for creating a unique atmosphere for exchanging and discussing research ideas.
References 1. Baader, F., Calvanese, D., McGuinness, D.L., Nardi, D., Patel-Schneider, P.F. (eds.): The Description Logic Handbook: Theory, Implementation, and Applications. Cambridge University Press, Cambridge (2003) 2. Ceravolo, P., Damiani, E., Viviani, M.: Bottom-up extraction and trust-based refinement of ontology metadata. IEEE Transactions on Knowledge and Data Engineering 19(2), 149–163 (2007) 3. Ceravolo, P., Fugazza, C., Leida, M.: Modeling semantics of business rules. In: Proceedings of the Inaugural IEEE International Conference On Digital Ecosystems and Technologies (IEEE-DEST) (February 2007) 4. Donini, F.M., Lenzerini, M., Nardi, D., Schaerf, A.: AL-log: Integrating Datalog and Description Logics. J. of Intelligent Information Systems 10(3), 227–252 (1998) 5. Donini, F.M., Nardi, D., Rosati, R.: Description logics of minimal knowledge and negation as failure. ACM Trans. Comput. Logic 3(2), 177–225 (2002) 6. Eriksson, H.-E., Penker, M.: Business Modeling With UML: Business Patterns at Work. John Wiley & Sons, Inc., New York (1998) 7. Gardner, T.: Uml modelling of automated business processes with a mapping to bpel4ws. In: Proceedings of the First European Workshop on Object Oriented Systems (2003) 8. Fingar, P., Smith, H.: Business process management. The third wave
32
E. Damiani et al.
9. Hofreiter, B., Huemer, C.: Transforming UMM business collaboration models to BPEL. In: Meersman, R., Tari, Z., Corsaro, A. (eds.) OTM-WS 2004. LNCS, vol. 3292, pp. 507–519. Springer, Heidelberg (2004) 10. Diaz, J.S., De la Vara Gonzalez, J.L.: Business process-driven requirements engineering: a goal-based approach. In: Proceedings of the 8th Workshop on Business Process Modeling, Development, and Support (2007) 11. Levy, A.Y., Rousset, M.-C.: CARIN: A representation language combining horn rules and description logics. In: European Conference on Artificial Intelligence, pp. 323–327 (1996) 12. Linehan, M.: Semantics in model-driven business design. In: Proceedings of 2nd International Semantic Web Policy Workshop (2006) 13. Marshall, C.: Enterprise modeling with UML: designing successful software through business analysis. Addison-Wesley Longman Ltd., Essex (2000) 14. Parsia, B., Sivrin, E., Grove, M., Alford, R.: Pellet OWL Reasoner (2003) 15. Rinderle, S., Wombacher, A., Reichert, M.: On the controlled evolution of process choreographies. In: ICDE 2006: Proceedings of the 22nd International Conference on Data Engineering (ICDE 2006), Washington, DC, USA, p. 124. IEEE Computer Society, Los Alamitos (2006) 16. Rosati, R.: DL+log: Tight Integration of Description Logics and Disjunctive Datalog. In: KR 2006, pp. 68–78 (2006) 17. W3C. OWL Web Ontology Language Overview, http://www.w3.org/TR/owl-features/
Part I
Internet Technology
Semi-automated Content Zoning of Spam Emails Claudine Brucks, Michael Hilker, Christoph Schommer, Cynthia Wagner, and Ralph Weires University of Luxembourg MINE Research Group, ILIAS Laboratory Dept. of Computer Science and Communication 6, Rue Richard Coudenhove-Kalergi, 1359 Luxembourg, Luxembourg Abstract. With the use of electronic mail as a communicative medium, the occurrence of spam emails has increased and kept stable; and although spam can be detected by spam filtering, an element of risk of a continuously network overload remains. To add insult to injury, spammers augment novel, clever and astute techniques, which currently leads to a neck-and-neck race between protection and injury. Constantly novel approaches are necessitated to enhance spam filters through a more efficient and effective prevention. In this respect, we argue for a zoning of emails to classify a text in regions, where each region bears a meaning. Here, we use this technique for spams to identify these zones and to evaluate if an email is a spam. We introduce content zoning and prove our approach by a prototypical implementation including first results. Keywords: Content Zoning, Spam Email Detection, Text Analysis and Statistics, Human-Computer Interaction.
1 Introduction Originally, the word spam had firstly been referred to a meat product and had been an abbreviation of Shoulder of Pork and hAM (SPAM), but later on, to Spiced Pork And Meat. By a sketch of Monty Python’s Comedy team, the meaning of the word spam has changed as consequence of spammed by a group of Vikings repeating a song containing . . . spam, lovely spam. . . Due to this excessive use, the word spam has become a term in the Internet, meaning any kind of unsolicited messages or junk emails or bulk emails. The distribution of spam emails is understood as spamming, and the sender is named a spammer. While speaking about spam emails, a multitude of different spam types exist. Unsolicited commercial mail are usually dubious emails that offer products to special price rates, like the advertisement of pharmaceuticals. Chain letters comprise texts with rumors/news instruct the recipient to forward. Malicious forms are worms and viruses, which cause computer damages, and phishing emails that may cause personal damages - mostly of financial type or by requesting sensitive data. Both normal and spam emails consist of several parts, for example the email header and the email content (see Figure 1): the header includes the technical information, protocols, or spam flags, the subject line, and the sender/recipient information. The email content includes the technical information (source code), a readable text (content), including images, and the signature. In the recent years, the importance of electronic messaging is growing and the question of how to control spam emails more and more arises. J. Filipe and J. Cordeiro (Eds.): WEBIST 2007, LNBIP 8, pp. 35–44, 2008. c Springer-Verlag Berlin Heidelberg 2008
36
C. Brucks et al.
Fig. 1. Source code of an electronic mail: Zone 1 represents the HEADER of the email including SUBJECT (marked in green). Zone 3 is the CONTENT, Zone 4 the SIGNATURE.
Many solutions already exist and improvements to avoid this kind of spam messaging are mostly of similar nature. In [6], honeypots accounts are created and positioned to receive spam. Spam filter engines often refer to a rule-based knowledge structure [1,7,10]. Intrusion Detection Systems (IDS) have been enabled [5] that simulate the natural body system control while generating different kinds of cells and cell communication as well. In this respect, each delivered email package, which is sent from one network computer to another, is controlled: if a certain (spam) pattern is present, then the packet is deleted from the network to minimize the risk of intrusion. Beside of that, spam filters exist that directly detect spams, for example by white lists and black lists that contain accepted and denied email addresses. Or, patterns are fixed that is foreseen as to be used by spammers. Another alternative is to analyse the header and the content of emails to classify in spam and non-spam. In fact, natural language filter systems remove emails that are written in a foreign language as well as user-defined rules. Indeed, the spammers continuously introduce novel techniques in order to fail spam filters and to hide spam emails. For example, spam emails can contain text hidden between complex HTML tags or replace characters, e.g. zeros by O’s - a challenge for common-used spam filters. Another example is the recent trend in spam emails of using scrambled words, which means that text - correctly understood by human beings and identified by machines
Semi-automated Content Zoning of Spam Emails
37
with access to natural language dictionaries (thesauri) - cannot be processed seriously by normal spam filters: Aoccdrnig to a rscheearch at Cmabrigde Uinervtisy, it deosn’t mttaer in waht oredr the ltteers in a wrod are, the olny iprmoetnt tihng is taht the frist and lsat ltteer be at the rghit pclae. . . 1 In this article, we follow the idea of content zoning as it is shown in Figure 1 to improve the quality of spam filtering. With content zoning, a text becomes composed of different regions (zones) where each region receives a meaning. For example, a zone SUBJECT may describe the subject of an email, OFFER to the advertised product itself. Therefore, we introduce our content zoning solution and demonstrate it with COZO, an interactive content zoning system [3], on two disjunctive domains: finance and pharmacy. We conclude the paper with an outlook.
2 Content Zoning In [4,9], Argumentative Zoning - that refers to annotated zones to describe and to explain scientific documents - is presented. Its aim is to provide an intelligent library towards the generation of summaries on scientific articles and the detection of intellectual ownership of documents, e.g. for a better document management or for plagiarism detection. 2.1 Annotated Regions in Texts Our approach bases on similar principles to analyse spam emails, as means to separate spam from non-spam mails. A part of this work can also be seen in [3], which deals with the manual analysis of up-to-date spam emails to check if they contain significant patterns in the subject line and the content section. We focus on discovering content structures to describe common patterns in spam emails. With this, we intend to gain information that can be used to perform automatic recognition of zones for further work. Automatically derived information from spam emails can then support a classification of spam emails while separating them from non-spam emails. The aim of the content zoning technique is to provide a useful function or add-on to normal spam filters for the detection of patterns in the content of spam emails and because of the questions if spam emails have characteristic content structures. The separation into zones is done according to the given structure of the documents (e.g. different paragraphs) and especially because of the semantic meaning of the document parts. In the end, the zones are supposed to represent the various semantic units of the document. Besides the plain division of a document in zones, additional analysis can also be performed to gain further information about the separate zones. Rather simple to extract information could be statistics about the size and layout of zones but a more sophisticated analysis of their text content is also possible. The latter can lead to an extraction of the semantic content and purpose of a zone. Such kind of information can be used for various purposes, 1
http://www.mrc-cbu.cam.ac.uk/∼mattd/Cmabrigde/ - November 2007.
38
C. Brucks et al.
such as comparing documents to each other (regarding their analaysed zone structure and content). In the following, we refer to such information about the zones as zone variables. 2.2 Identification of Spam Emails We focus on a test set of spam emails that has been applied to the process of zoning; the test set mainly refers to the different categories of pharmaceutical (i.e. drug offerings) and financial email texts (more precisely, stock spam emails). In the following, we give a more detailed description of how the zones for these two types of spam emails are defined and which zone variables are used. A zone can be defined as a region in a document that is allocated to a specific information in this document. For example, when having a text document with a price offer, this price offer can be annotated by a tag; a logical name for that zone could be price offer. Of course, when having a huge document, zones can reoccur multiple times. Figure 1 has illustrated an example of a pharmaceutical spam email; each zone is represented by a color and an annotated text (HEADER, SUBJECT, CONTENT, SIGNATURE). Financial emails typically introduce themselves as texts of large content, where specific categories like the price, the offerings or the expectation of profits have their place including a very detailed and accurate description. Pharmaceutical texts, on the other hand, refer to texts of shorter size; its content includes in-note-form-like texts with direct offerings and less descriptive text. Sometimes, pharmaceutical spam emails are composed of an image or the text is written completely in white color. Generally, while analyzing financial and pharmaceutical spam emails and characterizing their content, respectively, it has become obvious that pharmaceutical spam emails have a lower amount of zones than financial spam emails [3]. To get granular on electronic spam mails of pharmacy and finance, we have observed that the most redundant text rubrics have been considered as to be the default zones. Since irrelevant text is often included in spam emails, a content may remain unsupported by zoning. The annotation tags have logical names, so that no supplementary information is needed to understand what the zone is about: – INFORMATION: The introductive zone that informs the reader about the intention of the email sending. As an example, the text motivates any kind of lack or absence. – OFFER: This zone represents the offer itself, for example a product or a service. It is accompanied with specific advantages and auspicious expectations. – PRICE: In the PRICE zone, the effective price is put. In many cases, it is a sub-zone of the zone OFFER. – LINK: This zone triggers as connective link to external web-sources. When clicking, the user becomes automatically connected. However, a definite and overall definition of the zones is hard to give, because content zoning alludes more to a informational positions in the text; it does not analyze content via text mining functionalities. We therefore introduce zone variables to identify parts of the text. A zone variable is defined as a parameter that describes specific information of a zone. By describing the zones with zone variables, more structured information can be
Semi-automated Content Zoning of Spam Emails
39
extracted of the zone content. The zone variables are one of the most important factors in content zoning, because they contribute to the statistical evaluation and finally to the detection of similarities in spam emails. In this work, two kinds of variables have been defined. Zone independent variables are defined as parameters which can be applied to every zone. Examples for zone independent variables are for example the position of the zone in the text, the length of the zone, expressed in number of characters, the number of words contained in the zone, or the most occurring word (under consideration of stopword elimination). Zone dependent variables on the other hand are defined as parameters, which are specific for a certain type of zone and hence can only be applied to zones of that type. Mostly, this kind of variables represents semantic parameters, for example top-level domain for links (e.g. .com, .net, .lu), telephone numbers for the address zones, value for price zones, currency for price zones (e, £, $) and so on.
3 The COZO Explorer The COZO Explorer (see Figure 2) integrates a spam thesaurus content-wide and a graphical user interface towards zoning in texts. Furthermore, it supports an automatic image sorting. 3.1 A Spam Thesaurus The Spam-Thesaurus refers to the content of electronic mails and follows the idea that transformation of different writing styles to a consistent form of spam keywords may
Fig. 2. The COZO Explorer consists of two panels: the reader panel (left) and the options panel (right) with the predefined 18 zones. A zone can be marked and colorized, the result is shown in the reader panel.
40
C. Brucks et al.
improve the prematurely detection of spams and an advancement of spam-filtering. In COZO, the thesaurus is a list that is continuously expanded by reading spam email streams. It consists of the most occurring words of all extracted subject lines plus their standardized meaning. Currently, the thesaurus has a keyword-list of more than one thousand entries, excluding its flections, which vary of lot: for example, the word antibiotics has no flection as the word Viagra compounds has more than 120. 3.2 The Graphical User Interface Most common used email types (HTML, EML, etc.) are supported, irrelevant information, like email header or HTML tags, can automatically be removed on demand; only the content remains. The user can zone the content. The results can be exported and stored in XML. In Figure 2, the zones are selected for a given text. To facilitate the zoning, 18 zones have been predefined. However, if there is no convenient zone for a certain email region, the user has the possibility to create his specific zones or to leave that region unzoned. Each zone is visualised by a different colour; for each zone variables, statistical information like the position of the zone in the email, the density of the zone, and the amount of characters in the zone, are computed. The output - definition of zones, calculated variables, and a picture of the zoned email with colours - is also stored in files for further analysis, e.g. picture matching. At this point of time, the zoning must be done manually.
4 Selected Results In the following, we will present some results from a set of spam emails captured from a period of nine months. The data set contains emails from various domains (e.g. credit offers, movie downloads, porn emails), but only spam emails from pharmacy and finance have been tested. 4.1 Zoning Pharmaceutical Spam Emails The mechanism of zoning is firstly applied to spam emails pre-classified as pharmaceutical emails. The number of the test spam set is 100, we have principally taken the four general zones (see Section 2.2) but a bit more adapted to the mentioned domain. For pharmacy, the OFFER zone represents the part where the pharmaceutical product is promoted, LINK the position where the linked homepage is placed; the zone INFORMATION refers to pharmaceutical product descriptions, PRICE to the indication of the price: interestingly, this zone occurs rather seldom. While zoning the spam emails, diverse observations can be done, for example that 64% of the spam emails share a sequential pattern of SUBJECT - OFFER - LINK INFORMATION whereas 16% a sequential pattern SUBJECT - OFFER - LINK. 8% has not a pattern at all. For the tests, no supplementary zones have been generated, the initial 18 zones mostly cover each email content. The zones of the emails from the data set have been colorized as in Figure 3. After having computed the zoning, the statistics are calculated and stored in a XML-file with tags like
Semi-automated Content Zoning of Spam Emails
41
Fig. 3. Content Zoning of an spam email from finance: the zones are colorized and marked and annotated by the rectangles. The pattern ZON E3 → ZON E4 is continuously present. 399580 ...
We have obtained different results, for example that pharmaceutical spam share quite often a similar zoning. Consequently, the order and the size of the zones are similar as well as the calculated variables of each zone. 4.2 Zoning Financial Spam Emails For financial spam emails, the number of the test spam set has been set to 100 as well. Here, we have established principally a list of the following 8 zones: – – – –
ZONE 1 (NAME): Complete name of the company whose shares are offered. ZONE 2 (DESCRIPTION): Description of the company’s offering. ZONE 3 (CURRENT): Current price of the stock. ZONE 4 (PROMISED): Promised price, i.e. the speculative price that the share will reach in a few days time. – ZONE 5 (PROMOTION): Promotion, i.e. arguments to motivate the reader to buy the shares. – ZONE 6 (FACTS): Facts about the company, i.e. a short historical review of the company and its products. – ZONE 7 (NEWS): News about the company and its activities.
42
C. Brucks et al.
– ZONE 8 (FORWARD): Forward looking statement, i.e. a clear statement that no responsibility is taken over. These different zones have been set up manually, and at the beginning 20% of the financial emails have been analyzed to discover any kind of sequential zoning pattern. As financial, especially stock market spam emails, are assumed to be similar in their structure, nine different zones could be determined quite easily and then matched to the rest of the stock mails. Surprisingly, diverse patterns of zone sequences have been discovered, for example that a ZON E3 is always followed by ZON E4 (see Figure 3)2 .
Fig. 4. Coincidental arrangement of zoned spam emails before the image sort has taken place; financial and pharmaceutical spam emails are disordered
4.3 Automated Image Sorting As another alternative for spam email classification, we have downloaded and deployed the freeware software product ImageSorter3 to inspect collections of zoned spam emails and to sort them according to their colorized zones. As we have observed, both financial and pharmaceutical zoned spam emails differ in the number, occurrence and size. We take advantage of this and comply with the idea of accomplishing classification as computing the similarity between images: and, the similarity between zoned spam emails is geared to the color gradient of images. The figures 4 and 5 show some zoned financial and zoned pharmaceutical spam emails before and after an image sort has been applied. Zoned pharmaceutical spam emails are placed to the top; they are separated from zoned financial spam emails that are placed to the bottom. 2
3
The presented rectangles are just for a better understanding; COZO colorizes the zones but does not place a rectangle around it. ImageSorter is a freeware software that is currently developed at the Center of HumanComputer Interaction at the FHTW Berlin.
Semi-automated Content Zoning of Spam Emails
43
Fig. 5. Sorted zoned spam emails where zoned financial spam emails are put to the bottom, pharmaceutical to the top
5 Conclusions There are many advanced possibilities to classify spam emails using content zoning, for example to more automate the content zoning process including the discovery of annotated labels or to define a stronger similarity metrics according to the calculated zone statistics (zone ordering, zone sizes, etc.). The expansion of the idea of image sorting in order to realize the comparison between electronic mails is another claim: since the actual comparison is of course not limited to basic (exact) picture matching but must include more robust image sorting techniques as well. With this, we are able to perform a comparison of a new email to our existing spam email types (financial and pharmaceutical here) to decide whether or not the email is spam. Acknowledgments. This work is funded by the University of Luxembourg within the project TRIAS. We thank all the colleagues and friends of the MINE research group within the Intelligent and Adaptive Systems Laboratory (ILIAS).
References 1. Androutsopoulos, I., Koutsias, J., Chandrinos, K., Paliouras, G., Spyropoulos, C.: An evaluation of naive bayesian anti-spam filtering (2000) 2. Bayerl, S., Bollinger, T., Schommer, C.: Applying Models with Scoring. In: Proceedings on 3rd International Conference on Data Mining Methods and Databases for Engineering, Finance and Other Fields, Bologna, Italy, September 25-27 (2002)
44
C. Brucks et al.
3. Brucks, C., Wagner, C.: Spam analysis for network protection. University of Luxembourg (2006) 4. Feltrim, V., Teufel, S., Gracas Nunes, G., Alusio, S.: Argumentative Zoning applied to Critiquing Novices Scientific Abstracts. In: Computing Attitude and Affect in Text: Theory and Applications, pp. 233–245. Springer, Dordrecht (2005) 5. Hilker, M., Schommer, C.: SANA security analysis in internet traffic through artificial immune systems. In: Proceedings of the Trustworthy Software Workshop Saarbruecken, Germany (2006) 6. Lambert, A.: Analysis of Spam. Master Thesis, University of Dublin (2003) 7. Rigoutsos, I., Chung-Kwei, T.H.: A pattern-discovery-based system for the automatic identification of unsolicited e-mail messages (spam). In: Proc. of the Conference on Email and Anti-Spam (CEAS) (2004) 8. Sun, Q., Schommer, C., Lang, A.: Integration of Manual and Automatic Text Classification A Text Categorization for Emails and Spams. In: Biundo, S., Fr¨uhwirth, T., Palm, G. (eds.) KI 2004. LNCS (LNAI), vol. 3238, pp. 156–167. Springer, Heidelberg (2004) 9. Teufel, S.: Argumentative zoning: Information extraction from scientific text, Phd Thesis, University of Edinburgh, England (1999) 10. Wittel, G., Wu, S.: On attacking statistical spam filters. In: Proc. of the Conference on Email and Anti-Spam (CEAS) (2004)
Security and Business Risks from Early Design of Web-Based Systems Rattikorn Hewett Department of Computer Science, Texas Tech University Abilene TX 79601, U.S.A [email protected]
Abstract. This paper presents a systematic approach for the automated assessment of security and business risks of web-based systems at the early design stage. The approach combines risk concepts in reliability engineering with heuristics using characteristics of software and hardware deployment design to estimate security and business risks of the system to be developed. It provides a mechanism that can help locate high-risk software components. We discuss limitations of the approach and give an illustration in an industrial engineering and businessto-business domain using a case study of a web-based material requirements planning system for a manufacturing enterprise. Keywords: Software risk assessment, software security, business risks, software architecture, internet-based system.
1 Introduction Web application systems play important roles in providing functions and business services for many organizations. The scope and complexity of web applications have grown significantly from small-scale information dissemination to large-scale sophisticated systems that drive services, collaboration and business on a global scale. Unlike conventional software systems, web-based systems grow and change rapidly. They also tend to be living systems having long life cycles with no specific releases and thus, require continuous development and maintenance [7]. Designing web applications involves not only technical but also social and political factors of the system to be built. Poorly developed web applications that continue to expand could cause many serious consequences. Because they are ubiquitous and deal with a huge and diverse population of users, they must be tolerant to errors, adverse interactions and malicious attacks. Even if they perform perfectly, they can still put the business of the enterprise deploying them at risk if they do not support sound business logic. Effective web-based system design involves both technical and managerial decisions that require a balanced integration of economic flexibility, sound engineering judgment and organizational support. A thorough understanding of risk can help provide the connection between technical concepts, user values and values created by the design. The ability to quickly estimate security and business risks early in the system development life cycle can help software developers and managers make risk-acceptable, J. Filipe and J. Cordeiro (Eds.): WEBIST 2007, LNBIP 8, pp. 45–59, 2008. c Springer-Verlag Berlin Heidelberg 2008
46
R. Hewett
cost-effective and informed decisions. In this paper, the terms security risk and business risk refer to the possibilities of adverse consequences, due to software behavior, to resource ownerships, and to business objectives of the system in which the software is used, respectively. Various techniques and guidelines to security risk analysis are available [10,16] and activities to increase security assurance of software are integrated throughout the software development life cycle (e.g., security requirements, code analysis, security and penetration testing). However, most developers perform security risk assessments of software systems at a system level after the systems have been implemented. For business challenges, efforts on incorporating business processes into enterprise applications are mostly in a requirement stage in order to create a design [4,6,13,15]. However, analyses of business implications of the design products are seldom addressed structurally. Software vulnerabilities can occur in codes (e.g., buffer overflow, SQL injection) and designs (e.g., internal controls that attackers can compromise). They can even exist in components that are not involved with security procedures. As many software-intensive infrastructure systems increasingly utilize COTS (Commercial Off-The-Shelf) components for rapid application development, ability to analyze security risks at a software component level becomes essential. As web applications are implemented on standard platforms, they inherit similar security weaknesses that are published and available publicly [1]. These known vulnerability data-bases are used widely as parts of basic principles for guidelines, standards, and vulnerability scanning tools such as Nessus [2]. Currently, like other information systems, most web-based systems rely on standard checklists, firewall rules and scanners for system security risk testing. One drawback of these tools is their lack of ability to assess vulnerabilities specific to custom web applications [17]. Most traditional risk analysis techniques are difficult to apply directly to modern software design, as they do not necessarily address potential vulnerabilities and threats at a component or environment level [18]. Particularly, web-based information systems are designed on multi-tier architectures spanning over multiple boundaries of trust. A vulnerability of a component depends on both its platform (e.g., J2EE on Tomcat/Apache/Linux, C# on a Windows .NET server) and environment (e.g., exposed network such as LAN versus secure DMZ). To adapt risk methodology to web application design, security risk analysis should consider risk factors from different design levels including software components, hardware platforms and network environments. Risk analysis provides assessment of risk to serve as a basis for decision-making to prevent, mitigate, control and manage risks of loss. This paper presents a systematic approach for automated security and business risk assessment of a web application system from its component-based designs. By combining risk concepts in reliability engineering with heuristics using characteristics of software design and hardware platforms, the security and business risks of a proposed system can be estimated. In contrast with rigorous risk analysis during the operational phase, our approach provides a simple early estimate of risks that can help locate design components that are at risk. Thus, mitigating strategies (e.g., allocating resources for rigorous testing, or redesigning the high-risk component) can follow.
Security and Business Risks from Early Design of Web-Based Systems
47
Section 2 describes related work. Section 3 defines relevant security terminologies and risk concepts, and describes the proposed approach for security risk assessment. Section 4 illustrates our approach to a web application in industrial engineering. Section 5 describes how the proposed approach can be adapted for assessment of business risks. Conclusion and future work are discussed in Section 6.
2 Related Work Much research in web information systems is concerned with techniques and methodologies for engineering web-based systems effectively by adapting principles from software engineering to web application design and development [3,12,7]. Risk methodologies have been employed to assess safety and performance associated with software systems in various application domains including industrial engineering, business and space science [14,5,19]. Risk methodology including FMEA (Failure Mode and Effect Analysis) has been applied to gain understanding about usability of web information systems [20]. Unlike our approach, none of these work addresses risks in security contexts. Identifying likelihoods in risk methodology in these web information systems are mostly subjective rather than analytical and objective. Our work is most similar to a methodology and framework proposed by Yacoub et al. [19]. However, instead of assessing the reliability of software components, we assess the security and business risks associated with software. Thus, their risk models contain different attributes than ours. Reliability risks are estimated from the complexity of software components whereas security and business risks are based on characteristics of software design as well as hardware platforms to be deployed.
3 Security Risk Analysis In this section, we define basic risk analysis concepts and relevant security elements in Section 3.1. Section 3.2 describes the proposed methodology. 3.1 Preliminaries Risk is defined as a multiplication of the probability of occurrence of an undesired event and its consequences [9]. Risk analysis is a process that aims to assess, manage and communicate risk information to assist decision-making. Risk assessment involves identification of adverse situations (or hazards or threats) and quantifying their risks by determining all likelihoods of their causes and the severity of their impacts. The risk measures obtained can assist in cost benefit study for resource planning to prevent, eliminate or mitigate these undesired situations. These are activities of risk management. Many techniques for risk assessment exist (e.g., FMEA, HAZOP (Hazard and Operability), FTA (Fault Tree Analysis) [8]), but they have been mainly applied to hardware rather than software artifacts. Risks associated with software (or software risks) are risks that account for undesired consequences of software behavior, in the system in which the software is used.
48
R. Hewett
Security and safety are two related perspectives of value judgments of acceptable risks. Security is concerned with the protection of ownerships of assets and privacy, while safety is more concerned with the protection of human lives and environments, factors that cannot be evaluated easily in monetary terms. Examples also include some assets in web applications e.g., trust and a company’s reputation. Sources of threats include malicious (e.g., deliberate attack, code injection) as well as non-malicious actions (e.g., natural disaster, human error, software system failure) that may lead to a security breach. Security attacks, whether intentionally targeted or not, can originate from internal and external threats depending on the degree of exposure of assets and how vulnerable the system is to attack. A system’s vulnerability refers to its weakness (i.e., specific attribute/environment) that makes a system more prone to attack or makes an attack more likely to succeed. Exploited vulnerabilities can compromise security. In finance, risks are estimated in terms of expected loss per year, ALE (Annualized Loss Expectancy), which is a product of the annual rate of occurrence and single loss expectancy. The former reflects probability, and the latter for consequences over a period of time. In security risk analysis, risks are estimated from adapted ALE formulae. Basic probability factors are related to exposure degrees, vulnerabilities and threats, whereas basic consequence factors are measured by asset values. Thus, for a given threat/adverse action, its security risk can be roughly estimated by: Risk = threat × exposure × vulnerability × severity where threat actually measures the threat likelihood. The first three contributes to probability measures, whereas the last measures its consequence. In this paper we omit threat, which can be easily integrated into our analysis at a later stage of the design. While the ALE concept works well in general financial and security risks, it may not be applicable for cases like terrorist attacks, which are rare or may never occur. 3.2 Methodology Quantitative risk assessment methodologies, such as the ALE-based approach, provide useful mecha-nisms to facilitate risk ranking for decision-making in risk management. However, they are often criticised for producing results that are too general (statisticlike) to be useful for specific decision-making, and producing results that are not credible enough as they rely heavily on subjective value assignments. The authors’ proposed methodology given below has two objectives: (1) to estimate security risks of a web-based software system from its early design configuration, and (2) to estimate probability measures more objectively. The latter is achieved by using heuristics based on relevant design attributes (more details later). Because web applications are typically designed with multiple architectural layers and trust boundaries, our risk methodology must incorporate relevant attributes from multiple design levels (e.g., component, network and hardware platforms). Figure 1 shows basic steps in the methodology. The methodology requires two inputs: a software design and a hardware deployment design. We assume that the software design is expressed in terms of an extended activity
Security and Business Risks from Early Design of Web-Based Systems
49
Fig. 1. Proposed risk methodology
or workflow diagram, where the software components of the web applications and use case actors are represented as well as activities in use cases [3]. The software design includes all use cases, each of which may have one or more scenarios. The hardware deployment design provides information about hardware platforms including host or server, their attributes (e.g., services, operating environments), and network attributes (e.g. IP addresses for network configuration, and LAN, WAN or DMZ for connection). Security scanning tools use this information to identify vulnerabilities of each host in the system. The hardware deployment plan also indicates software components designated to use, or run on, corresponding hosts. Examples of a workflow diagram and a hardware deployment diagram are shown in Figures 3 and 4, respectively. As shown in Figure 1, the first two steps use software design characteristics as heuristic measures of probability factors of risk. Step 1) computes interaction rates among software components for each scenario based on the design workflow. For simplicity, we assume that each outgoing interaction from the same component occurs equally likely (when insights about interaction likelihoods are available, appropriate weights can be applied). Step 2) constructs a graph whose nodes represent software components. Transition rates among components are computed by taking scenario likelihoods into account for combining interaction rates from each scenario. In security risk contexts, scenario likelihoods are heuristic measures of probabilities of potential threats (e.g., when scenarios involve adverse actions), whereas interaction rates are heuristics that partially measure degrees of component exposure. Step 3) uses characteristics of hardware deployment design as a heuristic measure of an additional probability factor of security risk. We first determine vulnerability sources of each host. This can be done by a table lookup procedure on publicly available vulnerability databases for vulnerability sources based on relevant host and network attributes (e.g., IP addresses, services). Examples of vulnerability sources include Apache Chunked-Code software on web servers, operating environments Windows XP SP2 and
50
R. Hewett
JRE 1.4.2 for running JAVA programs on application servers, and Oracle and TNS Listener software for database servers. Such vulnerabilities and their sources can be easily identified by scanning tools when hardware platforms to be deployed are already in operation. Next step is to measure vulnerabilities of each software component. Each component may require one or more hosts. For example, a component that runs on one application server may also need to query data on another database server. From the deployment design, we can determine all hosts required by each component. Let Hi be a set of hosts required by software component i and vh be a number of vulnerabilities of host h. For component i, a number of vulnerabilities, vi and its normalization to a vulnerability ratio, Vi are estimated as: vh and Vi = vi / vi . vi = h∈Hi
i
Here vulnerability ratios of software components are heuristic measures of likelihood factors of security risks. The more vulnerable software component is, the more likely that the system security can be compromised. Step 4) combines all probability factors obtained in Steps 2) and 3) to estimate the likelihood of each component being attacked (or security breach). Here we interpret this likelihood as a chance of a component under investigation to be the first being attacked in the system. Thus, all components and links along the path to reach the component are secured (unexposed to attack and unexploited vulnerability). The likelihood of a node being secured is estimated by subtracting its exposure degree, estimated by its weaknesses (i.e., vulnerability ratio), from one.
Fig. 2. Likelihoods of security breach in components
We elaborate this step by a small example of a dependency graph shown in Figure 2, where nodes S, A, B, C and D represent components of a software design. A doublelined arrow indicates that S is an entry point to software system execution. Each component and transition is annotated with a corresponding vulnerability ratio (in bold) and a transition rate, respectively. As shown in Figure 2, the likelihood of S to breach security is determined by the vulnerability ratio of S itself. The chance of A being attacked (via S → A) is estimated by its chance to be reached without attacks (i.e., S being secured and S transits to A) and its exposure factor (weaknesses). The former is estimated by the likelihoods of S being secured and S transiting to A, whereas the latter is estimated by A’s vulnerability ratio. Thus, the chance of A being attacked is estimated
Security and Business Risks from Early Design of Web-Based Systems
51
by (1 − 0.3) × 0.2 × 0.1 = 0.014. Similarly, the likelihood estimate of security breach in B is 0.224. The chance of C being attacked is estimated by the sum of the chances of C being attacked via S → A → C (i.e., (1 − 0.3) × 0.2 × (1 − 0.1) × 1 × 0.1 = 0.0126) and via S → B → C (i.e., (1 − 0.3) × 0.8 × (1 − 0.4) × 0.1 × 0.1 = 0.00336). This gives 0.01596. In general, the estimated probability along an access path is analogous to propagation of probabilistic effects in Bayesian network by using Bayes Rule [11]. The transition rate plays a similar role to a conditional probability of a destination being attacked given a source being reached. For simplicity, we do not explicitly represent access privilege. We assume that any activity in the same component can be equally accessed. However, not every component deployed by the same host can be accessed equally. We also assume that network connections between hosts do not have vulnerabilities. Such vulnerabilities can be easily incorporated in our proposed approach. Note that since all probability factors (transition rate, vulnerability ratio) are no greater than one, the longer the path to reach a component is, the harder it is to attack (i.e., its chance of being attacked decreases). For a path that contains cycles, we consider the maximum likelihood that can be obtained. Thus, the heuristic function we use for estimate likelihoods conforms to intuition in security contexts. Step 5), shown in Figure 1, estimates consequences of each adverse action in each software component using approaches similar to HAZOP or FMEA. In this paper we focus on probability estimation. Thus, we apply a typical severity analysis technique, categorize and quantify component consequences according to security standards [9]. By using the maximum severity obtained in each component, we can compute an estimate of security risks of each component of the system being design. For a risk estimate of an overall system, we estimate the likelihood of a system by multiplying the likelihood of a system to safely terminate by a vulnerability ratio of overall system components, and the system severity. We estimate the system severity by taking a majority of severity categories of components in the overall system. Section 4 illustrates our methodology in a case study of web application design.
4 A Case Study This section presents an illustration in an industrial engineering and business-to-business (B2B) domain with a case study, adapted from [12], of a web-based material requirements planning system for a manufacturing enterprise. 4.1 Web Application Design A web-based material requirements planning (MRP) system takes customer’s product order through the B2B web front and performs material requirements planning. The goal of the system is to respond quickly and effectively to changing requirements to minimize delay in production and product delivery. The system automatically updates orders and predicts the daily planned orders, which may be overwritten by a manager. The daily planned orders are sent to a job shop that simulates real-time scheduling and production. At the same time, the system determines materials required for the daily
52
R. Hewett
planned orders and sends orders of raw materials to appropriate suppliers. Job shop simulator reports on finished parts/products to the MRP system, where the manager may adjust some initial parameters (e.g., lead time, lot size) for scheduling production plan to meet customer order requirements. A supplier can view material order status and acknowledge the receipt of the order. The web application design consists of seven software components: MRP, MPS (Master Production Schedule), BOM (Bill of Material), JOB (Job shop simulator), CI (Customer Interface), MI (Manager Interface), and SI (Supplier Interface). MRP validates logins, records product orders and determines customer’s billing. MPS schedules appropriate daily production plan from a given product demand (e.g., from customer order) and other input parameters. The details of MPS are beyond the scope of this paper. BOM determines parts required for each product to be produced and sends order for raw materials to suppliers. JOB is a job shop component that simulates real-time scheduling and production in a manufacturing plant. CI, MI and SI facilitate web interfaces for different users. Customers and suppliers are required to login via a web front, while a manager can do remote login to MRP.
Fig. 3. Workflow diagram of a customer use case
To focus our illustration on risk methodology, we simplify a B2B process by assuming that all payment between a customer and a manufacturer, or a manufacturer and a supplier are off-line transactions. We represent the design in an extended workflow diagram as described in Section 3.2. To make it easy to understand our analysis approach, we present the workflow in three use cases, each of which has one scenario. The three use cases are customer orders products, manager adjusts demands/requirements, and supplier acknowledges material order. We assume that the likelihood of occurrence of each scenario is 0.4, 0.2, and 0.4, respectively. Figure 3 shows a workflow diagram of a process to handle customer orders for products. We assume customers have established business relationships with the manufacturing enterprise and therefore MRP is restricted for use with these customers only. Thus, customers and suppliers are required to login via a web front. The workflow starts at
Security and Business Risks from Early Design of Web-Based Systems
53
a dark circle and ends at a white circle. It represents both sequential (by directed link) and concurrent (indicated by double bar) activities among different components. For example, after determining parts required for products production, BOM sends results to job shop to simulate production of product orders and concurrently it places order of raw materials to appropriate suppliers. The workflow diagram is extended to include software components responsible for functions/executions of these activities. The components are shown in oval shapes on the right of the figure. The arrows between components indicate the interactions among them.
Fig. 4. Hardware architecture and deployment plan
Second part of the design is a hardware deployment plan. Figure 4 shows the design of hardware platforms required. Like many web applications, the MRP system has a three-tier architecture. The presentation layer requires a web server, W . The business logic layer contains three application servers: A1 − A3. Finally, a data layer employs four database servers: D1 − D4, each facilitates data retrieval from different databases as summarized at the bottom of the figure. The servers/hosts are connected via LAN with a firewall between the web server, W and the LAN access to the three application servers. The deployment plan annotates each server with software components that rely on its services as appeared in a square box on top of each server. For example, as shown in Figure 4, a manager can login remotely to MRP via MI, both of which run on application server A1. MRP and MI allow him to review production status and adjust product demands. Since MRP is also responsible for login validation, it needs to retrieve password data. Thus, unlike other application servers, A1 needs to be connected with D1. Next section describes the analysis for assessing security risks associated with each of these software components from a given design. 4.2 Estimating Security Risks The first step for the proposed risk methodology is to calculate interaction rates from each scenario of the web application design. As shown in Figure 3, MRP has one interaction to MPS but has a total of four interactions from it (two from MRP to a terminal state, one to CI and one to MPS). Thus, an interaction rate from MRP to MPS is 1/4 as shown in table M1 of Figure 5. M1 represents interaction rates among components obtained from a “customer orders product” scenario, where T is a terminal state in the
54
R. Hewett
Fig. 5. Tables of interaction rates in each scenario
Table 1. Vulnerability Measures
workflow. Similarly, we can obtaintables M2 and M3 representing interaction rates in the manager and supplier scenario, respectively. Step 2) constructs a dependency graph of software components. For each pair of components, we compute a transition rate using the corresponding interaction rates obtained from each scenario (M1 , M2 and M3 ) and a given likelihood of each scenario, as described in Figure 1. In Step 3), determine vulnerabilities of hardware platforms from a given deployment plan shown in Figure 4 by applying a scanning tool (if the required host and network configurations of the system are already in operation), or (otherwise) a table lookup procedure to public vulnerability databases. Table 1(a) shows vulnerability sources and their counts on our application deployment plan. As described earlier, the table lookup for vulnerability sources are based on software and its environments required on each host to provide their services. Next, to find a vulnerability ratio, Vi of component i, we determine host services required by the component. For example, the customer interface, CI requires the web server W for login and ordering products, and also requires query retrievals of product information from the database D2 for browsing. Similar results can be obtained for other components as summarized in Table 1(b). The second to last column shows the number of vulnerabilities associated with each component. This is estimated by summing a total number of vulnerabilities in each of the hosts required in the component. The last column shows vulnerability ratios with a total ratio of one for overall system vulnerability. Figure 6 shows the resulting graph obtained from Step 2) with annotated transition rates and vulnerability ratios (in bold), heuristic estimates of vulnerability of the design components. Here we assume that at the starting point S (a solid circle), the system has no vulnerability, whereas the terminating point T (a white circle) has a vulnerability ratio of the whole system (i.e., one). Step 4 determines likelihoods of attacks to (or security breach in) each component using the approach as described in the example of
Security and Business Risks from Early Design of Web-Based Systems
55
Fig. 6. A dependency graph of software components
Table 2. Severity analysis of the web application design
Figure 2. For example, the chance of CI being attacked (via S → CI) is 1×0.4×0.09 = 0.036. Note that the path to CI from MRP (via S → CI → M RP → CI) is ignored since it produces a lower likelihood and thus, not the worst case as desired. The chance of MRP being attacked is computed as a sum of likelihoods of MRP being reached from CI, MI, SI, and JOB, i.e., 0.06 + 0.02 + 0.04 + 0.01 = 0.14. A summary of component security breach likelihoods is given in the first row of Table 4. In Step 5), we estimate consequences of adverse actions in each component by categorizing consequences into five categories: CT (catastrophic: 0.95), CR (critical: 0.75), MG (marginal: 0.5), and MN (minor: 0.25) [9]. The estimated severity for each component is shown in Table 2. By selecting the maximum severity obtained in each component to be the component severity, and applying standard quantification to it, we can obtain risk of each component as shown in Table 3. Here MRP is the most high-risk component with 34.31%
56
R. Hewett Table 3. Risk analysis results
risk, whereas MI is the lowest with 2.64% risk. The risk of overall system is computed by multiplying the likelihood of reaching the terminating point, T with its vulnerability and a severity of the overall system. Here we use, CR, which is a severity category of a majority of all components. The result in Table 3 shows that the security risk of overall system is about 50%, which can be used as a baseline to compare with alternatives. These risks give relative measures for assisting decision-makings on security options from different designs.
5 Business Risk Analysis Although both safety and security can impact the business success of the organization deploying the software, business risks have different additional characteristics that deserve to be considered separately. Despite many technology advances, business application designers are facing with organizational challenges in customizing the application to fit in with existing business processes and rules in order to ensure effective coordination of work across the enterprise. This section illustrates how the proposed risk estimation framework in Section 3 can be applied to analyzing the business implications from the design. The risk methodology for business risks can be simplified from the methodology shown in Figure 1 by omitting Step 3 to computer system vulnerabilities and eliminating vulnerability factor in Step 4. However, more detailed severity analysis is necessary. In particular, we are interested in analyzing consequences of a business failure. The analysis can help identify counter measures for design improvement. This is accomplished by distinguishing different types data flow terminating points of a given component-based software design. For example, consider the web application design described in Section 4 in our case study. Figure 7 illustrates three possible types of exits: • Quit before enter (q1 ): the customer either did not have login information at hand or had no time, or did not like the early login compulsion (especially for first time visitors). For business perspectives, one counter measure to prevent losing potential customers is by removing the early login requirement. • Quit after product order (q2 ): the customer did not have sufficient credit payment, or changed decisions (e.g., due to products or cost after adding shipment charge). One strategy to counter measure failure at this stage could be to give customers more delivery and shipment options including discounts on large orders. Alternatively, the inventory could be running out of items. In such a case, a counter measure can be marketing research and increasing inventory or supply of high demand.
Security and Business Risks from Early Design of Web-Based Systems
57
Fig. 7. Determining business implications from software data flow design
• Quit after payment or after material order (q3 ): this is a normal exit that meets the business objectives, i.e., to obtain transactions from customer and to allow a manager to order the supply materials required for the products. Consequently, the separation of different types of failures yields a slightly different dependency graph of software component from our previous security analysis. A new dependency graph is shown on the right of Figure 7. The likelihoods of each software component can be estimated in a straightforward manner. For severity analysis, we analyze consequences of failures. Table 4 gives an example of failure consequences of business contexts in our example where q1 is not considered as a failure state. Severity analysis is organization dependent and in practice, quantifying severity can be assessed from experiences (e.g., using expected loss per year). If such data is not available, a common practice is to categorize severity into four standard categories: catastrophic (CT), critical (CR), marginal (MG), and minor (MN), each of which is respectively quantified by 0.95, 0.75, 0.50 and 0.25. For example, as shown in Table 4, quitting after payment in q2 results in loss of one order transaction, which is critical but not as severe as loss of opportunity for numerous new customers as quitting before enter in q1 . The rest of the proposed steps can be applied similarly to obtain business risks fornot meeting business objectives as designed. These are reflected in quitting in q1 and q2 .
Table 4. Severity of different stages of business failures
58
R. Hewett
6 Concluding Remarks This paper attempts to provide a general framework for security risk analysis of web application software design. We present a systematic methodology for automated preliminary risk assessment. Our focus here is not to estimate precise risk measures but to provide a quick and early rough estimate of security risks by means of heuristics based on characteristics of software design and hardware platforms to help locate high-risk components for further rigorous testing. Our risk model is far from complete partly due to inherent limitations in the design phase. Although, we illustrate our approach to web applications, it is general to apply to other software-intensive systems and to extend the framework and the methodology to include additional security elements in the model. The approach has some limitations. Besides obvious requirements on knowledge about system usage scenarios and component-based design, the approach is limited by insufficient data available in the early phases of software life cycle to provide precise estimation of the system security. However, the risk analysis methodology suggested in this paper can be used to estimate security risks at an early stage, and in a systematic way. Finally, the approach does not consider failure (to satisfy security criteria) dependencies between components. Thus, risks of a component connected to attacked components are determined in the same way as those that are not attacked. We plan to extend this framework to address this issue by providing mechanisms that take user roles, access privileges and vulnerability exploits into account. Additional future work includes refinement of explicit severity analysis and incorporation of attack scenarios in the risk models. Acknowledgements. This research is partially supported by the Sheldon Foundation. Thanks to Phongphun Kidsanayothin and Meinhard Peters for their help in the previous version of this paper whose parts were presented at the 3rd Conference of Web Information Systems and Technology.
References 1. Bugtraq (October 2006), www.securityfocus.com/archive/1 2. Nessus Vulnerability Scanner (October 2006), www.nessus.org 3. Barna, P., Frasincar, F., Houben, G.-J.: A workflow-driven design of web information systems. In: ICWE 2006: Proceedings of the 6th international conference on Web engineering, pp. 321–328. ACM, New York (2006) 4. Bleistein, S.J., Cox, K., Verner, J.: Requirements engineering for e-business systems: Integrating jackson problem diagrams with goal modeling and bpm. In: 11th Asia Pacific Software Engineering Conference, Busan, Korea (2004) 5. Cortellessa, V., Appukkutty, K., Guedem, A.R., Elnaggar, R.: Model-based performance risk analysis. IEEE Trans. Softw. Eng. 31(1), 3–20 (2005); Senior Member-Katerina GosevaPopstojanova and Student Member-Ahmed Hassan and Student Member-Walid Abdelmoez and Member-Hany H. Ammar 6. Csertan, G., Pataricza, A., Harang, P., Doban, O., Biros, G., Dancsecz, A., Friedler, F.: BPM based robust E-Business application development (2002) 7. Ginige, A., Murugesan, S.: Web engineering: An introduction. Multimedia 8, 14–18 (2001)
Security and Business Risks from Early Design of Web-Based Systems
59
8. Haimes, Y.Y.: Risk Modeling, Assessment, and Management. Wiley-IEEE (2004) 9. ISO. Risk Management - Vocabulary - Guidelines for Use in Standards. ISO Copyright Office, Geneva (2002) 10. Landoll, D.J.: The Security Risk Assessment Handbook: A Complete Guide for Performing. CRC Press, Boca Raton (2006) 11. Pearl, J.: Graphical models for probabilistic and causal reasoning. In: Handbook of Defeasible Reasoning and Uncertainty Management Systems, vol. 1, pp. 367–389 (1998) 12. Qiang, L., Khong, T.C., San, W.Y., Jianguo, W., Choy, C.: A web-based material requirements planning integrated application. In: EDOC 2001: Proceedings of the 5th IEEE International Conference on Enterprise Distributed Object Computing, Washington, DC, USA, p. 14. IEEE Computer Society Press, Los Alamitos (2001) 13. Russell, N., van der Aalst, W.M.P., ter Hofstede, A.H.M., Wohed, P.: On the suitability of uml 2.0 activity diagrams for business process modelling, pp. 95–104. Australian Computer Society, Inc., Hobart (2006) 14. Shahrokhi, M., Bernard, A.: Risk assessment/prevention in industrial design processes. In: 2004 IEEE International Conference on Systems, Man and Cybernetics, vol. 3, pp. 2592– 2598 (2004) 15. Singh, I., Stearns, B., Johnson, M.: Designing enterprise applications with the J2EE platform, p. 417. Addison-Wesley Longman Publishing Co., Inc., Amsterdam (2002) 16. Stoneburner, G., Goguen, A., Feringa, A.: Risk management guide for information technology systems. NIST Special Publication, pp. 800–830 (2002) 17. van der Walt, C.: Assessing Internet Security Risk, Part 4: Custom Web Applications, securityfocus.com (October 2002) 18. Verdon, D., McGraw, G.: Risk analysis in software design. Security & Privacy Magazine 2, 79–84 (2004) 19. Yacoub, S.M., Cukic, B., Ammar, H.H.: Scenario-based reliability analysis of componentbased software. In: ISSRE 1999: Proceedings of the 10th International Symposium on Software Reliability Engineering, Washington, DC, USA, p. 22. IEEE Computer Society Press, Los Alamitos (1999) 20. Zhang, Y., Zhu, H., Greenwood, S., Huo, Q.: Quality modelling for web-based information systems. In: FTDCS 2001: Proceedings of the 8th IEEE Workshop on Future Trends of Distributed Computing Systems, Washington, DC, USA, p. 41. IEEE Computer Society, Los Alamitos (2001)
Designing Decentralized Service Compositions: Challenges and Solutions Ustun Yildiz and Claude Godart INRIA-LORIA BP 239 Campus Scientifique F-54506 Vandœuvre-l`es-Nancy, France {yildiz,godart}@loria.fr
Abstract. This paper reports a new approach to the decentralized execution of service compositions. The motivation of our work is to provide a systematic solution for the advanced configuration of P2P service interactions when they are involved in a composition. In order to do so, we consider a service composition as a centralized workflow specification and derive corresponding cooperating process in a such way that the dependencies of the centralized specification are implemented as P2P interactions between underlying services that execute the derived processes. More precisely, we present the most important issues of deriving operation and propose corresponding solutions that run counter to naive intuition.
1 Introduction Services oriented architectures are emerging as a new standardized way to design and implement business processes within and across enterprise boundaries. Basic XML-based service technologies (e.g. SOAP [1], WSDL [2]) provide simple but powerful means to model, discover and access software applications over the Web as analogous to wellknown concepts of traditional RPC-like and component-oriented computing. The term composition is used to describe the design and implementation issues of the combination of distinct services into a coherent whole in order to fulfill sophisticated tasks that cannot be fulfilled by a single service. For developing business processes based on service compositions, there are two essential initiatives around the banner of Orchestration and Choreography. Orchestration approach considers executable processes (e.g. WS-BPEL [3] specifications) that specify the dependencies of composed services through a single execution unit as analogous to centralized workflow management [4]. In contrast to orchestration, choreography approach (championed by WS-CDL [5]) is more decentralized in nature. It considers a global declarative composition that captures the interactions in which composed services can involve each other. Global choreographies are used to verify the consistency of composed services that work in a collaborative manner. As in application software modelling, different composition approaches are good for different things. Orchestration based compositions, for example, are quite useful for control and monitoring purposes as they govern compositions from a centralized site, but are less useful for data intensive compositions for scalability limitations. Choreography based compositions are useful in elucidating P2P interactions and parallelism, but are more J. Filipe and J. Cordeiro (Eds.): WEBIST 2007, LNBIP 8, pp. 60–71, 2008. c Springer-Verlag Berlin Heidelberg 2008
Designing Decentralized Service Compositions: Challenges and Solutions
61
difficult to establish and control. The focus of this paper is on the design and implementation of decentralized service compositions that bridge the gap between these two approaches. Intuitively speaking, establish peer-to-peer (P2P) interconnections between composed services is a major requirement in the management of business processes [6] [7]. However, the efficient configuration of service interactions is a difficult problem. With respect to two essential approaches, this difficulty has multiple reasons. Following the orchestration approach, composed services are coordinated from a single execution unit and they are isolated from each other. This prevents their P2P interactions. Following the choreography approach, services can establish P2P interactions. However, their interactions may not be adequate to those required by process designers or other contextual criteria. Decentralized workflow management has been always appealing in the conception and the implementation of sophisticated configurations between process participants [8] [9] [10]. The principal reason that prevents their wide spread use is their architectural assumptions that make them unreasonable in the web context. Following the approach of decentralized workflow management, we developed a new approach to the decentralized execution of service compositions. This approach is based on deriving cooperating processes from a workflow-like centralized specification of a composition. Derived processes are deployed on composed services to implement the dependencies of the centralizes specification as P2P interactions between underlying services. In this paper, we discuss the most important issues of deriving operation. First, we present the process specifications that the deriving operation is intuitive. Next, we consider sophisticated specifications that combine control, data and conversational aspects, most of which, there is no intuitive deriving operation. Furthermore, we present our approach for deriving processes from the latter. The remainder of this paper is organized as follows. Section 2 reviews the approach and the related work. Section 2 reviews the related work and presents the background of our approach. Section 3 introduces presents the formal concepts that we employed. Next, section 4 is the core core of contribution while the last section concludes.
2 Background Figure 1 illustrates the overview of our approach. The left-hand side describes a composition as a centralized process specification. In this kind of implementations, service are isolated from each other. A centralized workflow engine executes the composition specification. Consequently, the messages that correspond to the fulfillment of control and data dependencies of composition flow back and forth to the centralized workflow engine. The left-hand side of the same figure illustrates the execution of the same composition with the derived fragments. In contrast to existing solutions, this approach is reasonable and inexpensive in many ways. Intuitively, a composition specification that includes service interactions can be modeled by an organization and can be easily received and executed by another one even over a different execution engine. This is a secondary outcome of standardization efforts as they provide a common understanding for service compositions. However, the mobility of standard process models or process instances between different organizations and tools has not been handled by a decentralization point of view.
62
U. Yildiz and C. Godart
Service
Service
Service
Service
d6 a
d1
Centralized Workflow Engine
a
a
3
6
a
a
0
1
a
8
d0
9
a
a
10
11
a 2
d2
a 12
a 5
a
a
d4
17
Workflow Engine a
a
13
14
a
a
15
16
a
a
a
18
19
20
7
a
d5
4
d3
Service
Service Service
Service
Service Service
(a)
(b)
Fig. 1. Comparison of centralized and decentralized service compositions
The idea of decentralized execution setting is not new. Issues related to decentralized execution of workflows has been considered as very important in the relevant research literature a decade ago [11]. If we consider previous works that are not relevant to service oriented paradigms and technologies, much of them are conceived to prevent the overloading of a central coordinator component. Their basic proposition is to migrate process instances from an overloaded coordinator to a more available according to some metrics [12] or execute sub-partitions of a centralized specification on distinct coordinators [13]. Although efficient, these approaches do not consider P2P interactions of workflow elements. The implementation of decentralization is relatively new for service oriented applications. A simple but crucial observation is that some of the recent approaches deal with decentralization by using centralized services or additional software layers what can be a strong hypothesis in the Web context rather than reasoning decentralization with decentralized processes [9]. Some recent works provide several different process partitioning techniques similar to ours. For example, [14] focuses on the partitioning of variable assignment activities of a BPEL processes. But, they do not discuss how they deal with the elimination of dead paths (Dead Path Elimination [15]). The works presented in [16] [17] are similar, but limited to the decentralization of control flow. Our work is also relevant to traditional program partitioning for distributed computing. However there are important differences. First, the aim of a decentralized execution is not to parallelize the execution as the business logic of centralized composition must be preserved. From this point of view, decentralized service composition is much more relevant to P2P computing rather than program parallelization. The motivation of our approach highlights the problems of arriving at a definition of an accurate operation that derives cooperating processes to be executed by composed services (i.e. how a control or data dependency of the centralized specification can be reestablished and preserved in a decentralized composition?).
3 Formal Considerations This section describes the formal underpinnings of our approach that requires rigorous semantics on modeled processes. It is important to clarify what we do not intent to
Designing Decentralized Service Compositions: Challenges and Solutions
63
propose a new process language or extends an existing one. The presented concepts are applicable to a variety of process modeling languages and execution engines. A process can be represented as a message passing program modeled by a directed acyclic graph [15]. We assume that the centralized composition is modeled as a workflow. At this point, we assume that workflow models are structured processes [18]. A structured process considers control connectors that abstract the control flow of a process as properly nested split/join pairs that characterize control connectors (OR-split, OR-join, AND-split, AND-join). Definition 1 (Process). A process P is a tuple (A, C, Ec , Ed , D) where A is the set of activities, C is the set of control connectors, Ec ⊆ A × C × A is the set of control edges, Ed ⊆ (A × A) × D is the set of data edges and D is the set of data elements. A process P has a unique start activity, denoted with as , that has no predecessors, and has a unique final activity, denoted with af , with no successors. A control edge from an activity ai to another activity aj means that aj can not be executed until ai has reached a certain execution state (termination by default). A data edge exists between two activities if a data is defined by an activity and referenced by another without interleaving the definition of the same data. For example, a data received from a service si with the execution of an activity ai can be used as the input value of an activity aj that consists of an interaction with a service sj . In this case, there is a data edge between ai and aj . It should be noted that our process modeling does not consider control edges between control connectors. We associate source, target: Ec ∪ Ed → A functions return the target and source activities of control and data edges. With respect to control flow we can also define a partial order, denoted by <, over A, with < ⊆ A × A such that ai < aj means there is a control path from ai to aj . A service oriented process consists of synchronous/asynchronous interactions with services and locally executed activities such as variable assignments. We can define a function ser: A → S that returns a service interacted with an activity. As the dynamic service binding is supported, the return value of function ser can be a real service instance if there is already a bound service and if not it can be another value that identifies the service that is supposed to interacted with that activity. Thus, we can define a process activity as below. Definition 2 (Activity). An activity a is a tuple (in, out, s, type) where in, out ∈ D that are input and output values of a, s ∈ S is the interacting service with S is the set of services, and type is the type of activity with ΣA is the set of activity types. The set of activity types can be defined as follows type ∈ {read, write, write/read, local} where read activities receive data from services, write activities send data to services, write/read activities consists of synchronous interactions that sends a data and waits the reply and finally, local refers to activities that do not consist of service interactions. During the execution of sophisticated processes, some services can be used with conversations. This means that different operations of the same service can be invoked several times with distinct activities. These activities characterize the conversation-based interactions with the same service. The set of conversational activities belonging to a service si is denoted by Asi . For each activity, with respect to its incoming and outgoing control and data edges, we can identify two sets that include
64
U. Yildiz and C. Godart
corresponding source and target activities pointed by these edges. The preset of an activity a is •a = {ai ∈ A|∃d ∈ Ed ∨ ∃c ∈ Ec , (source(d) = ai ∧ target(d) = a) ∧ (source(c) = ai ∧ target(c) = a)}. The postset of an activity a is a• = {ai ∈ A|∃d ∈ Ed P ∨ ∃c ∈ Ec , (source(d) = ai ∧ target(d) = a) ∧ (source(c) = ai ∧ target(c)=a)}.
4 Deriving Cooperating Processes Cooperating processes of a centralized specification can be derived by employing various data structures and algorithms. Our method associates the activities of the centralized specification to the processes of the the services that they refer to. So, the elements of Asi are included in the process Psi that is executed by si . In order to preserve the semantics of the centralized specification with P2P interactions, the activities dispatched in different cooperating processes must be wired. Figure 2 depicts a simple wiring example of two control dependent activities across their corresponding cooperating processes. Activities a1 and a2 consist of the respective invocations of s1 and s2 . In a decentralized composition, a1 and a2 are included respectively in Ps1 and Ps2 that s1 and s2 execute. When Ps1 executes a1 , it must inform Ps2 about the termination of a1 to allow it to execute a2 . Figure 2 illustrates wiring activities that respectively send and receive a corresponding control message. The same example can be considered for a data dependency that interconnects two activities on the same control path. However, this time the exchanged message would contain a process data and not a control messages.
a1:s1
a2:s2
P
a1:s1 Ps1
w as1
r as1
s2(message)
s2(message)
a2:s2
Ps2
Fig. 2. Wiring two activities of a centralized specification between two cooperating processes
4.1 Dealing with the Consistency of Derived Processes Although based on simple principles, the deriving operation can lead to the situations that are not adequate to the consistency of cooperating processes. Figure 3 illustrates an example where a data dependency that interconnects two activities can cause a blocking situation between corresponding cooperating processes. The same figure illustrates also how this situation can be dealt by propagating the state of activities at the source of data edges to processes that must be informed about the state of the latter. Actually,
Designing Decentralized Service Compositions: Challenges and Solutions
65
Fig. 3. Wiring two activities dependent by a data edge where the source activity may not be executed
this technique is an extension of the current Dead Path Elimination (DPE) procedure implemented in current workflow management systems. We extended it towards the decentralized execution in order to prevent blocking situations[19]. Next subsection overviews this procedure. Dead Path Elimination and Insuring Global Soundness of Cooperating Processes. The extension of dead path elimination requires the extension of formal dependencies. For example, which activities must be taken by DPE after the evaluation of transition conditions to outgoing paths from an OR-split point? This requires a set of dependencies between an OR-split activity and activities that are involved in DPE. We call this as a extended postset denoted ai •. Similarly, for an activity that can be taken by several DPE, the source activities that trigger DPE must be considered. We gather them within the extended preset denoted •ai . Figure 4 illustrates the extended postset activities of activity a1 . The process that executes a1 informs directly the corresponding processes of activities that are on the paths taken by DPE. In the same time, the status of activities that are on the paths taken by DPE are propagated to processes that expect their outcome. If a2 is not an OR-split activity and its outgoing control edges evaluates to true, its process sends true control messages to the processes of the control dependent postset activities. As analogous to the elements of an extended postset, the extended preset of an activity includes the OR-split points that DPE that takes it or the source of its incoming data edges can be triggered. Figure 5 illustrates the extended preset activities of activity a19 . 4.2 Putting the Deriving Operation into Practice Our method associates the activities of the centralized specification to the derived processes of the services that they refer to. So, the elements of Asi are included in the process Psi that is executed by service si . In order to preserve the semantics of the centralized specification with P2P interactions,the activities dispatched in different processes must be wired. The above examples illustrated the dependencies that must be
66
U. Yildiz and C. Godart
d6 d1 a3
a8
a9
a10
a11
a17
a6
d0 a0
a1
d4 a2
d2
a13
a14
a15
a16
a12
a18
a5
a19
a20
a19
a20
a7 a4
d5 d3
Fig. 4. The elements of the extended postset of a1
d6 d1 a3
a8
a9
a10
a11
a17
a6
d0 a0
a1
d4 a2
d2
a13
a14
a15
a16
a12
a18
a5 a7 a4
d5 d3
Fig. 5. The extended preset activities of a19
considered while activities are wired. Now, we detail the deriving of cooperating processes from the centralized specification and the extended dependencies that presented in the previous section. The wiring operation is considered separately for preset and postset dependencies. 4.3 Wiring with Postset Activities Here we describe how a process behaves after the execution of one of its activities to wire it with its dependent activities. The structure a i • describes a process fragment that a process executes following the execution of its activity ai . ai can be an activity that precedes a OR-split point or not. If so, it is the starting point of DPE. In our approach to decentralized execution, DPE is achieved with P2P interactions. The process that executes ai informs directly the corresponding processes of activities that are on the paths taken
Designing Decentralized Service Compositions: Challenges and Solutions
67
Algorithm 1. Wiring with postset activities Require: ai , Centralized Process Ensure: ai • 1: for all aj ∈ai,c • do 2: for all an ∈ A − {ai } s.t. (an < ai ) ∧ (aj < an ) do 3: for all ak ∈ •an,c do OR/false || 4: a ← aw i→n (ak (f )) j 5: end for OR/false || ← aOR/false 6: a
j
7: 8: 9: 10: 11: 12: 13: 14: 15: 16: 17:
j
OR/false empty(a ) j end for for all an ∈ A s.t. ∃ak ∈ •an,ws , aj < ak ∧ ak < ai do for all ak ∈ •an,ws s.t. (aj < ak ) ∧ (ak < ai ) do OR/false || a ← aw (a (f ))
i→n
j
k
end for end for for all an ∈ A s.t. ∃ak ∈ •an,ws , cc ∈ {OR-split}, (aj < ak ) ∧ (ak < ai ) ∧ (cc < ak ) ∧ (ai < cc) ∧ (ak < cc) do OR/true || a ← aw (a (t))
i→n
j
k
end for OR/false OR/true ai • ← (aw ) ⊕ (aw ) i→j (ai (f )) + aj i→j (ai (t)) + aj ||
18: end for 19: for all aj ∈ ai,st • ∪ai,wt • do ||
ai • ← aw 20: i→j (d = aj .edge) 21: end for +
22: ai • ← F ai
by DPE. In the same time, the status of activities that are on the paths taken by DPE are propagated to the processes that expect their outcome. If ai is not an OR-split activity and its outgoing control edges evaluates to true, its process sends “true” control messages to the processes of the control dependent extended postset activities. Algorithm 1 defines the part of the structuring operation of a i •. In the notations, ⊕, ||, + denote respectively exclusive, concurrent and sequential append operations to a process fragment. The algorithm operates on each activity of the centralized specification. It should be noted that a i • is executed following the execution of ai . If ai is skipped, because it is on a path taken by DPE, the activities of a i • are skipped in the same process. Informally, the algorithm can be described as follows: For each control dependent activity aj that follows ai , two fragments that correspond to “true” and “false” transiOR/false (line 2-8, and line tion conditions of a are derived. These are characterized as a j
j
OR/true OR/false 9-13) and a (line 14-16). a include activities that send “false” messages to j j processes that execute activities that are on the paths taken by DPE (line 4). Moreover,
68
U. Yildiz and C. Godart
activities that send “false” messages to processes that have incoming data edges from OR/false (line 11). the activities of DPE are included in a j
OR/true a includes activities that send “true” messages to processes that have incoming j data edges from the activities that are not taken by DPE (line 15). Note that, we take nested OR-split/OR-join fragments into account. These two fragments are structured sequentially after the corresponding “true” and “false” control messages that can be sent to aj . The latter are exclusively structured (line 17). The last step of a i • is sending the outputs of ai to the processes that expect them (line 20). An unique activity F ai characterizes an empty activity added at the end of a i •. It is wired with the empty start activities of other activities a∈Asi that follow ai with respect to their partial order. This operation is detailed in the next subsection. We explain the algorithm with examples taken from our previous papers [19][20].
Example: Activities a0 and a11 depicted in figure 4 consist of the invocations of different s1 ’s operations. In Ps1 , after the execution of a0 , one of the paths that go to a1 and a2 can evaluate to “false”. In Ps1 , a0 is followed by two concurrent branches that are followed by two exclusive branches that correspond “false” and “true” con trol messages to be sent to the processes that include a1 and a2 . We can denote a 0• w w w = [(aw 0→1 (a0 (t)) + 1) ⊕ (a0→1 (a0 (f )) + 2)] || [( a0→2 (a0 (t)) + 3) ⊕ (a0→2 (a0 (f )) + F 4)] + aw 0→4 (d1 ) + a0 . The activities that send control messages to the processes that execute the elements of a0,c • are followed either by activities that fulfill DPE or inform processes about the status of activities that are not on the paths taken by DPE. Thus, w w w w 1 = ∅, 2 = [aw 0→4 (a1 (f ))||a0→5 (a4 (f ))|| a0→7 (a1 (f ))||a0→9 (a5 (f )) ||a0→9 (a8 (f ))] + w w w w (a (f ))], 3 = [a (a (t))], 4 = [a (a (f )) || a (a (f ))||a [aw 0→16 5 0→10 2 0→6 2 0→3 2 0→10 (a3 (f )) w w F || aw (a (f ))] + [a (a (f )) || a (a (f ))]. The activity a is connected to 0 0→10 6 0→10 2 0→16 3 S a11 . a11 is the last element in Ps1 . If there were other ai ∈ As1 that a0
Designing Decentralized Service Compositions: Challenges and Solutions
69
Algorithm 2. Wiring with preset activities Require: ai , Centralized Process Ensure: •ai +S
1: •ai ← ai 2: for all aj ∈•ai,c do 3: for all ak ∈ A s.t. ∃cc ∈ {OR-split}, c ∈ Ec , c = (ak , cc) ∧ ai < cc} do OR ⊕ 4: •a ← ar (aj (f ))
k→i
i
5: 6:
end for || OR •ai ← arj→i (aj (t)) ⊕ •a i
OR 7: empty(•a i ) 8: end for 9: for all aj ∈ •ai,ws do 10: for all am ∈ A s.t. ∃cc ∈ {OR-split}, c ∈ Ec , c = (am , cc) ∧ (cc < aj ) ∧ (aj < cc)} do ⊕ r aOR 11: j ← am→i (aj (f )) 12: end for 13: for all an ∈ A s.t. ∃cc, cc ∈ {OR-split}, c ∈ Ec , c = (an , cc) ∧ (cc < aj ) ∧ (aj < cc) ∧ (cc < cc ) ∧ (cc < aj )} do
14: 15: 16:
r r OR aOR j ← aj ⊕ (an→i (aj (t)) + aj→i (d = aj .edge)) end for || OR a+ i ← •aj OR empty(•a ) +
17: 18: end for
j
+ − 19: •ai ← (a+ i + ai ) ⊕ (ai ) 20: for all aj ∈ •ai,ss do +
||
•ai ← arj→i (d = aj .edge) 21: 22: end for
The algorithm can be explained as follows: If ai is on paths that can be taken by DPE, the processes that execute activities that precede OR-split points can send the messages that contain “false” message for all elements of •ai,c (line 2-8). If the immediate precedent activity is executed, the “true” message is received only from the process that executes it (line 6). The activities that collect status of control dependent preset activities are structured exclusively as there is a single process that can send the corresponding message (line 4 and line 6). According to the content of control mes− sages, ai can be executed or skipped. a+ i and ai characterize the beginnings of two exclusive branches that respectively execute or skip ai (line 19). a+ i is followed with + + the execution of fragment a . a includes activities that collect “false” or “true” mesi
i
sages that can come from the processes that can start DPE that the activities of •ai,ws are on (line 9-18). The activities that collect “true” messages are followed by activities i is the collection of data that are sent that read the sent data (line 14). The last step of •a by processes that execute the elements of •ai,ss (line 20-22). It should be noted that the latter do not require the execution of activities that test the status of source activities. a− i that corresponds to the start of the exclusive branch that is executed if ai is taken
70
U. Yildiz and C. Godart
by DPE, is interconnected to F ai . Consequently, this operation enables the following activities of the same process with respect to their partial order. Example: Suppose that a4 and a16 depicted in figure 5 consist of the invocations of different s2 ’s operations. In Ps2 , a4 is preceded by two OR-split points which means that its path can evaluate to “false” at these two different points. Consequently, the processes that execute a0 or a1 can send “false” control message for the control dependent preset activities of a4 . If the path of a4 is not taken by DPE, the process of a1 sends “true” control message to Ps2 . After the reception of control messages, a4 can be executed or skipped. If it is executed, its inputs must be collected from processes that execute source activities. Thus, a+ 4 is followed with the execution of an activity that receives the data sent by the process that executes a0 . As •a4,ws is empty, there is no need to test the source of incoming edges. We can denote • a4 − r = S a4 + [ar0→4 (a1 (f )) ⊕ ar1→4 (a1 (f )) ⊕ ar1→4 (a1 (t))] + [(a+ + a (d )) ⊕ a 1 04 4 4 ]. S The activity a− is interconnected to a . For a , there are four preset activities 16 16 4 (a3 , a5 , a10 , a11 ) that provide inputs to a16 . Two of them may not be executed (a3 , a5 ). Consequently, the “false” status of these activities can be sent to Ps2 by the processes that can start DPE. The “true” status message can be sent only by the process that executes the first OR-split point that precedes the source activity of data dependence. Activities a10 and a11 are strong source activities. Consequently, there is no necessity to test their status. The activities that collect d4 and d5 are executed beS r fore executing a16 . We can denote •a 16 as follows •a 16 = a16 + a15→16 (a15 (t)) + r r r r [(a0→16 (a3 (f )) ⊕ a2→16 (a3 (f )) ⊕ (a2→16 (a3 (t)) + a3→16 (d3 )))||(ar0→16 (a5 (f )) ⊕ ar1→16 (a5 (f )) ⊕ (ar1→16 (a5 (t)) + ar5→16 (d2 )))] + [ar10→16 (d5 )||ar11→16 (d4 )].
5 Conclusions This paper has taken on the challenge of designing a technique for translating a centralized composition specification to a set of dynamically deployable cooperating processes. The proposed approach aims to allow orchestrated services to have P2P interactions through the processes that they execute. When coupled with an efficient process transformation technique, we believe that the common process representation brought forward by standards can deal with architectural concerns of decentralized execution settings without making unreasonable assumptions about composed services. More precisely, we presented the most important issues of deriving operation and propose corresponding solutions that run counter to naive intuition. Our contribution stands between orchestration and choreography initiatives that govern the implementation of service oriented applications. We demonstrated how an orchestration description can be effective with a decentralized execution as analogous to choreography that requires much more sophisticated design and run-time reasoning.
References 1. Box, D., Ehnebuske, D., Kakivaya, G., Layman, A., Mendelsohn, N., Nielsen, H.: Simple Object Access Protocol (SOAP) 1.1 (2000), http://www.w3.org/TR/SOAP 2. Christensen, E., Curbera, F., Meredith, G., Weerawarana, S.: Web Services Description Language (WSDL) 1.1. W3C. 1.1 edn. (2001), http://www.w3c.org/TR/wsdl
Designing Decentralized Service Compositions: Challenges and Solutions
71
3. IBM, BEA Systems, Microsoft, SAP AG, Siebel Systems: Business Process Execution Language for Web Services, version 1.1 (updated 01 February 2005) (2005), http:// www-128.ibm.com/developerworks/library/specification/ws-bpel/ 4. Georgakopoulos, D., Hornick, M.F., Sheth, A.P.: An overview of workflow management: From process modeling to workflow automation infrastructure. Distributed and Parallel Databases 3, 119–153 (1995) 5. Kavantzas, N., et al.: Web Services Choreography Description Language Version 1.0 (2005) 6. Yildiz, U., Godart, C.: Enhancing secured service interoperability with decentralized orchestration. In: Proceedings of the 23rd International Conference on Data Engineering Workshops, ICDE Workshops, pp. 725–733. IEEE Computer Society Press, Los Alamitos (2007) 7. Yildiz, U., Godart, C.: Information flow control with decentralized service compositions. In: Proceedings of the 5th IEEE International Conference on Web Services, ICWS, pp. 7–19. IEEE Computer Society Press, Los Alamitos (2007) 8. Alonso, G., Kamath, M., Agrawal, D., Abbadi, A.E., Gunthor, R., Mohan, C.: Exotica/ fmqm: A persistent messagebased architecture for distributed workflow management. In: IFIP Working Conference on Information System Development for Decentralised Organisations, Trondheim (1995) 9. Schuler, C., Turker, C., Schek, H.J., Weber, R., Schuldt, H.: Peer-to-peer execution of (transactional) processes. International Journal of Cooperative Information Systems 14, 377–405 (2005) 10. Atluri, V., Chun, S.A., Mazzoleni, P.: A chinese wall security model for decentralized workflow systems. In: ACM Conference on Computer and Communications Security, pp. 48–57 (2001) 11. Sheth, A., Kochut, K.: Workflow applications to research agenda: Scalable and dynamic work coordination and collaboration systems. In: NATO Advanced Study Institute on Workflow Management Systems and Interoperability, Istanbul, T¨urkiye (1997) 12. Bauer, T., Reichert, M., Dadam, P.: Intra-subnet load balancing in distributed workflow management systems. Int. J. Cooperative Inf. Syst. 12, 295–324 (2003) 13. Wodtke, D., Weißenfels, J., Weikum, G., Dittrich, A.K., Muth, P.: The mentor workbench for enterprise-wide workflow management. In: SIGMOD Conference, pp. 576–579 (1997) 14. Chafle, G.B., Chandra, S., Mann, V., Nanda, M.G.: Decentralized orchestration of composite web services. In: WWW Alt. 2004, pp. 134–143 (2004) 15. Leymann, F., Roller, D.: Production Workflow - Concepts and Techniques. PTR Prentice Hall, Englewood Cliffs (2000) 16. Baresi, L., Maurino, A., Modafferi, S.: Workflow partitioning in mobile information systems. In: MOBIS, pp. 93–106 (2004) 17. Sadiq, W., Sadiq, S., Schulz, K.: Model driven distribution of collaborative business processes. In: IEEE International Conference on Services Computing, SCC (2006) 18. Kiepuszewski, B., ter Hofstede, A.H.M., Bussler, C.: On structured workflow modelling. In: Wangler, B., Bergman, L.D. (eds.) CAiSE 2000. LNCS, vol. 1789, pp. 431–445. Springer, Heidelberg (2000) 19. Yildiz, U.: On Dead Path Elimination in Decentralized Process Executions (Research Report RR-6131, INRIA (February 2007), http://hal.inria.fr/inria-00132928 20. Yildiz, U., Godart, C.: Centralized versus decentralized conversation-based orchestrations. In: CEC/EEE, pp. 289–296 (2007)
Grid Infrastructure Architecture: A Modular Approach from CoreGRID Augusto Ciuffoletti1 , Antonio Congiusta2 , Gracjan Jankowski3 , Michał Jankowski3 , Ondrej Krajiˇcek4 , and Norbert Meyer3 1
3
INFN/CNAF, Bologna, Italy [email protected] 2 DEIS, Universit`a della Calabria, Italy [email protected] Computer and Automation Research Institute, Poznan, Hungary {gracjan,jankowsk,norbert}@man.poznan.pl 4 Masaryk University, Czech Republic, Brno [email protected]
Abstract. The European Union CoreGRID project aims at encouraging collaboration among european research institutes. One target of such project is the design of an innovative Grid Infrastructure architecture, specifically addressing two challenging aspects of such entity: scalability and security. This paper outlines the results of such activity, ideally extending the content of the official deliverable document. Keywords: Grid Computing, Grid Information Service, Workflow Management, Checkpointing, Network Monitoring.
1 Introduction According to [8], a Grid is a complex architecture consisting of a collection of resources, which are made available at user level through a number of services. Such definition opens the way to a number of functional components, whose definition is of paramount importance for the design of a Grid: their semantics anticipate the capability of a Grid to make an efficient use of the resources it contains, to offer differentiated levels of quality of service, and, in essence, to meet user needs. Given the complexity and importance of such infrastructure, its design should address modularity as a primary feature: services provided by the Grid infrastructure should be precisely defined as for their interface and semantics, and form an integrated architecture which is a framework for their implementation. Modularity makes viable the independent evolution of each component, and allows the customization of the overall infrastructure. In order to guarantee interoperability among components, standard interfaces are not sufficient. In fact, the capabilities of a certain functional component should be well understood, and agreed in the community that develops other interoperating services: typical requirements address resource access, workflow management, and security. Such semantics should be compatible with the expected needs of the user, be it a human or a Grid-aware application. J. Filipe and J. Cordeiro (Eds.): WEBIST 2007, LNBIP 8, pp. 72–84, 2008. c Springer-Verlag Berlin Heidelberg 2008
Grid Infrastructure Architecture: A Modular Approach from CoreGRID Streams
Fig. 1. Integration between the functional components of our framework. Each component is a distributed entity that contributes to resource management exchanging descriptors with other components. Persistent information flows are encapsulated into streams, represented by session descriptors).
In addition, past experiences [12] prove that there is a tradeoff between portability and reuse of legacy tools: when functionalities that were not designed for integration are included into an existing project, the whole project tends to inherit all portability problems of the legacy parts. A plugin oriented approach does not solve the problem, but tends to complicate the design, and may even restrict portability. Taking into account such problems, we indicate a wrapper oriented approach: legacy tools are not directly included in the design, but accessible through interfaces that comply with portability requirements of the hosting environment. The agent that implements such functionality (the “wrapper”) is in charge of publishing portability issues that characterize the specific resource. One key issue in the design of a Grid environment is the technology used to support the Grid Information System (GIS). It is more and more evident that a unique technology (for instance, a relational database) cannot satisfy all needs, and may exhibit real scalability limits in case of take off of the Grid technology [3]. Here we propose a differentiated strategy for such vital component, splitting its functionality into a directory service, and a streaming support. The monitoring infrastructure provides input to the GIS: we describe such infrastructure decomposed into resource and middleware monitoring, workflow monitoring and network monitoring.
74
A. Ciuffoletti et al.
Another key aspect of a Grid infrastructure is job submission. According to the GGF guidelines in [13], we consider a unique component that performs batch submissions, scheduling and local queuing, workload monitoring and control. However, such component needs support for checkpointing and accounting, two activities that appear to require capabilities that need to be addressed specifically. We introduce two components that implement such functionalities. The resulting Grid infrastructure should address both the need of e-science applications, mostly oriented to storage and computation intensive applications with moderate scalability, and emerging industrial applications, where the demand is variegated and includes the management of a large number of small jobs: in this perspective, flexibility is mandatory to allow customization. Since we want to follow a clean design strategy, we address interoperation and integration issues since the early steps, using the GIS as a backbone. As a consequence, the adoption of a programming style and tools that support polymorphism is mandatory: the “wrapper oriented” approach indicated above helps on this way. In Section 2 we indentify the functional components, and in section 3 we consider a GIS which provides an integration backbone. In figure 1 we depict a schematic view of our proposal.
2 Functional Components of a Framework Architecture The focus of a Grid infrastructure is on resource management: the goal is to compose the operation of basic services into higher level tasks. To this purpose, the Grid infrastructure accepts and processes task descriptions that articulate a stepwise composition of computing activities. The use of appropriate basic services, whose availability is constantly monitored by a Resource Monitoring component, is scheduled after unfolding the dependencies between atomic computational tasks. Resource scheduling extends not only in the name space, to determine which resource is to be used, but also in time, describing when a certain resource will be busy on a certain task. The operation of assembling resources in order to perform a complex task is associated to the Workflow Analyzer component, whose role is to accept the operational description of a complex task, to manage, and to monitor its unfolding. The unfolding of a workflow must be sufficiently flexible, in order to cope with unanticipated events that may affect resources, either improving or degrading their performance. The appropriate way to cope with such events is the logistic re-organization of workflow execution, which usually entails the displacement of stateful computations, by re-instantiating services whose state corresponds to an intermediate computational step. Two basic functionalities are offered: the registration of a snapshot of an intermediate state of a service, and the re-instantiation of the same service with the given intermediate state. All resources in a workflow participate to such reorganization, and the resulting workflow execution must be consistent with the expected semantics. The Checkpoint Manager component is in charge of supporting the logistic re-organization of a workflow, preserving the relevant state information of a component services in preparation for the reconfiguration of the supporting low level services. Specific checkpointing indications are inserted in the operational description provided to the Workflow Analyzer.
Grid Infrastructure Architecture: A Modular Approach from CoreGRID
75
Since resources are similar to goods, their sharing must be controlled accordingly, taking into account property and commercial value. In that sense, the Grid infrastructure provides identities to Grid users, and defines service semantics according to the identity of the user, thus enforcing individual property. Using the same tools, the usage of a certain service is quantified, and a commercial value associated with it. The User and Account Management component is appointed with such aspects. The whole Grid infrastructure hinges upon the Grid Information System (GIS), which supports the integration between the parts of this distributed entity. From an abstract point of view, the content of the Grid Information System represents the state of the Grid, and is therefore dynamic. However, while some data remains constant for long periods of time, other are updated frequently, for instance when such information represents the residual (or preemptable) share of a resource. The activity of a component is pervasive, and many distinct agents contribute to its implementation: for instance, each site can provide a Workflow Analyzer agent in charge of accepting user requests. Such approach fits naturally with security requirements, which are based on mutual identification among agents. Here we give a summary of the functionalities each component offers, and we outline their internal structures: we use as a reference the work of the partners of the CoreGRID Institute on Grid Information, Resource and Workflow Monitoring. 2.1 Workflow Analyzer The Workflow Analyzer cares about workflows management under several aspects such as mapping, scheduling, and orchestration of workflow tasks against the available, dynamic Grid resources. To such purpose, it has close interaction with the Grid Information System in order to discover and allocate appropriate resources. But, at the same time, it is also a source of information coming from the monitoring of the workflows being executed: most of such information is reused by the Workflow Analyzer itself for adjusting the ongoing executions. A Grid workflow can be specified at different levels of abstraction: in [6] abstract workflows and concrete workflows are distinguished, the difference being whether resources are specified through logical files and logical component names or through specific executables and fully qualified resources or services. According to this approach the workflow definition is decoupled from the underlying Grid configuration. For this reason, the mapping phase (also referred to as matchmaking) is particularly important for selecting the most suitable resources that better satisfy the constraints and requirements specified in the abstract workflow, also with regard to quality of service and workload. The mapping process produces a concrete workflow that is suitable to be executed by a workflow engine, providing for scheduling and execution management capabilities. It is worth observing that, in case of dynamic scheduling, it is possible to re-invoke the mapping process at runtime, in order to modify the concrete workflow instance as a result of relevant events modifying the status of candidate resources. However, it is important to instrument job descriptions before actual execution, in order to ensure that the workflow execution is suitably checkpointed: succinct requirements about workflow recoverability are in the Workflow description provided by the user.
76
A. Ciuffoletti et al.
When the workflow enters the running state, the Workflow Analyzer monitors its advancement, and takes appropriate actions in response to relevant events. During the workflow execution, monitoring is essentially related to the observation of the workflow status. In particular, information about the execution of each single job included in the overall workflow is reported by the monitoring system. Typical information is constituted by services execution status, failures, progress, performance metrics, etc. In case of failure, the workflow execution service itself tries to recover the execution, for example by reassigning the work to a different host in case of host failure. To implement fault tolerance on a more refined extent, it is necessary whenever possible to trigger checkpoint recording, and drive the restart of one or more jobs from the last available checkpoint. The decision whether to checkpoint or restart a workflow is made on the basis of information from Resource monitoring. 2.2 Checkpointing The Checkpointing component is built around the idea of Grid Checkpointing Architecture [9,10], a novel concept that defines Grid embedded agents and associated design patters that allows the integration of a variety of existing and future low-level checkpointing packages. The emphasis has been put to make the GCA able to be integrated with other components and especially with the upper layer management services, for instance the Grid Broker or the Workflow Analizer. The main idea of the GCA boils down to provide a number of Checkpoint Translation Services (CTS) which are treated as drivers to the individual low-level checkpointing packages. The CTSes provide a uniform front-end to the upper layers of the GCA, and are customized to the underlying low-level checkpointers. When the CTS is deployed on a Computing Resource, the GCA has to be informed about its existence. To fulfill this requirement each CTS is registered at component named Grid Checkpointing Service (GCS), and the GCS further exports the information about the available CTSes and related properties to the GIS, so that the GIS becomes the mechanism that connects the GCA with the external Grid environment. Additionally, from the point of view of the Workflow Analyzer, the GCS is the gateway to which a checkpoint request has to be sent. When the GCS receives the request of taking the checkpoint of a given job, it forwards the request to the appropriate CTS. The GCS is able to find the adequate CTS using the information that the execute-job-wrapper registers when a checkpointable job is started. The execute-job wrapper is a special program provided together with an associated CTS. The component that is in charge of submitting the user’s job to the given Execution Manager replaces the actual job with the adequate execute-job wrapper and passes to the second the original job and the global identifier assigned to this job. Which execute-job wrapper is to be used depends on which CTS has been matched to the job’s requirements, according with GIS records. When the execute-job wrapper is started it appropriately configures the GCA environment and finally with help of exec() syscall replaces itself into the original job. When the Workflow Analyzer decides that a given job is to be recovered, then the job has to be resubmitted in an adequate way: one relevant issue is that the job can be
Grid Infrastructure Architecture: A Modular Approach from CoreGRID
77
resubmitted only to a Computing Element associated with a CTS of the same type that was used to checkpoint the job. When a proper Computing Element is found the job is resubmitted to it, but instead of resubmitting the original job itself the recovery-job-wrapper is resubmitted. The original job, as well as the identifier of the checkpoint that is to be used in the recovery process, are passed to the recovery-job-wrapper as the arguments. The recovery-job wrapper is the counterpart of the execute-job wrapper used for the recovery activity. The recovery-job wrapper starts fetching the image, and the subsequent actions are similar to those performed by the execute-job wrapper. As a last step, the recovery-job wrapper calls the appropriate low-level checkpointing package to recover the job using the image indicated by the calling Workflow Analyzer. The GCA shares the motivations of the Grid Checkpoint and Recovery working group of the GGF [15], which is to include the checkpointing technology into the Grid environment. However, the GCA focuses mainly on legacy checkpointing packages and, notably those that are not Grid-aware, while the GridCPR “is defining a user-level API and associated layer of services that will permit checkpointed jobs to be recovered and continued on the same or on remote Grid resources”. Therefore, while GridCPR works on a specification that future Grid applications will have to adhere to in order to make them checkpointable, our effort is towards the integration of existing products into a complex framework. 2.3 User and Account Management The User and Account Management component [7] offers a controlled, secure access to grid resources, complemented with the possibility of gathering data from resource providers in order to log user activity for accounting and auditing purposes. These objectives are realized introducing authorization, ensuring an appropriate level of job isolation and processing logging data. A virtual environment encapsulates jobs of a given user and grants a limited set of privileges. Job activity is bound to a user identity, which is qualified as a member of an organization. The User and Account Management component is a pluggable framework, that allows combining different authorization methods (e.g. gridmap file, banned user list, VO membership based authorization) and different implementations of environments (virtual accounts, virtual machines, and sandboxes). The configuration of the framework is quite flexible and depends on detailed requirements which may vary between the resources, so the administrators may tune local authorization policy to the real needs and abilities. The internal architecture of an agent consists of 3 modules: an authorization module, a virtual environment module and a virtual workspace database. The authorization module performs authentication first (based on existing software, such as Globus GSI). The authorization is done by querying a configurable set of authorization plugins. The virtual environment module is responsible for creation, deletion and communication with virtual environments modeled as Stateful Resources. The module is also pluggable, so it is possible to use different implementations of VE. The database records operations on the virtual environments (time of creation and destruction, users mapped to the environment, etc.). These records together with the standard system logs and accounting data, provides complete information on user actions and resource usage.
78
A. Ciuffoletti et al.
2.4 Resource Monitoring The information on resources and accompanying middleware is provided by the Resource Monitor. Resource Monitor component collects data from various monitoring tools available on the grid. We do not presume any particular monitoring approach, since the current state of the art provides quite wide range of monitoring toolkits and approaches. It is however a difficult task to integrate and process monitoring information from various monitoring tools. Moreover, we cannot assume any scale of the resulting infrastructure thus scalability of the proposed solution, both in terms of amount of monitored resources and required processing throughput for monitoring data, must be emphasized. To achieve the desired level of scalability, with security and flexibility in mind, we propose the design of a Resource Monitor based on the C-GMA [11] monitoring architecture. C-GMA is a direct extension of the GMA [16] specification supported by the Open Grid Forum. The key feature supplied by the C-GMA is the introduction of several metadata layers associated with services, resources and monitoring data. The metadata may specify the data definition language used by the services, the non-functional properties and requirements imposed by the services and resources (such as security and QoS-related requirements) and others. The metadata are used in the matchmaking process implemented by the C-GMA architecture, which is essentially a reasoning on provided metadata about the compatibility of the services and data described by them. When the examined parties are considered compatible, the “proposal” is sent to them to initiate a potential communication. In this way, and with the introduction of various translation components, the C-GMA architecture enables the exchange of monitoring data between various monitoring services. The Resource Monitor service leverages this functionality by connecting to the C-GMA monitoring architecture and using translation services for various monitoring toolkits it collects the monitoring data and supplies them to the Grid Information System. Special attention is paid to Network Monitoring, since scalability issues appear as challenging. We have identified one basic agent, the Network Monitoring Element, which is responsible of implementing the Network Monitoring Service [5]. Network Monitoring Elements (NMEs) cooperate in order to implement the Network Monitoring component, using mainly lightweight passive monitoring techniques. The basic semantic object is Network Monitoring Session, which consists in the measurement of certain traffic characteristics between the Domains whose NMEs participate in the session. To improve the scalability of the Network Monitoring Service, the NMEs apply an overlay Domain partition to the network, thus decoupling the intra-domain network infrastructure (under control of the peripheral administration), from the inter-domain infrastructure (meant to be out of control of the peripheral administration). According to the overlay domain partitioning network, monitoring sessions are associated to Resources denoted as Network Elements (NE), corresponding to inter-domain traffic classes. The overlay domain partition is maintained in an internal distributed database, which allows the coordination among Network Monitoring Elements. The management of network monitoring sessions includes the control of periodic sessions, as configured by
Grid Infrastructure Architecture: A Modular Approach from CoreGRID
79
network administrators, and of on-demand sessions dynamically configured by applications, and uses a scalable peer to peer mechanism to diffuse updates.
3 Integration between Functional Components The central idea of the proposed architecture is to convey all the data through the Grid Information Service in order to have a standard interface across the different administrative sites and services (see [1,2] for a similar approach). One relevant feature of a data repository, and of the Grid Information System, is the volatility of its content. At one end we find “write once” data, that are not subject to update operations and have a relatively low volatility. At the other hand we find data that are frequently updated. The functionality associated to the Grid Information System is a mix of both: while certain data, like a Workflow description, fall in the “write-once” category, other kind of data, like resource usage statistics, fall into the category of data that are frequently updated: a solution that devises a common treatment for both kinds of data suffers of a number of inefficiencies, first the lack of scalability. Therefore our first step is to recognize the need of distinct solutions for persistent and for volatile data. One criteria to distinguish the two kinds of data is the length of the time interval during which the information remains unchanged, under normal conditions. Here we assume that a significant threshold is given by the typical job execution time: we consider as persistent those pieces of information that, under normal conditions, remain valid during the execution of a job. We call such informations descriptors: starting from the specifications of the components that compose our framework given in previous sections, we now classify the descriptors that are exchanged among them, and that collectively represent the persistent content of the Grid Information System. Workflow Descriptor. It is acquired from a user interface by the Workflow Analizer component. It has the function of indicating the stepwise organization of a Grid computation. It contains high level indications about the processing requested at each step, as well as dependencies among individual steps. It should be designed in order to hide all unnecessary details, for instance package names or versions, and focus on the functionality (for instance, “fast fourier transform”, or “MPEG4 compression”). During workflow execution, such structure is used by the Workflow Analizer component in order to monitor workflow execution. Job Descriptor. It is produced by the Workflow Analizer component, and fed to various other components: it is used by the Checkpointing component in order to prepare the execution environment with checkpointing facilities, and by the User and Account Management component in order to associate an appropriate environment to its execution. The Job description is used by the Workflow Analizer component in order to instruct resources about their activity, and during workflow execution, to monitor workflow advancement. Checkpoint Image Descriptor. It is produced by the Checkpointing component (in case of the GCA, this descriptor is produced by the CTS) upon recording a new
80
A. Ciuffoletti et al.
checkpoint. The descriptor contains the bookkeeping data regarding the newly created image. The data can be used by the Workflow Analizer in order to find the identifier of the image that is to be used in order to perform recovery and migration. The GCA itself, basing on the descriptor, is able to fetch the image to the node on which the given job is to be recovered. Checkpoint Provider Descriptor. It is produced by the Checkpointing component. The descriptor advertises the location of service that provides access and unified interface to a particular low-level checkpointing package. The Workflow Analizer uses such descriptior to find the node that provides the desired checkpointing package, as specified in job descriptor. Upon recovery, the descriptor allows finding the nodes offering the same package used for checkpointing. Session Descriptor. It is produced by a generic component, and supports the exchange of volatile data, as described below. User Descriptor. It is produced and used by the User and Account Management component. It contains a description of a user, like its name, institution, reachability, role, as well as security related data, like public keys. The Workflow Analysis component uses such data to enforce access restrictions when scheduling a Workflow. Environment Descriptor. It is produced and used by the User and Account Management component. It contains references to he descriptions of the resources associated to a given processing environment, as well as the access modes for such resources. This may correspond, for instance, to what is needed to run a specific kind of job, and to the identities of the users that are allowed to operate within such environment. The Workflow Analysis component uses such data in order to process a workflow description. Resource Descriptor. It represents usual resource descriptions, including storage, processing, network and network monitoring elements. The identification of a resource includes its network monitoring domain. The Workload Analyzer uses such descriptions in order to schedule job execution, and allocate checkpoint storage. The management of descriptors relies on a directory-like structure. Such structure cannot be concentrated in replicated servers, but distributed in the whole system based on local needs. Functional components that need to have access to such data should address a proxy, which makes available the requested information, or add/delete a descriptor. An LDAP directory provides a first approximation of such entity: however, descriptors are not organized hierarchically. A better alternative is an adaptive caching of those descriptors that are considered locally relevant: for instance, the descriptor of a monitoring session might be cached in a GIS proxy near the monitored resource. Descriptors are diffused in the system using a low footprint, low performance broadcast protocol, and cached wherever needed. The volatile data is represented by data that change during the execution of a job: a typical example is the workload of a computing element. Such data are produced by one of the components described in the previous section, and made available to a restricted number of other components. The storage into a globally accessible facility, included a distributed relational database, seems inappropriate since the information is usually transferred from a source to a limited number of destinations. The concept that is usually applied to solve such kind of problems is the multicast.
Grid Infrastructure Architecture: A Modular Approach from CoreGRID
81
A multicast facility appropriate for diffusing volatile data of a Grid Information System has many points in common with a Voice over IP infrastructure: the container of the communication is similar to a Session (as defined in the SIP protocol). In contrast with a typical VoIP application, the data trasfer within a session in mainly uni-directional and requires a low bandwidth with moderate real time requirements: we call streams the information flows associated to the trasport of volatile data within a Grid Information System. All of the components outlined in section 2 are able to initiate or accept a session with another component: security issues are coped with using the descriptors associated with the agents. E.g., a Resource will accept a call only from a Workflow analyzer that submitted a job. Here we outline some of the relevant streams: Resource usage Stream. It is originated by a resource, like a Storage Element, and summarizes the performance of the resource, as well as the available share of it. Typical callers are the Workflow Analyzer, either during the resource selection or the execution phase. Workflow Advancement Stream. It is originated by a Workflow Analyzer component, and reports the caller about the workflow advancement. Typical callers are user oriented interfaces. One characteristic of a session, that makes it not interchangeable with a directory service, is that the establishment of a session has a relevant cost, which is amortized only if the session persists for a significant time interval. For this reason we include sessions in the number of entities that have a descriptor recorded in the Grid Information Service. Such descriptor advertizes the existence of a given session: it is a task of the callee to create and make available an appropriate Session descriptor, as outlined above. Sessions can be activated on demand, or be permanently available: such option depends on the balance between the workload needed to activate a new session on demand, and of keeping it warm for connection. E.g., Network Monitoring sessions will be mostly activated on demand, while Storage usage statistics can be maintained permanently active.
4 Comparison with Other Works The architecture we propose takes into account the goals and achievements of a number of scientific, as well as industrial projects that accepted the challenges proposed by the design of an effective grid infrastructure. One outstanding project which is being developed to meet the requirements the scientific community is gLite: it is developed within the European EGEE project, the successor of DATAGRID. Its purpose is to capitalize tools and experience matured in the course of DATAGRID, in order to assemble a Grid infrastructure usable for high performance computation, first the LHC experiment on schedule for the next year. We consider gLite [12] as a precious source of experience about a real scale Grid environment. We considered as relevants the inclusion of a number of features that are not considered, or considered at an embrional level, in gLite. Namely, we introduce a specific component that takes into account job checkpointing, we adopt a more powerful
82
A. Ciuffoletti et al.
workflow description language (but gLite is working towards a DRMAA [13] compliant interface), we take into account the task of workflow monitoring under scalability requirements, also considering networking resources, we differentiate the functionality of the GIS into a high latency directory service, and a multicast real-time streaming service. Overall, with respect to gLite, we considered the need for a wide portability: although such problem is not overly relevant for the environment for which gLite has been developed, we considered it relevant in a broader scope. To improve portability we suggest the realization of an integrated framework for the whole infrastructure, hosting legacy components encapsulated in specific wrappers. With respect to implementations based on the DRMAA proposed standard [13] we consider the interactions between Resource Management and Checkpointing, since we observe that the resource management is the component in charge of instructing the resource about activities relevant to recovery and relocation of running jobs. Therefore we describe an interface between a component in charge of managing a transparent management of checkpoints, and another in charge of interpreting user requests. The N1GE by Sun [4] is considered as a relevant representative of the industrial effort towards the implementation of a Grid infrastructure. Such project recognises the problems arising from the adoption of a monolythic relational database, and adheres to the DRMAA standards as for job descriptions. In order to overcome scalability limits imposed by a monolythic databases, it adopts a more flexible commercial database, Berkeley DB [3]. In our proposal we identify the kind of services of interest for our infrastructure, and indicate complementary solutions, that cannot be assimilated to a relational database. This should improve scalability and resource usage. The focus of the GPE [14] prototype by Intel is to bridge users from non-Grid environments, and to provide an interface that will remain sufficiently stable in the future, shielding the user from the changes of a still evolving middleware technology. Therefore the focus is on the provision of a powerful interface that adapts to several kinds of users. In order to take advantage of legacy tools, like UNICORE [17], security issues are delegated to a specific component, the Security Gateway, that enfoces a secure access to sensitive resources. In our view this is a source of problems, since the presence of a bottleneck limits the performance of a system. Instead, we indicate a pervasive attention to security issues, in order to implement appropriate security issues inside each agent. We pay special attention to a Grid resource that is often overlooked: the network infrastructure. Such resource is difficult to represent and to monitor since, unlike other resources, its complexity grows with the square of system size. Yet this resource has a vital role in a distributed system as a whole, since its availability determines its performance, and directly reflects on jobs performance.
5 Conclusions CoreGRID is an European project whose primary goal is to foster collaboration among european organizations towards the definition of an advanced Grid architecture. One of the tasks that contributes to this achievement is targeted at the description of an Integrated Framework for Resource and Workflow Monitoring. In order to enforce integration since the early steps, the research and development activities from several research groups are included in the same container, with frequent and planned meetings.
Grid Infrastructure Architecture: A Modular Approach from CoreGRID
83
This paper presents an early result on this way, after two years from the beginning of the project. We have tried to understand the problems left opened by other similar initiatives, specifically aiming at scalability and security issues, and identified the actors inside our framework. The research groups have produced relevant results for each of them that are only summarized in this paper; instead, we focus on the integration among such actors, based on descriptors advertised in the Grid Information Service. Acknowledgements. This research work is carried out under the FP6 Network of Excellence CoreGRID funded by the European Commission (Contract IST-2002-004265).”
References 1. Aiftimiei, C., Andreozzi, S., Cuscela, G., Bortoli, N.D., Donvito, G., Fantinel, S., Fattibene, E., Misurelli, G., Pierro, A., Rubini, G., Tortone, G.: GridICE: Requirements, architecture and experience of a monitoring tool for grid systems. In: Proceedings of the International Conference on Computing in High Energy and Nuclear Physics (CHEP 2006), Mumbai India (2006) 2. Andreozzi, S., De Bortoli, N., Fantinel, S., Ghiselli, A., Rubini, G., Tortone, G., Vistoli, C.: GridICE: a monitoring service for Grid systems. Future Generation Computer Systems Journal 21(4), 559–571 (2005) 3. BerkeleyDB. Diverse needs, database choices. Technical report, Sleepycat Software Inc. 4. Bulhes, P.T., Byun, C., Castrapel, R., Hassaine, O.: N1 grid engine 6 – features and capabilities. Technical report, SUPerG (2004) 5. Ciuffoletti, A., Polychronakis, M.: Architecture of a network monitoring element. Technical Report TR-0033, CoreGRID (2006) 6. Deelman, E., Blythe, J., Gil, Y., Kesselman, C., Mehta, G., Vahi, K., Blackburn, K., Lazzarini, A., Arbree, A., Cavanaugh, R., Koranda, S.: Mapping abstract complex workflows onto grid environments. Journal of Grid Computing 1(1), 25–39 (2003) 7. Denemark, J., Jankowski, M., Matyska, L., Meyer, N., Ruda, M., Wolniewicz, P.: Usermanagement for virtual organizations. Technical Report TR-0012, CoreGRID (2005) 8. Foster, I., Kesselman, C., Nick, J., Tuecke, S.: The physiology of the grid: An open grid services architecture for distributed systems integration (2002) 9. Jankowski, G., Januszewski, R., Mikolajczak, R., Kovacs, J.: Grid checkpointing architecture - a revised proposal. Technical Report TR0036, CoreGRID - Network of Excellence (2006) 10. Jankowski, G., Kovacs, J., Meyer, N., Januszewski, R., Mikolajczak, R.: Towards Checkpointing Grid Architecture. In: Wyrzykowski, R., Dongarra, J., Meyer, N., Wa´sniewski, J. (eds.) PPAM 2005. LNCS, vol. 3911. Springer, Heidelberg (2006) 11. Krajicek, O., Ceccanti, A., Krenek, A., Matyska, L., Ruda, M.: Designing a distributed mediator for the C-GMA monitoring architecture. In: Proc. of the DAPSYS 2006 Conference, Innsbruck (2006) 12. Laure, E., Fisher, S., Frohner, A., Grandi, C., Kunszt, P., Krenek, A., Mulmo, O., Pacini, F., Prelz, F., White, J., Barroso, M., Buncic, P., Hemmer, F., Meglio, A.D., Edlund, A.: Programming the grid with glite. Technical Report EGEE-TR-2006-001, EGEE (2006) 13. Rajic, H., Brobst, R., Chan, W., Ferstl, F., Gardiner, J., Haas, A., Nitzberg, B., Tollefsrud, J.: Distributed resource management application API specification. Technical report, Global Grid Forum (2004), http://www.ggf.org/documents/GWD-R/GFD-R.022.pdf
84
A. Ciuffoletti et al.
14. Ratering, R.: Grid programming environment (GPE) concepts. Technical report, Intel Corporation (2005) 15. Stone, N., Simmel, D., Kielmann, T.: An architecture for grid checkpoint and recovery (gridcpr) services and a gridcpr application programming interface. Technical report, Global Grid Forum draft (2005) 16. Tierney, B., Aydt, R., Gunter, D., Smith, W., Swany, M., Taylor, V., Wolski, R.: A grid monitoring architecture. Technical Report GFD 1.7, Global Grid Forum (2002) 17. UNICORE, UNICORE plus final report. Technical report, BMBF Project UNICORE Plus (2003)
The Privacy Advocate: Assertion of Privacy by Personalised Contracts Michael Maaser, Steffen Ortmann, and Peter Langendörfer IHP GmbH, Im Technologiepark 25, 15236 Frankfurt (Oder), Germany {maaser,ortmann,langendoerfer}@ihp-microelectronics.com
Abstract. Privacy has been a hot topic in research for several years. A lot of different approaches to protect privacy have been proposed recently. Among these there are several tools for negotiation of privacy contracts. In this paper we present our privacy negotiation framework called “The Privacy Advocate”. It consists of three main parts: the policy evaluation unit, the signature unit and the preferences. In addition, our framework supports an interface for negotiation strategies, so that they are independent of the framework. The preferences can be expressed with a combination of P3P and APPEL. The tests we executed using a state of the art PDA clearly indicate that our framework can be used on mobile devices. The completion of a successful negotiation usually takes about 2 sec. including message transfer via an 802.11b wireless link. This involves multiple evaluations of proposals. Each of them is done in less than 250 msec. Keywords: Privacy, negotiation, P3P, privacy contract, user preferences, personalisation, mobile devices.
Privacy negotiation tools should provide service providers and service users with means to express their own requirements. Please note, by the term negotiation we really mean a kind of bargaining about the content of the resulting privacy contract. Unlike other tools or protocols the requirements do not specify a list of alternatives which can be chosen from exclusively. Our approach describes requirements as negotiation intervals by stating their boundaries but support use of discrete values as well. We made both negotiation parties capable to offer counterproposals based on those intervals and former offers. Such intervals are especially useful for measurable data such as charges and position information. In this paper we present our privacy negotiation framework, named Privacy Advocate (PrivAd). PrivAd provides means for service providers as well as users to negotiate individual privacy contracts. Thereby PrivAd is fully interoperable with already existing P3P based infrastructure. Existing tools such as Privacy Bird [7] can effectively communicate with servers using PrivAd as well as PrivAd clients can process static P3P policies from regular web servers. This paper is structured as follows. Section 2 provides a short state of the art. The extensions of P3P that are needed to enable negotiations are briefly discussed in section 3. In the following section we present details of PrivAd and section 5 outlines its implementation. The paper concludes with a summary and an outlook on further research.
2 Related Work There are several privacy related tools that are based on P3P and APPEL specifications. AT&T’s Privacy Bird is a free plug-in for Microsoft Internet Explorer. It allows users to specify privacy preferences regarding how a website collects and stores data about them. When the user visits a website, the Privacy Bird analyzes the provided policy and indicates whether or not the policy matches the user’s preferences. The Microsoft Internet Explorer 6 and Netscape 7 embed a similar behaviour. They allow users to set some options regarding cookies and are capable of displaying privacy policies in human readable format. PAWS also relies on P3P and APPEL but allows discovery of privacy policies within ubiquitous environments [8]. The policy evaluation is not done on the mobile device. All these tools are a valuable step into the right direction, but they still lack means to individualize privacy policies. There are approaches that allow users to choose from a given number of alternatives presented by the service provider [9, 10]. Such behaviour does not characterize a real negotiation to us. These approaches rely on the fact that a non empty intersection of the user requirements and the policies offered by the service provider is given a priori. Such approaches are widely used in protocols such as SSL and TLS. In contrast to this we consider both negotiation opponents as equal and they both are enabled to offer proposals and counterproposals. [11] presents an approach for negotiating privacy policies using negotiation trees. This is similar to our approach but the prototype is not capable of automatic negotiation.
The Privacy Advocate: Assertion of Privacy by Personalised Contracts
87
3 Negotiation Extensions To enable negotiation scenarios the following documents are needed: Service and user preferences that describe the respective requirements, proposals and counterproposals from preferences and finally a mutually agreed privacy contract as the outcome of a successful negotiation. Preferences are secret policies of each party that are never exposed to the opposite side. The secret preferences specify boundaries that span a certain room to negotiate. These boundaries do not only define certain values or enumerations but can specify also single- or double-bounded intervals of real numbers or any other value. Surely unique requirements of users and service provider have to be considered too. In order to enable individualization a user has to state permissions and prohibitions whereas the service providers state requirements and facultative requests. This is done in the secret preferences of each party by respective tags that are added to the statements. Proposals or counterproposals contain complete envisioned contracts or parts of those. Further they may contain requests for deletion of certain parts. Such documents are exchanged alternately during the negotiation process. A privacy contract enumerates all data, purposes, recipients etc. whose release has been agreed on by both opponents during the negotiation. Privacy contracts must not contain any interval. All values have to be stated explicitly. Our vision of a privacy contract is mostly resembled by a P3P policy. Therefore we decided to enhance the P3P standard with different language features to enable negotiation possibilities. Since P3P does not support a description of intervals as needed for the preferences, we added tags to describe those intervals first. This is done by specifying boundaries with tags like and . Boundaries can be set for measurable data i.e. accuracy of location. For remuneration of service usage we integrated a special element to represent the amount of money to be paid for a service. Besides the necessary interval it contains unit and currency tags to enable respective conversion. Due to space limitation we omitted detailed description of all extensions. For further information please refer to [12]. 3.1 Requirements and Facultative Requests For service providers, the secret preference contains the data they want to collect and the purposes, retentions and recipients they want to use the data for. Above all, the concept allows dividing different classes of data items to accentuate the need for them. Thus the data items have to be classified into requirements and facultative requests. Requirements state all data needed to fulfil a service whereas facultative requests specify optional data items, i.e. useful or attractive information that are not essential for service usage. Statements in the secret preference of service provider that are not tagged as facultative requests are classified as requirements automatically. Logically proposals are never acceptable for the service providers until all requirements are fulfilled.
88
M. Maaser, S. Ortmann, and P. Langendörfer
3.2 Permissions and Prohibitions While requirements and facultative request at service provider preferences are selfexplanatory, the permissions and prohibitions in user preferences contain additional implications. Users are able to define statements in their preferences, which are explicitly prohibited or allowed. In this way users are capable of prohibiting single data or groups of data for certain purposes, recipients or retentions. It is also possible to prohibit data or groups of data for all purposes or all recipients etc. That is, prohibitions specify domains that are explicitly prohibited. Preferences that are solely composed of prohibitions, implicitly allow all non-stated data. Similarly, permissions explicitly allow data for certain purposes or recipients. If user preferences contain permissions only, all non-stated data are implicitly prohibited. Obviously, explicitly allowed or prohibited content of statements is more significant than comparable implicit ones. 3.3 Privacy Contracts Successful negotiations finish with a mutually agreed and signed policy, called privacy contract. Digital certificates analogous to those used by SSL are the basis for the used signatures. Since the privacy contract is an individual document regarding each pair of user and service provider it must be digitally signed by both parties. The P3P standard does not provide an opportunity to integrate digital signatures. Thus, to ensure backward compatibility, we use XML comments to integrate the digital signatures. This way we avoid misinterpreted policies by standard P3P applications. Our negotiation protocol is not capable to choose from a list of signature algorithms, analogue to SSL, yet. Thus, both negotiation parties have to use a predefined signature algorithm. Currently this is fixed to SHA/DSA.
4 Privacy Advocate The Privacy Advocate (PrivAd) is a framework that is capable to negotiate on the basis of P3P privacy policies including the introduced extensions. It is designed for both negotiation parties and may be used by service users as well as the service provider side. Hence we implemented a PrivAd client and a PrivAd server. The framework consists of a Policy Evaluation Unit, a Signature Unit, a Repository for preferences and signed policies, mechanisms for http communications and an interface for exchangeable Negotiation Strategies, see Figure 1. The graphical user interface of the PrivAd client allows the user to monitor and control the negotiation process and was implemented for demonstration purposes only. Surely an embedded version of PrivAd does not need such a graphical user interface and works unseen in background. The Signature Unit verifies signatures of incoming documents and signs outgoing ones. The Policy Evaluation Unit is the core of PrivAd and the negotiation concept. Every incoming proposal or policy has to be evaluated and compared with the preferences by this unit. The last component, the Negotiation Strategy, is responsible for calculating counterproposals. It is not a part of the framework. We intentionally made this part exchangeable to provide means of using alternative
The Privacy Advocate: Assertion of Privacy by Personalised Contracts
89
strategies. Exemplary we implemented two strategies described in Section 5.5, but the framework could be completed with any Negotiation Strategy. To create a negotiation enabled antagonist we embedded PrivAd into a Tomcat Web Server. To exchange negotiation messages between two negotiation parties or to collect a policy of a static device, the transport protocol http is used. Our Privacy Negotiation Protocol (PNP) used by PrivAd is a stop-and-go protocol in which each incoming policy triggers its evaluation and sending of an answer. Figure 1 displays the interaction of a PrivAd client with a PrivAd server or a regular P3P enabled web server respectively.
Fig. 1. Structure of the Privacy Advocate
4.1 Negotiation Process This section describes a single turn of the bargaining procedure, when a PrivAd client receives a policy as a proposal or counterproposal from a negotiation enabled service provider using PrivAd server. The Signature Unit verifies the signature of any received policy first. Incorrect signatures lead to stop the negotiation immediately. Policies that contain correct signatures are forwarded to the Policy Evaluation Unit that compares the content of the policy with the client’s preferences. The Policy Evaluation Unit returns a Boolean value that states whether or not the received policy matches the client’s preferences. In case of a match PrivAd has 2 options: 1.
2.
PrivAd signs the policy and sends it back to the service. If the incoming policy was already single signed both sides have a mutually signed contract now and may continue with fulfilling the service. PrivAd orders the Negotiation Strategy to create a counterproposal although the policy was acceptable. The Negotiation Strategy may create a counterproposal that contains more favourable terms and conditions for the client’s side which is then send back.
In case of a mismatch with own preferences, PrivAd has 3 options to response: 1. 2.
PrivAd cancels the negotiation because the policy was unacceptable. PrivAd orders the Negotiation Strategy to create a counterproposal that matches own preferences and returns it to the sender.
90
M. Maaser, S. Ortmann, and P. Langendörfer
3.
PrivAd cancels the negotiation because an integrated assessment function indicates that the negotiation process does not seem to advance and become successful.
Note that the Policy Evaluation Unit and the Negotiation Strategies are described in separate subsections later. The PrivAd server handles incoming proposals analogous. To provide more flexibility, PrivAd uses an independent Negotiation Strategy to create counterproposals. As seen above, PrivAd does not need a strategy to evaluate incoming documents but for appropriate responses only. It is possible that a complete bargaining needs many negotiation rounds. Therefore PrivAd is capable to store and adjust former proposals if necessary. Thus, counterproposals do not need to contain the complete policy. Only changes need to be submitted. If a negotiation party agrees to the proposal with all changes applied, a new contract proposal containing all involved statements is generated. That contract proposal is signed and sent to the opposite side. Finally the case to «negotiate» with regular P3P enabled web servers has to be considered, see section 5.1 for further details. PrivAd sends no document to such server but receives their static P3P privacy policy. This policy is evaluated by the Policy Evaluation Unit as usual. PrivAd client notifies the user about the acceptance of the policy via a message on the screen. Due to the static server no counterproposal is calculated in case of a mismatch with own privacy requirements. So, PrivAd provides the same functionality as existing P3P tools. In the case of especially tenacious strategies it may happen that the negotiation will never find an agreement, despite such agreement theoretically exists. These cases cannot be prevented by design. Thus PrivAd needs to provide an assessment function that allows to monitor progress of the negotiation as well as to calculate, whether a negotiation should be cancelled. The assessment function is independent of the strategies. Our current assessment function accounts the percentage of accepted statements to check the convergence of the negotiation process. If this percentage of the current negotiation round is less than the minimum of last five negotiation rounds, PrivAd cancels the current negotiation. In principle each user may define an own assessment function or adjust the number of negotiation rounds of the currently used function. 4.2 Policy Evaluation This section describes the procedure of checking a proposal against negotiation preferences. It is essential that every incoming document can be checked automatically. Therefore an algorithm and a data-model were developed, which determine the acceptance of a document in relation to own preferences. Surely these documents have to conform to P3P with or without the introduced extensions. As stated before, the two possible results of this check are the acceptance or denial of a proposal. For further explanations the following naming conventions are constituted. Statements of arrived proposals are called paragraphs and statements of preferences are called directives. The policy evaluation passes three different stages starting with the evaluation of individual elements of a statement up to the complete proposal.
The Privacy Advocate: Assertion of Privacy by Personalised Contracts
91
4.2.1 Elements of Statements The paragraphs and directives are interpreted as vectors of sets. These sets are purposes, recipients and data_group. Thus, elements of a statement are represented as follows. statement = ({}purposes,{}recipients,{}data_group) Depending on the use of the statement the sets may also be Ω. Ω specifies all possible members of a set, i.e. all purposes that are imaginable. Ω may be used in the user preferences to prohibit or allow data without referencing a certain purpose or recipient. Hence, it can be assumed that each element of a statement, which works as a paragraph, is neither empty nor Ω. Obviously empty sets or Ω are allowed to be used in directives only. Considering paragraphs and directives as vectors of sets, just these sets have to be compared. That is, we compare purposes from paragraphs with purposes from directives just etc. We call sets from directives with permissions permitted sets and sets from directives with prohibitions prohibited sets. The result of this comparison is one of the following logical values. • `not affected` Æ The compared sets are disjoint. • `definitely permitted` Æ The set of the paragraph is a subset of the permitted set. Hence it is allowed. • `possibly prohibited` Æ The intersection of the paragraph’s set and the prohibited set is not empty but the paragraph’s set contains elements that are not elements of the prohibited one. So final decision whether or not the paragraph can be accepted requires additional calculations, see section 4.2.2. • `possibly permitted` Æ The intersection of the paragraph’s set and the permitted set is not empty but the paragraph’s set contains elements that are not elements of the permitted one. So further processing is required analogous to `possibly prohibited `. • `definitely prohibited` Æ The set of the paragraph is a subset of the prohibited set. Thus it must be prohibited. Table 1. Logical values of sets in paragraphs to the respective sets in directives Directive is
permit
prohibition / {}+
§ ∩ D ≠ Ø, § ∩ D ≠ §, §∩D≠D
POSSIBLY PERMITTED
DEFINITELY PROHIBITED
POSSIBLY PROHIBITED
§∩D=Ø
NOT AFFECTED
NOT AFFECTED
NOT AFFECTED
§=D
DEFINITELY PERMITTED
DEFINITELY PROHIBITED
DEFINITELY PROHIBITED
§⊂D
DEFINITELY PERMITTED
DEFINITELY PROHIBITED
POSSIBLY PROHIBITED
§⊃D
POSSIBLY PERMITTED
DEFINITELY PROHIBITED
DEFINITELY PROHIBITED
D=Ω
DEFINITELY PERMITTED
DEFINITELY PROHIBITED
DEFINITELY PERMITTED
D=Ø
POSSIBLY PERMITTED
DEFINITELY PERMITTED
DEFINITELY PROHIBITED
Ø≠D≠Ω
prohibition / {}&
92
M. Maaser, S. Ortmann, and P. Langendörfer
Each element gets one such value assigned and is further processed in the next steps to finally determine the acceptance of the proposal. For easier notation the examined set from the directive is D and the respective set from the paragraph is §. According to this notation a lookup table, see Table 1, is used to determine logical values. Just see the headline to define the kind of directive and the first column to assign the set of the paragraph to determine the result. For determination of logical values it must be taken into account whether a directive is a prohibition or a permit. Since prohibitions forbid the combination of certain data for any of the stated purposes or recipients, we have to distinguish OR-sets {}+ and ANDsets {}&. While OR-sets prohibit all possible combinations of the contained elements, an AND-set only prohibits the given combination of the set. Thus, the OR-set is the more restrictive one. In permits this classification of sets is not necessary because every subset of them is also permitted. 4.2.2 Grading of Paragraphs After determining the logical values of sets in paragraphs, the next step of evaluation has to be done. The logical values of sets in a paragraph have to be concatenated to determine the logical value of the complete paragraph in relation to the directive. We use two additional lookup tables to define this concatenation operation. One lookup table is responsible for prohibiting directives and the other one for permits. Example: A user prohibits the combination of her data A and B for the purpose P by a prohibiting directive. A service requests data A for the purposes P and Q. According to Table 1 the requested data is `possibly prohibited` and the purposes are `definitely prohibited`. These logical values are concatenated resulting in the value `possibly prohibited`. Hence, the paragraph is acceptable because it requests not all of the prohibited data. A special focus lies on paragraphs, which got the logical values `possibly permitted` or `possibly prohibited`. All paragraphs which have been evaluated to those logical values have to be analyzed again. Please notice that both values are not equivalent. In the following we provide an example for `possibly prohibited` paragraphs. Suppose a directive with prohibited data {name, address, birth date}. The combination of this data could be misused to falsify an eBay account. Additionally assume a proposal containing two paragraphs. One paragraph requests the name and address, the other one requests the birth date. Both paragraphs are evaluated to `possibly prohibited` (refer Table 1) because the requested data_groups are only subsets of the prohibition {name, address, birth date}. So these individual paragraphs would still be acceptable, despite their combination is violating the directive. To avoid such cases, all `possibly prohibited` paragraphs are merged as long as possible in order to determine effectively, whether they are `definitely prohibited` or not. Thus, these two paragraphs have to be merged into one paragraph if possible. Two paragraphs that differ in one set only, can be merged to a single paragraph where the differing sets are conjoint. In our example the merged paragraph contains data {name, address, birth date} after finishing the merge operation and is checked again. Obviously this turns out to be `definitely prohibited`. Consequentially the logical values of both original paragraphs are changed to `definitely prohibited`.
The Privacy Advocate: Assertion of Privacy by Personalised Contracts
93
4.2.3 Aggregation of Acceptance of Proposal After the logical values for all paragraphs of a proposal are determined the results have to undergo another evaluation step. It is checked whether a certain paragraph was evaluated to `definitely prohibited` or `possibly permitted`. If such a paragraph exists in the proposal, the Policy Evaluation Unit signals the non-acceptance of the proposal. Otherwise it could be agreed to. In [13], the used data-model and the policy evaluation are defined in more detail. 4.3 Protocol Issues To exchange negotiation documents we used http as transport protocol. This protocol is widespread and ensures backwards compatibility with P3P. PrivAd uses the httppost method to pick up and send proposals or policies to negotiation enabled servers. If PrivAd detects no authorization to send an http-post request towards the negotiation opponent, a regular P3P enabled web server is assumed. Thus PrivAd will get the static policy, if available, by using the http-get method. That enables PrivAd to verify every existing static P3P policy. Also our PrivAd server responds with a standard policy at receiving an http-get request. That enables the server to answer requests of clients that are unable to negotiate such as most existing P3P tools. If an http-post request is received, the PrivAd server supposes a negotiation enabled client on the other side. Therefore PrivAd client sends an initial request to negotiation enabled server by http-post to initiate the negotiation process. This post request can be empty or contain a proposal.
5 Implementation 5.1 Mobile PrivAd Client The main restriction while implementing PrivAd client was to ensure that mobile devices can use it. Therefore the PrivAd client was implemented compatible to JDK 1.1.8. We use the JVM CrEme [14] to run PrivAd on PDA’s. The PrivAd client PDA version contains a graphical user interface to visualize the negotiation process on the PDA for plausibility checks and demonstration. In this GUI every transferred document and the preferences of the client are shown. The user has to choose a file with preferences and a URL as a target for negotiation. Pushing a button to run the next negotiation round enables the user to control the progress of the negotiation process. All policy evaluation and proposal generation is done by the PrivAd client. Note the code size takes about 150 Kbytes only. The privacy negotiation process was tested on a Toshiba Pocket PC e740 WLAN with a 400 MHz XScale processor and 64 MB RAM. For this test we used simple policies that were compiled with less than 4 statements on the PrivAd server and the mobile PrivAd client. For the air link we used 802.11b with a nominal data rate of 11 Mbit/sec. We measured results for accepted as well as for rejected policies. Comparing the results for different transmission speeds of the network shows clearly that the communication time dominates the time needed to complete the negotiation process. Using a wired connection the complete negotiation cycles were finished in
94
M. Maaser, S. Ortmann, and P. Langendörfer
less than 300 msec whereas it took about 2000 msec using the wireless connection. The pure calculation of proposals and counterproposals takes less than 250 msec on the mobile PrivAd client. 5.2 PrivAd Server The server side is implemented as a Java Servlet that runs in a Servlet container. For our reference application we used Apache Tomcat [15]. Similar to its client the PrivAd server is backward compatible to existing P3P tools. Therefore the server is able to behave like a regular P3P enabled web server, if it is necessary. On a first request the PrivAd server will respond with its initial policy. If it is an http-post request, the PrivAd server initiates the negotiation process with this client. By receiving a simple http-get request the server only replies with a default policy without any negotiation interests which is fully compatible to P3P standard. Hence the PrivAd server works interoperable with already existing tools like Privacy Bird, Internet Explorer or Netscape. Using http-post for policy exchange may inhibit options for invaders. This carries limited risk, because incoming messages are parsed and validated against the PrivAd schema. In worst case there will be an error accounted by parsing the element. Since all policies are interpreted as text and not as executables, the risk, that malicious code can be executed in system, is kept minimal. 5.3 Digital Signature All contracts have to be signed digitally. This will block man-in-the-middle-attacks, give proof of identity and offers more integrity. Incorrect signatures lead to cancellation of the negotiation process. In addition the Signature Unit signs an acceptable policy to generate a single signed contract proposal. Accepted contract proposals from the opponent are signed to generate a mutually signed contract. For digital signing we implemented the SHA/DSA algorithm in the Signature Unit. We rely on existing PKI as SSL/TLS also does. For future versions we intend to integrate selection means for different signature algorithms. 5.4 Negotiation Strategies Creation of a counterproposal depends on the Negotiation Strategy in PrivAd and its secret preferences. Simple Negotiation Strategies, for both negotiation parties are developed and integrated. For demonstration we implemented two strategies where both sides try to find a mutual contract by accommodating as fast as possible. To start we point out operation methods of a simple service provider strategy for PrivAd server. Consequently, the service provider creates and proposes an initial policy, which contains all requirements and facultative requests. Additionally, most profitable values from the intervals in the preferences are chosen in this policy to gather a maximum of information and money. On receiving a counterproposal it is applied to
The Privacy Advocate: Assertion of Privacy by Personalised Contracts
95
previous proposals to get a combined proposal and the provider strategy has to react in a more complex way: • The strategy acquires the logical values of all statements that the Policy Evaluation Unit has determined before. Logically statements that are neither explicitly nor implicitly prohibited do not require a counterproposal. • All acceptable statements in the proposal are mapped to respective statements in the preferences. • If all requirements are satisfied already, the strategy creates a new contract proposal from the combined proposal. All additionally acceptable statements are surely added to gain as many data as possible. The PrivAd server signs it and sends that new contract proposal back to the user. • If not all requirements are satisfied, a new contract proposal or counterproposal has to be created and proposed. Therefore all requirements are requested. • Unacceptable proposed statements are updated by new values matching the intervals in the preferences. Here the strategy may reduce the proposed charge for the provided service by 10 percent or similar. • If proposals instruct to discard statements that are requirements, the provider strategy proposes exactly these statements again since otherwise the service cannot be provided. However, the service provider never signs a policy, where not all minimum requirements from its preferences are included. • The created counterproposal is passed to PrivAd server that sends it back to the opponent. While service providers normally try to retrieve the highest possible amount of data, the users naturally negotiate in the opposite way. • The user gets an initial policy by querying the service provider at first time. • The PrivAd client checks the signature. Proposals are applied to previous ones if those exist. Then the acceptance of the combined proposal is evaluated. Thereby the Policy Evaluation Unit marks all statements in the proposal with logical values to determine prohibited ones. • If the proposed policy does not contain any prohibited statement this policy is signed by the Signature Unit and is returned. If this policy was a contract proposal, it is also signed by the PrivAd client and becomes a mutually signed contract. In that case both sides have a mutually signed contract now and may start using the service. Surely also an acceptable proposal could be counterproposed here. Thereby the strategy may cut the prizes for the service or adapt the proposed values to improved ones. • If it contains prohibited statements the strategy component generates a counterproposal. Prohibited paragraphs are the base for calculating the counterproposal. Below we display an operation breakdown on a single prohibited paragraph. As mentioned before, statements in preferences are called directives whereas statements in proposals are called paragraphs. o First, the prohibiting directives for this paragraph are searched in the client’s preferences. o Special focus is set on merged paragraphs. Certainly all prohibited paragraphs that were merged into one single paragraph have to be
96
M. Maaser, S. Ortmann, and P. Langendörfer
rejected in their isolated form as well. Since all paragraphs are numbered a rejection can be accomplished by returning an empty paragraph with that particular number. o Finally, the prohibited paragraph is intersected with the prohibiting directive. Here intersection means that all values of the paragraph are adjusted to match the intervals in the directive. The strategy may try to approximate the values of the counterproposal to the preferred boundary of the interval. o If the strategy has still not found an appropriate intersection, the prohibited paragraph has to be rejected without a replacement. o This process is repeated for every prohibited paragraph of the unaccepted policy. • Last step is to create a counterproposal containing all changed and newly generated paragraphs. All paragraphs, that were not listed or explicitly rejected in counterproposal are supposed to be accepted. • The created counterproposal is passed to the PrivAd client that sends it to the opponent.
6 Summary and Outlook This article determined the situation in stating and contracting privacy in the Internet. Lacking negotiation capability has been identified as a shortcoming. A privacy negotiation framework that allows mobile users and service providers to negotiate about data and its potential recipients, purpose etc. was presented. This framework PrivAd is downwards compatible with P3P and enables negotiation scenarios between PrivAd clients and static servers as well as between known P3P tools and PrivAd servers. The process of evaluating a privacy policy and the calculation of counterproposals was explained. Tests, done with a state of the art mobile device, prove that our mobile PrivAd client is functional. Successful negotiations, finishing with a mutually signed privacy contract, take about 2 seconds via wireless link and about 0.3 seconds via a wired link, including all message transfer. In the future the two negotiation parties and their types of requirements shall be merged, in order to create tools which can perform both sides of negotiation at a time. This will enable negotiations between enterprises and in the peer-to-peer domain.
References 1. Cranor, L.F.: Beyond Concern: Under-standing Net Users’ Attitudes About Online Privacy. In: Vogelsang, I., Compaine, B.M. (eds.) The Internet Upheaval: Raising Questions, Seeking Answers in Communications Policy, pp. 47–70. The MIT Press, Cambridge (2000) 2. Cranor, L.F., Langheinrich, M., Marchiori, M.: The Platform for Privacy Preferences 1.1 (P3P1.1) Specification (2006), http://www.w3.org/TR/P3P11/ 3. Cranor, L.F., Langheinrich, M.: A P3P Preference Exchange Language 1.0 (2002), http://www.w3.org/TR/P3P-preferences/
The Privacy Advocate: Assertion of Privacy by Personalised Contracts
97
4. Thibadeau, R.: A Critique of P3P: Privacy on the Web. The eCommerce Institute, School of Computer Science, Carnegie Mellon University (2000) 5. Bennicke, M., Langendörfer, P.: Towards Automatic Negotiation of Privacy Contracts for Internet Services. In: Proceedings of 11th IEEE Conference on Computer Networks. IEEE Society Press, Los Alamitos (2003) 6. Druecke, R.: Attitudes to Privacy at using mobile phones (in german), Technical Report, Mobile Internet Business, Nr. 3 (2006) ISSN 1861-3926 7. Privacy Bird, AT&T Corporation (2006), http://privacybird.com 8. Langheinrich, M.: A Privacy Awareness System for Ubiquitous Computing Environments. In: Borriello, G., Holmquist, L.E. (eds.) UbiComp 2002. LNCS, vol. 2498. Springer, Heidelberg (2002) 9. Preibusch, S.: Implementing Privacy Negotiation Techniques in E-Commerce. In: 7th IEEE International Conference on E-Commerce Technology, IEEE CEC 2005, Technical University Munich, Germany (2005) 10. El-Khatib, K.: A Privacy Negotiation Protocol for Web Services. In: Workshop on Collaboration Agents; Autonomous Agents for Collaborative Environments Halifax, Nova Scotia, Canada (2003) 11. Yee, G., Korba, L.: Privacy Policies and their Negotiation in Distance Education. In: Instructional Technologies: Cognitive Aspects of Online Programs. Idea Group Inc. NRC 46555 (2004) 12. Maaser, M., Ortmann, S., Langendoerfer, P.: NEPP: Negotiation Enhancements for Privacy Policies. In: W3C Workshop on Languages for Privacy Policy Negotiation and Semantics-Driven Enforcement, Ispra, Italy (2006) 13. Maaser, M., Langendoerfer, P.: Automated Negotiation of Privacy Contracts. In: 29th Annual International Computer Software and Applications Conference (COMPSAC 2005), Edinburgh, Scotland, UK. IEEE Computer Society, Los Alamitos (2005) 14. CrEme, version 3.25, NSICOM (2004), http://www.nsicom.com 15. Tomcat, version 5.5, Apache Software Foundation (2005), http://www.tomcat. apache.org
Efficient Queries on XML Data through Partitioning Olli Luoma Nice Business Solutions Finland Oy Lemmink¨aisenkatu 32 B, P.O. Box 115, 20521 Turku, Finland [email protected]
Abstract. The query evaluation algorithms of practically all XML management systems are based on structural joins, i.e., operations which determine all occurrences of parent/child, ancestor/descendant, preceding/following etc. relationships between node sets. In this paper, we present a simple method for accelerating structural joins which is very easy to implement on different platforms. Our idea is to split the nodes into disjoint partitions and use this information to avoid unnecessary structural joins. Despite its simplicity, our proposal can considerably accelerate XPath evaluation on different XML management systems. To exemplify this, we describe two implementation options of our method - one built from the scratch and one based on a relational database - and present the results of our experiments. Keywords: XML, XPath, structural joins, spatial indexes.
1 Introduction In many modern applications areas, such as bioinformatics and Web services, there is a need to efficiently store and query large heterogeneous data sets marked up using XML [1], i.e., represented as a partially ordered, labeled XML tree. This imposes a great challenge on data management systems which have traditionally been designed to cope with structured rather than semistructured data. In recent years, a lot of work has thus been done to develop new methods for storing and querying XML data using XML query languages, such as XPath [2] and XQuery [3]. At the heart of the XPath query language, there are 12 axes, i.e., operators for tree traversal. In this paper, we present a method for accelerating the evaluation of the major axes, i.e., the ancestor, descendant, preceding, and following axes, through partitioning. We make use of the observation that starting from any node, the main axes partition the document into four disjoint partitions. Our idea is to partition the nodes more accurately, store the partition information, and use this information to filter the nodes that cannot be contained in the result before actually performing the axis. Since our method is based on simple geometric properties of the preorder and postorder numbers of the nodes, it is somewhat similar to the XML indexing approaches based on spatial structures, such as R-trees [4] or UB-trees [5]. The method described in this paper, however, is very easy to implement even using B-trees, and thus it can easily be used to accelerate XPath evaluation in XML management systems built on relational databases, such as XRel [6] and XPath accelerator [4]. Furthermore, our approach can J. Filipe and J. Cordeiro (Eds.): WEBIST 2007, LNBIP 8, pp. 98–108, 2008. c Springer-Verlag Berlin Heidelberg 2008
Efficient Queries on XML Data through Partitioning
99
also be tailored to be used with other node identification schemes, such as the order/size scheme. The rest of our paper is organized as follows. In Section 2, we take a short look at the related work and in Section 3, we discuss the basics of the XPath query language. Our partitioning method is presented in Section 4. In section 5, we discuss the implementation of our method and in Section 6, the results of our experimental evaluation. Section 7 concludes this article.
2 Related Work In the XML research literature, there are numerous examples of different XML management systems. These systems can roughly be categorized into the following three categories: 1. The flat streams approach handles the documents as byte streams [7,8]. Obviously, accessing the structure of the documents requires parsing and consumes a lot of time, but this option might still be viable in a setting where the documents to be stored and queried are small. 2. In the metamodeling approach, the XML documents are first represented as trees which are then stored into a database. With suitable indexes this approach provides fast access to subtrees, and thus the metamodeling approach is by far the most popular in the field of XML management research. Unlike in the flat streams approach, however, rebuilding large parts of original documents from a large number of individual nodes can be rather expensive. The partitioning method discussed in this paper also falls into this category. 3. The mixed approach aims at combining the previous two approaches. In some systems, the data is stored in two redundant repositories, one flat and one metamodeled, which allows the stored documents to be queried and the result documents to be built efficiently, but obviously creates some storage overhead. This problem could be tackled by compression to which XML data is often very amenable. There are also examples of a hybrid approach in which coarser structures of the XML documents are modeled as trees and finer structures as flat streams [9]. A lot of work has been carried out to accelerate structural joins [10], i.e., operations which determine all occurrences of parent/child, ancestor/descendant, preceding/following etc. relationships between node sets. Some approaches are based on indexes built on XML trees, whereas others aim at designing more efficient join algorithms. The former category includes the so called proxy index [11,12] which effectively partitions the nodes into overlapping partitions so that the ancestors of any given node are contained within the same partition. Thus, when the ancestors of a given node are retrieved it is sufficient to check the partitions to which the node is assigned. Conversely, descendants can also be found efficiently since there is no need to check the partitions to which the node is not assigned. However, this approach is suitable for accelerating only ancestor/descendant operations. The method discussed in this paper, on the contrary, can also accelerate preceding/following operations.
100
O. Luoma
The other group of methods include, for example, the staircase join [13]. The idea of the staircase join is to prune the set of context nodes, i.e., the initial nodes from which the axis is performed. For instance, for any two context nodes n and m such that m is a descendant of n, all descendants m are also descendants of n, and thus m can be pruned before evaluating the descendant axis. Another method based on preprocessing the nodes was proposed in [14]. In this approach the nodes were partitioned somewhat similarly to the method discussed in this paper. However, both of these methods require a considerable amount of programming effort since they work by preprocessing the data rather than by building indexes. In the context of relational databases, for example, this would mean either reprogramming the DBMS internals or programming a collection of external classes to implement the join algorithms. Our method, on the contrary, requires very little programming effort even when implemented using a relational database, which we regard as the main advantage of our approach. The nodes of an XML tree can also be partitioned according to some other criteria than mere structure. In almost all XML management systems, an index is built on the names of the nodes, which essentially partitions the nodes according to their tags and makes it possible to locate nodes with a given tag very efficiently. An index can also be built on the element paths of the nodes or some more selective structural criterion. The F&B index [15] and the schema graph [16], for example, partition the nodes into equivalence classes such that two nodes are considered similar if they share the same element path and their subtree structures are similar. One should notice that the aforementioned methods and the method discussed in this paper are not exclusive but rather complementary.
3 XPath Basics As mentioned earlier, XPath [2] is based on a tree representation of a well-formed XML document, i.e., a document that conforms to the syntactic rules of XML. A simple example of an XML tree corresponding to the XML document <e>kl <e>ez is presented in Fig. 1. The nodes are identified using their preorder and postorder numbers which encode a lot of structural information1. The tree traversals in XPath are based on 12 axes which are presented in Table 3. In simple terms, an XPath query can be thought of as a series of location steps of the form /axis::nodetest[predicate] which start from a context node - initially the root of the tree - and select a set of related nodes specified by the axis. A node test can be used to restrict the name or the type of the selected nodes. An additional predicate can be used to filter the resulting node set further. The location step n/child::c[child::*], for example, selects all children of the context node n which are named “c” and have one or more child nodes. As a shorthand for axes child, descendant-or-self, and attribute, one can use /, //, and /@, respectively. 1
In preorder traversal, a node is visited before its subtrees are recursively traversed from left to right and in postorder traversal, a node is visited after its subtrees have been recursively traversed from left to right.
Efficient Queries on XML Data through Partitioning
101
1, 10 b
2, 2 c
3, 1 d=”y”
4, 6 c
5, 3 d=”y”
8, 9 c
6, 5 e
9, 8 e
7, 4 ”kl”
10, 7 ”ez”
Fig. 1. An XML tree Table 1. XPath axes and their semantics Axis Semantics of n/Axis parent Parent of n. child Children of n, no attribute nodes. ancestor Transitive closure of parent. descendant Transitive closure of child, no attribute nodes. ancestor-or-self Like ancestor, plus n. descendant-or-self Like descendant, plus n, no attribute nodes. preceding Nodes preceding n, no ancestors or attribute nodes. following Nodes following n, no descendants or attribute nodes. preceding-sibling Preceding siblings of n, no attribute nodes. following-sibling Following siblings of n, no attribute nodes. attribute Attribute nodes of n. self Node n.
Using XPath, it is also possible to set conditions for the string values of the nodes. The XPath query //record=‘‘John Scofield Groove Elation Blue Note", for instance, selects all element nodes with label “record” for which the value of all text node descendants concatenated in document order matches “John Scofield Groove Elation Blue Note”. Notice also that the result of an XPath query is the concatenation of the parts of the document corresponding to the result nodes. In our example case, query //c//*, for example, would result in no less than <e>kl<e>ez<e>kl <e>ez.
4 Our Partitioning Method As mentioned earlier, we rely on pre-/postorder encoding [17], i.e., we assign both preorder and postorder numbers for the nodes. This encoding provides us with enough information to evaluate the four major axes [4] as follows: Proposition 1. Let pre(n) and post(n) denote the preorder and postorder numbers of node n, respectively. For any two nodes n and m, n ∈ m/ancestor::* iff pre(n) <
102
O. Luoma
10
Post
Post
10
Pre
Pre 10
10
Fig. 2. Examples on the major axes (left) and partitioning using value 4 for p (right)
pre(m) and post(n) > post(m), n ∈ m/descendant::* iff pre(n) > pre(m) and post(n) < post(m), n ∈ m/preceding::* iff pre(n) < pre(m) and post(n) < post(m), and n ∈ m/following::* iff pre(n) > pre(m) and post(n) > post(m). This observation is actually slightly inaccurate since the attribute nodes are not in the result of, for instance, the descendant axis. However, if all nodes are treated equally, the proposition holds. The observation is exemplified on the left side of Fig. 2 which shows the partitioning created by the major axes using node (6,5) as the context node. The ancestors of node (6,5) can be found in the upper-left partition, the descendants in the lower-right partition, the predecessors in the lower-left partition, and the followers in the upper-right partition. Based on the previous observation, the following proposition is obvious: Proposition 2. For any two nodes n and m and for any p > 0, pre(n)/p ≤ pre(m) /p and post(n)/p ≥ post(m)/p if n ∈ m/ancestor::*, pre(n)/p ≥ pre(m)/p and post(n)/p ≤ post(m)/p if n ∈ m/descendant::*, pre(n) /p ≤ pre(m)/p and post(n)/p ≤ post(m)/p if n ∈ m/preceding::*, and pre(n)/p ≥ pre(m)/p and post(n)/p ≥ post(m)/p if n ∈ m /following::*. Proposition 2 provides us with simple means for partitioning the nodes into disjoint subsets. The partitioning is exemplified in Fig. 2 (right) which shows the nodes of our example tree partitioned using value 4 for p. It should now be obvious that when searching, for instance, the ancestors of node (6,5), it is sufficient to check the nodes in the shaded partitions since other partitions cannot contain any ancestors. In what follows, values pre(n)/p and pre(n)/p where n is a node residing in partition P are simply referred to as the preorder and postorder number of P and denoted by pre(P ) and post(P ), respectively. Many minor axes can also benefit from partitioning. The ancestor-or-self and descendant-or-self axes, for example, behave similarly to ancestor and descendant axes, and thus they can be accelerated using the same partition information. The results of preceding-sibling and following-sibling, on the
Efficient Queries on XML Data through Partitioning
103
other hand, are subsets of preceding and following and the results of parent and child subsets of ancestor and descendant so they can also be accelerated using the partitioning. However, other minor axes than ancestor-or-self and descendant-or-self are generally very easy to evaluate, and thus using the partition information to evaluate them is often just unnecessary overhead especially if we have an index which can be used to locate the nodes with a given parent node efficiently [18]. One should also notice that in practice, the nodes are not distributed evenly and there are several empty partitions. Thus, the number of non-empty partitions is usually much smaller than p2 .
5 Implementation Options 5.1 Relational Implementation To test our idea, we designed a relational database according to the principles described in the previous section. As the basis of our implementation, we chose the XPath accelerator [4]. In order to store the partition information, we employed relation Part, and thus we ended up with the following relational schema; the underlined attributes serve as the primary keys: Node(Pre, Post, Par, Part, Type, Name, Value) Part(Part, Pre, Post)
As in the original proposal, the database attributes Pre, Post, Par, Type, Name, and Value of the Node relation correspond to the pre- and postorder numbers, the preorder number of the parent, the type, the name, and the string value of the node, respectively. The database attributes Type, Name, and Value are needed to support node tests and string value tests; the axes can be evaluated using database attributes Pre, Post, and Par. The partition information is contained in the Part table in which the database attributes Pre and Post correspond to the preorder and postorder number of the partition, respectively. The database attribute Part in the Node relation is a foreign key referencing to the primary key of the Part relation. For the sake of brevity, we do not describe the XPath-to-SQL query translation in detail; the details can be found in [4] and [6]. For our purposes, it is sufficient to say that in order to perform structural joins, these systems issue SQL queries which involve nonequijoins, i.e., joins using < or > as the join condition, which can easily lead to scalability problems [12,18,19]. In XPath accelerator, the XPath query n0 /descendant::*, for instance, is transformed into the following SQL query: SELECT DISTINCT n1 .* FROM Node n0 , Node n1 WHERE n1 .Pre>n0 .Pre AND n1 .Post
Notice that tuple variables n0 and n1 are joined completely using slow nonequijoins. However, we can use the partition information stored into table Part to replace some
104
O. Luoma
of the nonequijoins with much faster equijoins, i.e., joins using = as the join condition. With the partition information, the same query can be evaluated much more efficiently using the following piece of SQL: SELECT DISTINCT n1 .* FROM Node n0 , Node n1 , Part a1 , Part b1 WHERE a1 .Part=n0 .Part AND b1 .Pre>=a1 .Pre AND b1 .Post<=a1 .Post AND n1 .Part=b1 .Part AND n1 .Pre>n0 .Pre AND n1 .Post
In simple terms, we use tuple variable a1 for the partition in which node corresponding to tuple variable n0 resides. Variable b1 corresponds to the partitions which can contain descendants of n0 ; condition n1 .Part=b1 .Part restricts our search into the nodes residing in those partitions. Finally, we simply use condition n1 .Pre>n0 .Pre AND n1 .Post
Efficient Queries on XML Data through Partitioning
105
The structural joins within a partition, then, are carried out using algorithm joinPart: joinPart(n, P , axis) in: Node n, partition P , XPath axis axis out: Nodes in n/axis::* of p if axis = preceding then for each m ∈ P if pre(m) < pre(n) and post(m) < post(n) result ← result + m if axis = preceding-sibling then for each m ∈ P if pre(m) < pre(n) and par(m) = par(n) result ← result + m ... return result
Again, our algorithm is heavily simplified. However, it is easy to extend the algorithm to support node tests and string value tests by checking the type, name, and string value attributes of the node. Notice also that we could dramatically lower the number of structural joins by considering the fact that the all nodes in partition P such that pre(P ) < pre(n)/p and post(P ) < post(n)/p are guaranteed to be in n/preceding::*, and thus they could be added to the result without performing the structural joins. We also implemented this option but found out that is did not lead to considerable gains in evaluation times. However, the benefits obviously depend on the implementation and in some cases, this approach can further accelerate the joins.
6 Experimental Results 6.1 Relational Implementation We implemented the relational option using Windows XP and Microsoft SQL Server 2000 running on 2.00 GHz Pentium PC Equipped with 512 MB of RAM and standard IDE disks. The algorithms needed to build the databases, as well as the algorithms for query translation, were implemented using Java. We built indexes on Node(Post), Node(Name), Part(Pre), and Part(Post) and a clustered index on Node(Part). As our test data sets, we used the 1998 baseball statistics (52707 nodes) and the collection of Shakespeare’s plays (327129 nodes), both available at http://www.ibiblio.org/examples. The sizes of these documents were 640 kB and 7.47 MB, respectively; values 1,2,4, ..., 256 were used as a partitioning factor p. Fig. 3 presents the query performance for the major axes using a relatively large set of context nodes; in this figure, the word “Partitions” refers to the value of p or the number of partitions per (preorder or postorder) dimension. In the case of baseball.xml, nodes with name “PLAYER” (1226 nodes) and in plays.xml, nodes with name “SCENE” (750 nodes) served as the context nodes. One should notice that our method can lead to considerable performance gains in ancestor and descendant axes even with a very small amount of partitions. In the case baseball.xml and 16 partitions per dimension,
106
O. Luoma
for example, there were only 55 partitions in total. Thus, there were only 55 rows in the Part table, which is certainly acceptable considering that there are no less than 52707 nodes in the Node table. Even in the case of 256 partitions per dimension, there were only 772 rows in the Part table for “1998statistics.xml” and 859 rows for “plays.xml”. In the case of preceding and following axes, however, no real acceleration could be observed. These axes could be actually evaluated very efficiently even without the partitionin, which is due to the very sophisticated join algorithms implemented in SQL Server. Nevertheless, considering that the ancestor and descendant axes were very time-consuming to evaluate without the partition information, we are still convinced that the small amount of partition information can indeed pull its weight. 6.2 Native Implementation The evaluation of our native implementation was carried out using the same equipment and data sets which were used in the relational case. Furthermore, the same number of partitions per dimension were used but in these tests, we randomly selected the context nodes. Fig. 4 and Fig. 5 present the results obtained using the native implementation; all results are averages for 100 iterations. Overall, the ancestor and descendant axes behaved similarly to the relational case, i.e., they were considerably accelerated. However, the preceding and following axes were also accelerated by roughly a factor of two. This is actually intuitively clear since an average node has much more predecessors and followers than it has ancestors and descendants, and thus the preceding and following axes usually result in much larger node sets. In the case of ancestor axis, for example, the partitioning can help us to filter out a massive amount of nodes, i.e., the most of the descendants, predecessors, and followers, whereas in the case of preceding axis, a much smaller amount of nodes with respect to the size of the result can be filtered. One should also notice that ordering the nodes according to their preorder numbers favors the preceding axis over the following axis. After the preorder numbers have been checked in the case of preceding axis, we still have to filter out the ancestors using the postorder numbers of the nodes. In the case of following axis, on the contrary, the descendants have to be filtered, which is a harder task since a node in an 120000
30000 ancestor descendant preceding following
ancestor descendant preceding following
100000
20000
Time (ms)
Time (ms)
25000
15000 10000 5000
80000 60000 40000 20000
0 1
10 Partitions
100
0 1
10 Partitions
Fig. 3. Relational results 1998statistics.xml (left) and plays.xml (right)
100
Efficient Queries on XML Data through Partitioning 120000
700000 ancestor descendant preceding following
100000
ancestor descendant preceding following
600000 500000 Joins
80000 Joins
107
60000
400000 300000
40000
200000
20000
100000
0
0 1
10 Partitions
100
1
10 Partitions
100
Fig. 4. Number of joins for the major axes using random context nodes for 1998statistics.xml (left) and plays.xml (right); native implementation 300 ancestor descendant preceding following
200
Time (ms)
Time (ms)
250
150 100 50 0 1
10 Partitions
100
2000 1800 1600 1400 1200 1000 800 600 400 200 0
ancestor descendant preceding following
1
10 Partitions
100
Fig. 5. Times for the major axes using random context nodes for 1998statistics.xml (left) and plays.xml (right); native implementation
XML tree usually has more descendants than it has ancestors. Conversely, sorting the nodes according to their postorder numbers slightly favors the following axis.
7 Conclusions In this paper, we discussed a partitioning method which can considerably accelerate the evaluation of XPath axes in different XML management systems. Our method is based on simple properties of the preorder and postorder numbers of the nodes in an XML tree, which makes it very easy to implement. This was exemplified by presenting two different implementations of our idea which both indicated that alhtough our idea is simple and very easy to implement, it can lead to considerable performance gains. Acknowledgements. The author would like to thank Professor Jukka Teuhola of University of Turku (Finland) who provided his valuable comments during the preparation of this paper.
108
O. Luoma
References 1. W3C (World Wide Web Consortium): Extensible Markup Language (XML) 1.0, http://www.w3c.org/TR/REC-xml/ 2. W3C (World Wide Web Consortium): XML Path Language (XPath) 2.0, http://www.w3c.org/TR/xpath20/ 3. W3C (World Wide Web Consortium): XQuery 1.0: An XML Query Language, http://www.w3c.org/TR/xquery/ 4. Grust, T.: Accelerating XPath Location Steps. In: Proceedings of the 2002 ACM SIGMOD Conference on Management of Data, pp. 109–120 (2002) 5. Kr´atk´y, M., Pokorn´y, J., Sn´asˇ el, V.: Implementation of XPath Axes in the Multi-Dimensional Approach to Indexing XML Data. In: Proceedings of Current Trends in Database Technology, pp. 219–229 (2004) 6. Yoshikawa, M., Amagasa, T., Shimura, T., Uemura, S.: XRel: A Path-Based Approach to Storage and Retrieval of XML Documents Using Relational Databases. ACM Transactions on Internet Technology 1(1), 110–141 (2001) 7. Peng, F., Chawathe, S.S.: XPath Queries on Streaming Data. In: Proceedings of the 2003 ACM SIGMOD Conference on Management of Data, pp. 431–442 (2003) 8. Barton, C., Charles, P., Goyal, D., Raghavachari, M., Fontoura, M., Josifovski, V.: Streaming XPath Processing with Forward and Backward Axes. In: Proceedings of the 19th International Conference on Data Engineering, pp. 455–466 (2003) 9. Fiebig, T., Helmer, S., Kanne, C.-C., Moerkotte, G., Neumann, J., Schiele, R., Westmann, T.: Natix: A Technology Overview. In: Chaudhri, A.B., Jeckle, M., Rahm, E., Unland, R. (eds.) NODe-WS 2002. LNCS, vol. 2593, pp. 12–33. Springer, Heidelberg (2003) 10. Al-Khalifa, S., Jagadish, H.V., Patel, J.M., Wu, Y., Koudas, N., Srivastava, D.: Structural Joins: A Primitive for Efficient XML Query Pattern Matching. In: Proceedings of the 18th International Conference on Data Engineering, pp. 141–152 (2002) 11. Luoma, O.: Supporting XPath Axes with Relational Databases Using a Proxy Index. In: Proceedings of the 3rd International XML Database Symposium, pp. 99–113 (2005) 12. Luoma, O.: Xeek: An Efficient Method for Supporting XPath Evaluation with Relational Databases. In: Local Proceedings of the 10th East-European Conference on Advances in Databases and Information Systems, pp. 30–45 (2006) 13. Grust, T., van Keulen, M.: Tree Awareness for Relational RDBMS Kernels: Staircase Join. In: Intelligent Search on XML Data, Applications, Languages, Models, Implementations, and Benchmarks, pp. 231–245 (2003) 14. Tang, N., Yu, J.X., Wong, K.-F., L¨u, K., Li, J.: Accelerating XML Structural Join by Partitioning. In: Proceedings of the 16th International Conference on Database and Expert Systems Applications, pp. 280–289 (2005) 15. Kaushik, R., Bohannon, P., Naughton, J.F., Korth, H.F.: Covering Indexes for Branching Path Queries. In: Proceedings of the 2002 ACM SIGMOD Conference on Management of Data, pp. 133–144 (2002) 16. Luoma, O.: A Structure-Based Filtering Method for XML Management Systems. In: Proceedings of the 15th International Conference on Database and Expert Systems Applications, pp. 401–410 (2004) 17. Dietz, P.F.: Maintaining Order in a Linked List. In: Proceedings of the 14th Annual Symposium on Theory of Computing, pp. 122–127 (1982) 18. Luoma, O.: Modeling Nested Relationships in XML Documents using Relational Databases. In: Proceedings of the 31st Conference on Current Trends in Theory and Practice of Computer Science, pp. 259–268 (2005) 19. Luoma, O.: Efficient Methods for Storing and Querying XML Data with Relational Databases. PhD Thesis, University of Turku (2007)
Contract Based Behavior Model for Services Coordination Alberto Portilla1,2, , Genoveva Vargas-Solar1, Christine Collet1 Jos´e-Luis Zechinelli-Martini2, and Luciano Garc´ıa-Ba˜nuelos3 1
Laboratoire d’Informatique de Grenoble BP 72, 38402 Saint Martin d’H`eres, France 2 Research Center of Information and Automation Technologies Universidad de las Am´ericas, Puebla Sta Catarina Martir, 72820, San Andres Cholula, Mexico 3 Facultad de Ciencias B´asicas, Ingenier´ıa y Tecnolog´ıa Universidad Aut´onoma de Tlaxcala Calzada Apizaquito S/N Km. 1.5, 90300 Apizaco, Tlaxcala, Mexico
Abstract. A key step towards consistent services coordination to provide non functional properties. In this sense, transactional properties are particularly relevant because of the business nature of current applications. While services composition has been successfully addressed, transactional properties have been main- ly provided by ad-hoc and limited solutions at systems’ back end. This paper proposes a flexible transactional behavior model for services coordination. We assume that given a flow describing the application logic of a service based application, it is possible to associate to it a personalized transactional behavior in an orthogonal way. This behavior is defined by specifying contracts and associating a well defined behavior to the activities participating in the coordination. Such contracts ensure transactional properties at execution time in the presence of exceptions. Keywords: Transactional behavior, advanced transactional models, services oriented systems, services based applications, web services.
1 Introduction A key step towards consistent services coordination is to provide non functional properties. In this sense, transactional properties are particularly relevant because of the business nature of current applications. While services composition has been successfully addressed, transactional properties have been mainly provided by ad-hoc and limited solutions at systems’ back end. In service based applications, a service is a software component, available in a network, that offers some functions by exporting an application programming interface (API). Nowadays, services based applications capture complex processes for achieving a goal, they are built by coordinating interactions among several services.
Supported by the Mexican Education Council (PROMEP-SEP), the Autonomous University of Tlaxcala, M´exico and the Jenkins Excellence Fellowship Program at UDLA.
J. Filipe and J. Cordeiro (Eds.): WEBIST 2007, LNBIP 8, pp. 109–123, 2008. c Springer-Verlag Berlin Heidelberg 2008
110
A. Portilla et al.
Figure 1 shows an example of a service based application composed using services of several providers. In the Figure, boxes represent activities and each activity represents a service method call. Let us consider that this application is related to a trip reservation with the following logic: given reservation information it is necessary to get a bank authorization for the payment (a1 ), if the payment has been guaranteed, it is necessary to reserve the flight (a2 ), a hotel (a3 ) and a car (a4 ), otherwise the reservation is canceled (a5 ). In addition, there are non functional issues that must be considered, some are related to application semantics and others are related to the activity semantics. In this example, at the application level we can consider the following aspects: before making a reservation (flight, hotel or car) the payment must be authorized (a1 ); a reservation cannot be rejected (a5 ) and processed at the same time (a2 , a3 and a4 ). At the activity level we can consider: there are several alternatives for executing an activity. For example, flight reservation can be completed using Mexicana, Air France or British Airways.There are critical activities for the success of the application. For example, while flight reservation and hotel reservation are critical for the acceptance of the trip reservation, an unsuccessful car reservation can be tolerated. Some activities can be retried in case of failure, for example, obtaining the bank authorization or booking the flight.
Fig. 1. Service based application example
In particular, these aspects illustrate transactional behavior. Recall that a transaction is a unit of work composed of several operations that must be handled as an execution unit. Such execution unit can be successfully executed or not [22,5]. Even if such a behavior has been addressed for managing data (DBMS) and process oriented applications (transactional workflows), existing solutions are not well adapted for: – synchronizing highly autonomous services execution within transactions and, – for dealing with long duration resources blocking. Transactional behavior for services coordination is currently addressed in three ways: it is explicitly defined within the application logic by adding sequences of activities to its workflow. Such activities implement advanced transactional models (e.g. [30], [25], etc.); it is implemented by transactional protocols (e.g. BTP [12], WS-Tx [3], etc.); or it is specified by transactional policies expressed using transactional coordination languages (e.g., πt-calculus[21], [1], etc.). However, this leads to complex service based applications difficult to maintain and hardly adaptable to different application requirements [17,29,6]. This paper proposes a contract based behavior model for specifying transactional properties to services coordination. Our approach addresses atomic behavior based on
Contract Based Behavior Model for Services Coordination
111
atomicity contracts for associating atomicity properties and recovery strategies to sets of activities (i.e., spheres [4,17,20]). Our model separates the specification of coordination and transactional behavior as follows: – Application logic is captured by using a coordination approach based on workflows. – Transactional behavior is defined by contracts according to the behavior of services participating in a coordination and the semantic of the target service based application. Given a coordination describing the application logic of a service based application it is possible to associate it with a transactional behavior. Different from existing proposals our approach enables the definition of transactional behavior without implying the modification of the original application logic. Besides, our model is not based in any advanced transactional model, rather it enables such kind of behavior if necessary. Our approach proposes recovery strategies for ensuring transactional behavior of services coordination. Such strategies can be chosen according to specific application requirements. Thus, the same type of application can have different associated behaviors according to given application requirements. The remainder of the paper is organized as follows. Section 2 gives an overview of services coordination with transactional properties. Section 3 introduces the main concepts of the transactional behavior model that we propose. Section 4 illustrates the specification of transactional behavior for a target application through a use case. Section 5 discusses the extensibility of our model. Section 6 compares our work with existing approaches that provide transactional properties to services coordination. Finally, Section 7 concludes this paper and discusses future work.
2 Coordination Model with Transactional Behavior Properties A coordination model can be used for expressing an application logic defined by an orchestration. An orchestration specifies the workflow whereby activities are synchronized using ordering operators (see Figure 2).
Fig. 2. Coordination UML class diagram
112
A. Portilla et al.
Acoording to its role within the application logic, an activity has an associated behavior that specifies the way it should be handled under the presence of exceptional situations. 2.1 Service A service represents a set of functions provided by an autonomous software component through a port. It is formally defined as a quadruple S-name, F M , N M , P where: – S-name is the name of the service. – F M and N M are interfaces consisting of a finite set of methods that provide respectively functional and non functional aspects, for example XA/XOPEN [27]. A method is defined by a triple M =M -name, In, Out where: • M -name is the name of the method. • In and Out are respectively the sets of input and output parameters. • A parameter is defined by a tuple P -name, T ype, where: ∗ P -name is the name of the parameter. ∗ T ype represents a predefined type such as boolean, string, integer, etc. – P is the communication port used for interacting with the service. The trip reservation application previously introduced uses three services providers (see Figure 1): bank controls financial operations executed on accounts (Provider 1), travel agency rates and manages flight and hotel reservations (Provider 2), and Provider 3 manages car rental. 2.2 Orchestration An orchestration describes an application logic through control and data flows. In our work an orchestration is defined using Petri nets. A Petri net consists of places, transitions, tokens, and directed arcs. A place models the states among two transitions. A transition represents a task, hereinafter, refered as activity. A token represents an execution state. A directed arc is used to link places and transitions. Therefore, a Petri net is defined as a triple P L, A, F [28] where: – P L is a finite set of places. – A is a finite set of transitions, P L ∩ A = φ. – F ⊆ (P L x A) ∪ (A x P L) is a set of arcs. Using Petri nets the control flow of an orchestration is expressed by order operators (sequential routing, parallel routing, selective routing, and iteration) that relate activities. The data flow is embedded within the control flow. Figure 3 shows the orchestration of the trip reservation example by using Petri nets: places are represented as circles (p1 , p2 , p3 , p4 , p5 and p6 ), transitions are represented as boxes (a1 , a2 , a3 , a4 and a5 ), tokens are represented as black circles, and flow relationships are represented as arrowed lines.
Contract Based Behavior Model for Services Coordination
113
Fig. 3. Service based application with Petri nets
2.3 Activity Behavior Depending on its role within an application logic, an activity can have different behaviors at execution time in the presence of failures [14,9]. There are three possible scenarios: 1. An activity can be rolled back/undone, but possible side effects must be taken into account. In the trip reservation example (see Figure 1), undoing the flight reservation (a2 ) can imply an extra charge. 2. A committed activity can be compensated by a so called compensation activity that semantically undoes its actions but it does not necessarily gets back the application to the previous state before the activity was committed. For example, in trip reservation application, canceling a flight reservation can be undone with an extra charge which does not return the balance account to the previous state. However, there are some activities that cannot be undone, for example personalizing an item using laser engraving. 3. The execution of an activity can be retried several times within the same coordination. In our example, obtaining the hotel reservation can be executed several times assuming that hotels accommodation is not easy during hollidays. However, retrying an activity can be associated to other constraints, such as the number of times that an activity can be retried. Considering the above scenarios an activity can be1 (see Table 1): – Non vital. The activity does not need to be compensated if it has to be undone/rolled back after having committed. For example, let us assume an activity that uses the trip reservation information for sending commercial information. It can be defined as non vital because the reservation can be done even if the the customer does not receive the publicity. – Critical. There is no way to compensate or undo the effects of an activity once it has been committed. Recalling the trip reservation example, the activity bank 1
For the time being, this classification is non exhaustive.
114
A. Portilla et al.
authorization is critical because once a reservation has been authorized the bank cannot change its opinion, so the execution of the reservation cannot be stopped for this reason. – Undoable. A commited activity can be undone by a compensating activity without causing side-effects. Once it has been compensated the activity can be retried. In our example the activity hotel reservation can be defined as undoable activity, assuming that some hotel reservation can be canceled at most three days before the date of arrival. – Compensatable. A commited activity can be undone by a compensating activity with associated side-effects. Once it has been compensated the activity can be retried but other aspects must be considered such as application restrictions, monetary costs, temporal constraints, etc. In our trip reservation example the activity flight reservation can be undone by an activity that cancels the flight and that reimburses a percentage of the original amount to the costumer account. However, retrying flight reservation when canceled can imply costs that might not be acceptable for the customer. Table 1. Types of activities Activity Causes Can be Can be type side effects compensated retried Non vital No No Yes Critical No Undoable No Yes Yes Compensatable Yes Yes Maybe
Execution State. At execution time, an activity can have the following associated states (see Figure 4): – – – – – – –
Ready when the activity is ready to be executed. Blocked is used for intercepting activate call for some purposes. Active when the activity is being executed. Failed when the activity execution has unsuccessfully finished. Committed when the activity execution failed. Withdrawn when the activity has been withdrawn. Compensated when the activity execution has been semantically undone or rolled back.
We consider activities as atomic, this means that when executed they commit or fail.
3 Transactional Behavior Model Inspired in [4], activities within an orchestration can be grouped together into a behavior contract that can have specific associated transactional properties (e.g., atomicity) and execution strategies for ensuring such properties. For example, the activities used for
Contract Based Behavior Model for Services Coordination
115
Fig. 4. Activity execution transition state diagram
implementing the payment process can be executed atomically, so that non authorized purchases do not lead to the generation of an invoice. Furthermore, several behavior contracts can be specified and composed within an application logic so that its behavior can be controlled in a fined grained manner. 3.1 Contract A contract specifies a transactional behavior for a set of activities (e.g. atomicity, isolation, etc.). It specifies a recovery strategy that must be executed in case of failure according to the contract type. Grouping Activities. A contract groups together sequences of activities or contracts as shown in Figure 2. In the following let A be a set of activities. This set may be obtained from an orchestration defined as a Petri net P L, A, F . The set of contracts associated to A denoted C(A) is defined by following rules: 1. If ai ∈ A, then (ai ) ∈ C(A) is a simple contract and ai is called the component of the contract. For example c2 in Figure 5, where c2 = (a5 ). 2. If c1 , c2 , ...,cn ∈ C(A), (c1 ,c2 ,...,cn ) ∈ C(A) is a complex contract and c1 , c2 , ...,cn are called the components of the contract. For example c1 in Figure 5 where c2 = ((a2 ), (a3 ), (a4 )). Similarly, c3 is a complex contract where c3 = (c1 , c2 ). Behavior. The behavior of a contract c at execution time can be deduced from the behavior of its components according to the following rules: – c is non vital, if its components are non vital. For example, assuming that components (a2 ), (a3 ) and (a4 ) are non vital the contract c1 is non vital (see Figure 5). – c is critical, if it contains at least one critical component that has been committed. For example assuming that the component (a2 ) is critical and it has already been committed, then contract c1 is critical (see Figure 5). Note that a token in place p3 indicates that (a2 ) is committed for a given execution state.
116
A. Portilla et al.
– c is undoable, if all its components are undoable or non vital. For example assuming that (a2 ) is non vital and (a3 ) and (a4 ) are undoable, contract c1 is undoable (see Figure 5). – c is compensatable, if all its components are non vital, undoable or compensatable. For example assuming that (a2 ) is non vital, (a3 ) is undoable, and (a4 ) is compensatable, contract c1 is compensatable (see Figure 5). Note that the behavior of a contract is determined at run time since it depends on the behavior of its components and on their execution state. For example, consider contract c1 in Figure 5 with component (a3 ) defined as critical. Then c1 is compensatable before (a3 ) commits, and it becomes critical after (a3 ) has been committed.
Fig. 5. Contracts example
Execution State. The execution state of a contract c can be: – Ready when c is ready to be executed, this implies that all its components are ready to be executed. – Active when c is being executed, at least one of its components is being executed. – Failed when c finished unsuccessfully. At least one of its vital components fails. – Committed when c finished successfully. All vital contract components commit. – Compensated when c has been semantically undone. At least one of its vital components has been compensated. We also consider contracts as atomic: the initial state of any contract is ready, then it is active during its execution, but after its execution, a contract either fails or commits. Well Formed Contract. A contract is well formed if it is guaranteed that it will commit or fail. Thus a contract c is well formed if it contains at most one critical component, denoted as ccr , or all components within the contract are not critical. Note that when c contains a critical activity, once ccr commits c must commit because ccr cannot be compensated or undone. In such a case, for ensuring that c commits when ccr commits, it is necessary that the set of components to be executed after ccr must be retriable. For
Contract Based Behavior Model for Services Coordination
117
example, contract c1 in Figure 5 is well formed if (a2 ) is compensatable, (a3 ) is critical and (a4 ) is undoable. Safe Point. Given a simple contract c its safe point is either the most recent critical component committed within c or the activity/contract before the first component of c. The safe point represents the last consistent state of a coordination before the execution of c. If c fails, it is possible to cancel the components within c and restart the execution from the safe point. For example, a1 can be a safe point for contract c1 . Assuming that a1 is the safe point of c1 , it is also the safe point for contract c3 if contract c2 is well formed and contract c2 is undoable (see Figure 5). Recovery Strategies. There are three main recovery strategies that can be implemented in case of failure: forward execution, backward and forward recovery. Recovery strategies help to decide wether the execution of a coordination can continue after the occurrence of a failure. Given a well formed contract c with its associated safe point sp: – Forward execution. If a retriable or non vital component in c fails the execution of c can proceed. For example considering that all components of c1 are non vital (see Figure 5), its execution can proceed even if one of its components fails. – Backward recovery. If a component in c fails and forward recovery cannot be applied then previously committed components in c are undone by their corresponding compensating activities until a safe point is reached.For example, assume that the components in c1 are defined as follows: (a2 ) and (a3 ) are compensatable and (a4 ) is critical. If during the execution of c1 , (a3 ) fails, then (a2 ) and (a3 ) are succesively compensated until a1 is reached. – Forward recovery. It combines backward recovery and forward execution. If forward execution cannot be applied because a critical or compensatable component has failed then apply backward recovery until a safe point is reached and then apply forward execution with a different execution path. Using one of the above recovery strategies depends on the type of component that fails and the type of previously executed components. 3.2 Atomicity Contracts A classic example of a transactional behavior is the atomicity requirement. We used the notion of contract and recovery strategies for defining so called atomicity contracts. We show in the following the definition of a strict atomicity contract2 . Intuitively, an atomicity contract specifies that given a contrat c all its components must be successfully executed or no component at all. Details on the definition of atomicity contracts can be found in [23]. The strict atomicity contract is defined by the following rules. Given a well formed contract c: – Rule 1. If a critical component in c fails, the contract fails and backward recovery is applied. 2
In fact, we are relaxing the classic atomicity concept.
118
A. Portilla et al.
– Rule 2. If a component in c fails and the critical component of c has already been committed then forward execution is applied until c commits. – Rule 3. If a component in c fails and the critical component of c has not yet been committed, forward execution is applied until c commits. If the component has failed repeatedly c fails and backward recovery is applied.
4 Defining Transactional Behavior for an e-Commerce Application Consider the e-commerce application illustrated in Figure 6. A costumer first provides an order and her/his account information, then commercial information is sent. If the purchase is authorized by the bank, the purchase can proceed; otherwise the order is canceled. For processing authorized orders, the stock is checked and the purchase process starts. If products are available the purchase is applied to an account, an invoice is sent and the order is completed; at the same time the packet is wrapped and delivered. On the other hand, if products are not available, the payment and the shipment are canceled.
Fig. 6. E-commerce application with atomicity contracts
Thus the application logic (see Figure 6) of such an application specifies activities for defining and authorizing an order (Get Order(), Get Payment Information(), Bank Autho rization()); and four other execution paths for: – Sending commercial information: Send Commercial Info(); – Stop a rejected order: Cancel Order(); – Processing an authorized order: Process Order(), Check Stock(), Wrap() and Deliver() (if there is enough stock); or Cancel Shipment(), Cancel Payment() (if there is not enough stock). – Executing the purchase: Apply Charges(), Send Invoice(), Finish Order().
Contract Based Behavior Model for Services Coordination
119
4.1 Activities The company that uses this application has specific business rules that must be respected. For instance: – BR1 : the bank authorization can be only executed once during the process of an order. – BR2 : once a package has been delivered it cannot be returned to the company. – BR3 : The company has implemented packages recycling strategies, so if a package has been wrapped, the wrapping paper can be removed and used for other packages. – BR4 : Invoices can be canceled. – BR5 : If clients are not satisfied with the products they can return them to the company but extra charges are applied. Considering such business rules, we associated the following behaviors to activities (see Figure 6): – – – –
Bank authorization() and Deliver() are critical activities (BR1 ,BR2 ). Wrap() and Send invoice() are undoable activities (BR3 , BR4 ). Apply charges() is a compensatable activity (BR5 ).
The remaining activities are considered as non vital ones because no specific business rule is specified in the application.
It must be noted that activities behavior depends on the application semantics. If the application changes its business rules, activities behavior can change. An advantage of our approach is that this can be easily done because the application logic and the transactional behavior aspects are defined separately. 4.2 Contracts The business rules defined in the previous section concern activities behavior. Still, the following business rules concern more complex patterns that complete the application semantics. The authorization of the bank is requested once the costumer information has been obtained (BR6 ). An order is finished only when charges have been applied to an account and corresponding invoice has been sent (BR7 ). Once packages have been packed they must be shipped (BR8 ). The purchase is finished once the package is delivered and the order has been processed (BR9 ). Thus the following contracts and their associated behaviors must be defined: – c1 is a critical contract that groups components involved in the order authorization ( (Get Order()), (Get Payment Information()), (Bank Authorization()) ). It is critical because it contains the critical component (Bank authorization()). – c2 is a compensatable contract that groups activities used for implementing the order processing ( (Apply Charges()), (Send Invoice()), (Finish Order()) ). It is compensatable because (Apply Charges() ) is a compensatable component. – c3 is a critical contract that groups activities for delivering the package ( (Wrap() ), (Deliver()) ). It is critical because it contains a critical component (Deliver()).
120
A. Portilla et al.
– c4 is a critical contract that groups contracts c2 and c3 for ensuring that the purchase will finish only if the package is wrapped and delivered and the payment has been executed and the invoice has been sent. It is critical because c3 is critical. Note that all contracts are well formed because they contain at most one critical component. Once contracts have been defined, they must be associated to atomicity behaviors that specify strategies to be undertaken at execution time in the presence of failure in order to ensure specific properties. In the case of our example, we have identified contracts requiring strict atomicity properties in order to fulfil BR7 -BR9 . Coupled with the strict atomicity contract that we defined in Section 3.2, c1 ensures BR6 , c2 ensures BR7 , c3 ensures BR8 , and c4 ensures BR9 .
5 Model Extensibility We claim that our transactional behavior model is flexible and adaptable to application needs because our model is not based in an specific advanced transactional model or extends an existing coordination language. As a prof of this, let us analyze the strict atomicity contract introduced in section 3.2 with respect to activities within such an atomicity contract. Let c be a simple well formed contract, c={a1 , a2 , . . . , an }. Let us consider, N V as the set of non vital activities, CR as the set of critical activities, U N as the set of undoable activities, and CO as the set of compensatable activities. There are four possibilities for activities within c: 1) all activities within c are non vital, 2) all activities within c are non vital or undoable, 3) all activities within c are non vital undoable or compensatable, and 4) there is a critical activity within c: – Case 1. This is the trivial case because c always commits. – Case 2. It extends the saga model because undoable activities can be retried, if some one of them fails it must be retried. However note that c will always commits because its activities will commit in a finite time after some retries. – Case 3. It extends the saga model if we assume that activities can be retried in a limited way. When activities cannot be retried it implements saga model behavior. This behavior has been also provided by flexible transaction model. – Case 4. It implements a kind of pivotal model such as proposed in flexible transaction model. As long as a critical activity does not commit, forward execution or backward recovery can be applied. Yet, when a critical activity commits, only forward execution is possible. It must noted that although this kind of behavior was introduced by flexible transactional model, in this case it only corresponds to one alternative execution path. By defining one atomicity contract using our model we have implemented several transactional models according to activity properties. We also have implemented such transactional models without modifying the application logic.
6 Related Works Several research projects have addressed transactional behavior for information systems, first in the context of DBMS and after for process oriented systems.
Contract Based Behavior Model for Services Coordination
121
In DBMS, transactional behavior has been tackled successfully to data through the concept of ACID transactions [5,22,7,15] and advanced transactional models [13,10,31]. These approaches are well suited when data reside in one site and transactions lifetime is short. Besides they operate with homogeneous execution units and therefore they are not applicable directly to others environments such as Internet. In workflow systems there are several approaches that aim at ensuring consistency among computations using process as execution units. WAMO [8] introduces a complex transactional language for workflows. [17] introduces an approach to add atomicity and exception handling for IBM-FlowMark. [16] proposes a workflow definition language to implement the saga model. [6] addresses atomic behavior for workflows by means of contracts. [25] proposes a model based in flexible transaction model for handling concurrency and recovery in process execution. However these approaches address transactional behavior in an ad-hoc fashion. In Web services, transactions are used to ensure sound interactions among business process. Web services transactions (WS-Tx) [3] and business transaction protocol (BTP) [12] remain as the most accepted protocols for coordinating Web services. A coordination protocol is a set of well-defined messages that are exchanged between participants of a transaction scope. However, these approaches do not offer mechanisms to ensure correctness in the specification because there is no reference model and the developer implements transactional behavior. Other approaches provide transactional frameworks. [11] introduces a model for transactional services composition based on an advanced transactional model. [1] proposes an approach that consists of a set of algorithms and rules to assist designers to compose transactional services. In [30] the model introduced in [25] is extended to web services for addressing atomicity. These approaches support the definition of atomic behavior based in the states of termination of activities with execution control flow defined a priori, which makes them not adaptable. There are other approaches, transaction policies [26], πt-calculus [2] and transactional Web services orchestration [19], that address transactional behavior to web services based on existing protocols (BTP and WS-Tx) and therefore implement partially transactional behavior without a clear separation among application logic and transactional requirements. In contrast to Web services protocols and frameworks, we consider that transactional aspects can be separated from the coordination specification and tackled using a general point of view. In this way our model fulfills the current spirit of reusing practices existing in software engineering.
7 Conclusions and Future Work The main contribution of this paper is an approach that adds transactional behavior to services coordination. For modeling transactional behavior, we introduce a model based on behavior of activities and set of activities. The contract concept is used for addressing a transactional behavior to contracts. Our model can be added to existing coordination approaches with minor changes but must important respecting the application logic. Our behavior model can be scaled because contracts can be defined recursively.
122
A. Portilla et al.
We are currently formalizing our model to provide a general framework for specifying transactional properties for services coordination and for ensuring such properties at execution time. We are also conducting the implementation of a contract management engine that can be plugged in existing orchestration engines to enact transactional orchestrations (see [24,18] for details). Finally, we will explore how our behavior model can be extended to address other properties such as isolation, and durability. Acknowledgements. This research is partially related to DELFOS project of the Franco-Mexican Laboratory of Informatics (LAFMI) of the French and Mexican governments.
References 1. Bhiri, S., Godart, C., Perrin, O.: Reliable web services composition using a transactional approach. In: IEEE International, (ed.) O. e-Technology, e-Commerce and e-Service, March 2005, vol. 1, pp. 15–21 (2005) 2. Bocchi, L., Ciancarini, P., Rossi, D.: Transactional aspects in semantic based discovery of services. In: Jacquet, J.-M., Picco, G.P. (eds.) COORDINATION 2005. LNCS, vol. 3454, pp. 283–297. Springer, Heidelberg (2005) 3. Cox, W., Cabrera, F., Copeland, G., Freund, T., Klein, J., Storey, T., Thatte, S.: Web services transaction (ws-transaction). Technical specification, BEA Systems, International Business Machines Corporation, Microsoft Corporation, Inc. (November 2004) 4. Davies Jr., C.T.: Data processing spheres of control. IBM Systems Journal 17(2), 179–198 (1978) 5. Delobel, C., Adiba, M.: Bases de donn´ees et syst`emes relationnels. Dunod, Informatique (1982) 6. Derks, W., Dehnert, J., Grefen, P., Jonker, W.: Customized atomicity specification for transactional workflows. In: Proceedings of the International Symposium on Cooperative Database Systems and Applications, pp. 155–164. IEEE, Los Alamitos (2001) 7. Doucet, A., Jomier, G. (eds.): Bases de donn´ees et internet, Mod`eles, languages et syst`eme, 1st edn. Hermes, Informatique et Syst`emes d’information. Lavoisier (2001) 8. Eder, J., Liebhart, W.: The workflow activity model WAMO. In: Conference on Cooperative Information Systems, pp. 87–98 (1995) 9. Eder, J., Liebhart, W.: Workflow recovery. In: Proceedings of the International Conference on Cooperative Information Systems, CoopIS 1996 (June 1996) 10. Elmagarmid, A.K., Leu, Y., Litwin, W., Rusinkiewicz, M.: A multidatabase transaction model for interbase. In: Proceedings of the sixteenth international conference on Very large databases, pp. 507–518. Morgan Kaufmann, San Francisco (1990) 11. Fauvet, M.-C., Duarte, H., Dumas, M., Benatallah, B.: Handling transactional properties in web service composition. In: Ngu, A.H.H., Kitsuregawa, M., Neuhold, E.J., Chung, J.-Y., Sheng, Q.Z. (eds.) WISE 2005. LNCS, vol. 3806, pp. 273–289. Springer, Heidelberg (2005) 12. Furniss, P.: Business transaction protocol. Technical specification, OASIS (November 2004) 13. Garc´ıa-Molina, H., Salem, K.: Sagas. In: ACM (ed.) 9th Int. Conf. on Management of Data, San Francisco, California, USA, pp. 249–259 (1987) 14. Gray, J.: The transaction concept: Virtues and limitations (invited paper). In: Very Large Data Bases, 7th International Conference, Cannes, France, Proceedings, September 9-11, 1981, pp. 144–154. IEEE Computer Society, Los Alamitos (1981)
Contract Based Behavior Model for Services Coordination
123
15. Gray, J., Reuter, A.: Transaction processing: concepts and techniques. Morgan Kaufmann Publishers, San Francisco (1993) 16. Grefen, P., Vonk, J., Apers, P.: Global tansaction support for workflow management systems: from formal specification to practical implementation. Very Large Data Base Journal 10(4), 316–333 (2001) 17. Hagen, C., Alonso, G.: Exception handling in workflow management systems. IEEE Transactions on Software Engineering 26(10), 943–958 (2000) 18. Hern´andez-Baruch, V., Portilla, A., Zechinelli-Martini, J.-L.: Rose: A transactional services coordination engine. In: 8th Mexican International Conference on Computer Science (ENC 2007). ENC-SMCC. IEEE, Los Alamitos (2007) 19. Hrastnik, P., Winiwarter, W.: Twso transactional web service orchestrations. Journal of Digital Information Management 4(1) (2006) 20. Leymann, F., Roller, D.: Workflow-based applications. IBM Systems Journal 36(1) (1997) 21. Milner, R., Parrow, J., Walker, D.: A calculus of mobile processes, part i/ii. Journal of Information and Computation 100, 1–77 (1992) 22. Ozsu, M.T., Valduriez, P.: Principles of distributed database systems, 2nd edn. Prentice Hall, Englewood Cliffs (1999) 23. Portilla, A.: Services coordination with transactional properties. Master’s thesis, UDLAINPG, M´exico, Grenoble (June 2006) 24. Portilla, A., Vargas-Solar, G., Collet, C., Zechinelli-Martini, J.-L., Garc´ıa-Ba˜nuelos, L., Hern´andez-Baruch, V.: Rose: a transactional services coordination engine. In: Proceedings of the 23emes Journees Bases de Donness Avancees, BDA 2007 (October 2007) 25. Schuldt, H., Alonso, G., Beeri, C., Schek, H.-J.: Atomicity and Isolation for Transactional Processes. ACM Transactions on Database Systems (TODS) 27(1), 63–116 (2002) 26. Tai, S., Mikalsen, T., Wohlstadter, E., Desai, N., Rouvellou, I.: Transaction policies for service-oriented computing. Data Knowl. Eng. 51(1), 59–79 (2004) 27. The-Open-Group: Distributed Transaction Processing: The XA Specification, X/Open Company Ltd., U.K (1991) 28. van der Aalst, W., van Hee, K.: Workflow Management, Models, Methods, and Systems, 1st edn. The MIT Press, Cambridge (2004) 29. Vargas-Solar, G., Garc´ıa-Banuelos, L., Zechinelli-Martini, J.-L.: Toward aspect oriented services coordination for building modern information systems. In: Encuentro Internacional de Computacion 2003. ENC-SMCC. IEEE, Los Alamitos (2003) 30. Vidyasankar, K., Vossen, G.: A multi-level model for web service composition. In: ICWS, p. 462. IEEE Computer Society, Los Alamitos (2004) 31. Wachter, H., Reuter, A.: The contract model. In: Elmagarmid, A.K. (ed.) Database Transaction Models for Advanced Applications, ch.7, pp. 219–263. Morgan Kaufmann Publishers, San Francisco (1992)
Transparent Admission Control and Scheduling of e-Commerce Web Services Dmytro Dyachuk and Ralph Deters Department of Computer Science, University of Saskatchewan 110 Science Place, Saskatoon, Canada [email protected], [email protected]
Abstract. Web Services are applications that expose functionality to consumers via public interfaces. Since these interfaces are defined, described and consumed using XML-based standards, Web Services outperform other middleware approaches (e.g. CORBA, RPC) in terms of platform interoperability and ease of use. Web Services support the concept of loosely coupled components, which in turn enables the development of more agile and open systems. However, this flexibility comes at the price of reduced control over the usage of the services that are exposed via the interfaces. This paper focuses on the transparent scheduling of inbound requests by introducing a proxy that prevents clients from directly accessing the provider. By manipulating the order and volume of requests sent to the provider it becomes possible to improve throughput and mean response time and to ensure consistent performance in overload situation. Keywords: SOA, Web Services, Scheduling.
1 Introduction According to the Four Tenets of Service Orientation [7], services are characterized by having clear boundaries and autonomy. Service Orientation (SO) also replaces data types as a means to describe input/output of services by using contracts and schemas and defines service compatibility by use of policies. In a service-oriented system [fig. 1], services are offered by service providers that register them with registries (e.g. UDDI). Service consumers (aka clients) discover at runtime service providers by simply queering the registries. Upon discovering a service provider, the consumer obtains from the provider the meta-data of the service that is then used to establish a binding to the provider. Since services are high-level constructs that hide implementation details, consumers can easily bind to unknown services across platform and language barriers, resulting in a system with very dynamic functional dependencies between its components. Consequently SO supports very loose coupling between consumers and providers, allowing for agile and open systems. Compared to other middleware approaches such as RPC (e.g. ONC-RPC) and object-oriented middleware (e.g. CORBA) SO differs in its lack of access transparency, since there is a very clear notion between local and remote. Service-oriented middleware (e.g. Web Services[6]) enables developers to expose functionality in terms of services, which can be described in a declarative manner ensuring interoperable and platform independence. Using IDE tools (e.g. Visual Studio [5], J. Filipe and J. Cordeiro (Eds.): WEBIST 2007, LNBIP 8, pp. 124–136, 2008. c Springer-Verlag Berlin Heidelberg 2008
Transparent Admission Control and Scheduling of e-Commerce Web Services
125
Registry (UDDI Server in Web Services)
s er ov
Pu bli sh es
sc Di
Accesses
Service Provider
Service Consumer
Fig. 1. Service Oriented Architecture
Eclipse [2]) and frameworks (e.g. Axis [1]) it is fairly easy for programmers to expose application interfaces and/or consume existing services, resulting in an ever-increasing number of Web Service deployments. However, the ease with which components (e.g. legacy systems) can now be exposed and consequently combined, raises serious concerns in regards to the dependability of the resulting system. It is important to remember that providers of services expose local resources (e.g. legacy applications) to potentially new and unknown loads. This is particularly worrisome since Web Services platforms tend to implement PS (Processor Sharing) as the default scheduling policy [11] which can easily lead to server overload situations that in turn can cause ripple effects throughout a system (e.g. faultpropagation). This paper focuses on the use of scheduling as a means to ensure that the exposed services are protected from overload situations and that the throughput and mean response time are optimized. The remainder of the paper is structured as follows. Section two presents an overview on server behaviour and scheduling. Section three presents a benchmark and an experimental setup of a system used to evaluate the scheduling of service requests. This is followed by experimentation and evaluation sections that discuss the results of the experimentation. Section six discusses related work. The paper concludes with a summary and outlook.
2 Services and Load If we assume that service providers do not share resources with other service providers (e.g. no two providers expose the same data base), than every service provider can be modelled as a server. If these providers are capable of handling multiple requests simultaneously, incoming service requests must be assigned a separate thread. Since each server has a finite amount of resources and each consumer request will lead to a temporary reduction of server-resources, it is interesting to examine the server’s behaviour under various loads. Studies [12] show that servers respond in a common way to loads. If a server is gradually exposed to an ever-increasing number of service requests it is possible to observe
D. Dyachuk and R. Deters
Throughput
126
with thrashing without thrashing Number jobs of on server under-load
saturation (peak load)
over-load
Fig. 2. Behaviour of Server
three distinct stages, under-load, saturation and over-load. At the beginning the server experiences a load that is below its capacity (under-load) and consequently it is not fully utilized. As the number of requests is increased the throughput (number of completed jobs per time unit) increased. As the rate of incoming requests increases the server experiences its saturation point (peak load). This marks the point where the server is fully utilized and operating at its full capacity. The saturation point marks also the highest possible throughput. Further increases of the rate at which the request arrive will now lead to an overload or thrashing effect. The service capacity is exceeded and “an increase of the load results in the decrease of throughput” [12]. The main reasons for a server to experience thrashing are either resource contention (overload of the physical devices e.g. CPU, memory, hard drive, etc.) or data contention (locking). Since server over-load situations lead to a decline in throughput it is important to avoid them. Heiss and Wagner [12] proposed the use of adaptive load control as the means for preventing overloads. This is achieved by first determining the maximum number of parallel requests (e.g. maximum number of simultaneous consumers) and then buffering/balking the requests once the saturation point has been reached. The impact of this approach can be seen in figure 2. The darker curve shows the characteristic three phases a server can experience, under-load, saturation and over load. Using an admission control (grey curve), the trashing is avoided due to the buffering/balking of requests above peak load. Adding an admission control into an already existing Web Services system can be achieved by use of the proxy pattern. As shown in figure 3, adding a proxy that shields/hides the original provider enables a transparent admission control. The role of the proxy is to monitor the rate at which consumers issue requests. Once the request rate exceeds the capacity of the provider, a First In First Out (FIFO) queue is used to buffer the excess request. The transparent admission control is a very effective approach for handling requests bursts [12,10]. However a basic FIFO queue ignores that different requests impact the server in different way e.g. different resource utilization. Since service-oriented middleware (e.g. Web Services) tends to favor declarative communication styles (e.g. SOAP messages), it is fairly easy to analyze the server bound traffic, determine the request and estimate the impact each request will have
Transparent Admission Control and Scheduling of e-Commerce Web Services
Proxy( Admission Control)
SOAP requests
Consumers SOAP responses
127
Service
Fig. 3. Transparent Admission Control
Scheduling SOAP request
SOAP request
SOAP request
SOAP responses
SOAP request
SOAP requests
Consumers
Admission Control
Service
Waiting queue
Fig. 4. Transparent Scheduling
on the service provider. This in turn leads to the possibility of scheduling (re-ordering) of service requests [fig. 4]. Scheduling of requests opens a broad spectrum of possibilities, like maximizing service performance in terms of interactions per time unit, minimizing variance of service response times, etc. This paper focuses on SJF (Shortest Job First) scheduling as a means for optimizing average response time. SJF (Shortest Job First) is a scheduling policy which minimizes the response time of light requests, for the price of the heavier ones. All incoming service calls are put in a waiting queue and executed in the order of their size as shown in figure four. Smith [16] proved that SJF is the best scheduling policy for minimizing the average response time if in the information on the job sizes is known and all the jobs are available at time zero. The problem of minimizing average response time in the setups where the jobs are arriving over time is proven to be NP hard [13], nonetheless due to successful deployments [10,8] SJF presents a noteworthy heuristics. In this paper we limit the discussion of scheduling to SOAP encoded, synchronous RPC style, stateless Web Services. In addition rather than using a complex resource model, requests (jobs) will be characterized by the load they create on the provider (server). The SJF scheduler is highly dependent on a correlation value of predicted and observed job length - the better the correlation the better the optimality of the schedule. While a correlation close to one (perfect predictions) achieves the optimal schedule [9], a correlation close to minus one (always wrong predictions) results in the worst schedule (instead of minimizing the response time, it will be maximized). A correlation of zero (random guess, equal amounts of correct and wrong predictions) leads to a random scheduling [9]. The SJF scheduler is highly dependent on a correlation value of predicted and observed job length - the better the correlation the better the optimality of the schedule.
128
D. Dyachuk and R. Deters
3 TPC-APP To achieve a realistic and domain independent evaluation of the transparent scheduling, the TPC-APP benchmark of the Transaction Processing Performance Council TPC was chosen. According to TPC [4] the “TPC Benchmark App [3] is an application server and web services benchmark. The workload is performed in a managed environment that simulates the activities of a business-to-business transactional application server operating in a 24x7 environment..” [4]. Since there were no free test-suites available, a reimplementation of the benchmarks was implemented following the TPC specifications.
Fig. 5. Topology of TPC-APP
As shown in figure 5, the TPC-APP scenario consists of clients, a third-party service, a bookstore service and a database server. The bookstore services and the database server are hosted on different machines (application server & DB server). TPC-App also introduces a third-party service (e.g. credit card service) to simulate external parties. The application server (bookstore service) exposes eight different methods shown in table one with their distribution in the client requests (e.g. 50% of all client calls are Create Order requests). Two methods are Writes (Create Order, Change Item), one is Read/Write (New Customer, Change Payment) and three are Reads (Order Status, New Products, and Product Detail). TPC-APP dictates that each of these methods should be treated as a transaction with ACID properties. Table 1. Bookstore Service Operations Operation New Customer Change Payment Method Create Order Order Status New Products Product Detail Change Item
Transparent Admission Control and Scheduling of e-Commerce Web Services
129
The database contained all the information operated by service, like customers, orders, inventory, etc. The size of the database was 80 Mb. The inventory table was scaled to 100 000 records, and the table describing orders and clients was proportional to the number of virtual clients. In order to reduce proxy complexity component inserted between clients and application server, the underlying communication protocol was changed from the HTTPS to HTTP.
4 Experiments Our TPC-APP implementation uses JSE as the officially allowed by TPC platform. The application server is Jakarta Tomcat 5.0.28 running on JSE 1.4. Tomcat’s default configuration is set to use a load balancing package. Load balancing was disabled to ensure a consistent behavior for the experiments (it also leads to a 15 ms speedup in average on each service call). Axis 1.3 served as framework for implementing the Web Services. All Web Services were implemented in a synchronous way and use the default HTTP 1.0 as the transport protocol. MySQL 4.1.12a was selected as the database server and the Connector J3.1.12 JDBC driver was used to link the services to the database. The application server, database server and virtual clients resided on separate machines of following configuration: Pentium IV 2.8 GHz, 2Gb of RAM. The third part services used Pentium 3, 600Mhz with 512 of RAM. The machines were connected with a 100Mb LAN and used XP SP 2 as their OS. The XP performance counters were used to collect the performance data. As the main metrics we chose throughput and average response time. The average response time is the mean value of all the response times of service during the measurement interval. Throughput is here defined as the number of successful service interactions per time unit. The measurement interval is 30 minutes. To emulate the behavior of multiple clients, a workload generator is used. The settings of the client session lengths were distributed according a discrete Beta distribution with shape parameters 1.5 and 3.0. [3]. In order to keep the number of simultaneous clients constant, new clients arrive after old clients have finished their session. According to TPC-APP the client’s business logic is ignored (takes zero time). Every client starts a new service call immediately after obtaining the results from the previous ones. The load on the service provider is a result of the number of simultaneous clients. HTTP 1.0 is the used transport protocol (closing the connection after each interaction). The timeout for the connection is set to 90 seconds. The clients are simulated by the workload generator and execute a sequence of requests that is determined by invoking a random generator at runtime (the setting represent the distribution shown in table one). The process of determining calls sequences has no memory and consequently the probability of the next operation being invoked does not depend on the previous call. Two types of experiments were conducted. The first experiment is used determine if posteriori knowledge (runtime statistics) can predict with sufficient accuracy the behavior of a job. The second type of experiments was used to determine the impact of scheduling.
130
D. Dyachuk and R. Deters
5 Evaluation 5.1 Job-Size Estimation In order to evaluate the sizes of the jobs we divided all the SOAP requests into classes. SOAP requests within the same class create approximately same loads on the service. Using the average response time of the class it is possible to predict the behavior of a request.
Cumulative Percent
100.0%
Complexity 1
80.0%
2 3
60.0%
4 5
40.0%
6 7
20.0%
8 0.0%
9 6 29
0 14
93
47
15
Fig. 6. Execution time of service methods (16 clients)
Cumulative Percent
100.0%
Method
80.0%
Change Item
60.0%
Change Payment Method
40.0%
Create Order New Customer
20.0%
New Products
0.0% 46
7
9
3 34
40
6 26
0 14
3 20
63
0
ResponseTime
Order Status Product Detail
Fig. 7. Execution time of “Order status“ with different SOAP request complexity (16 clients)
Trace studies of the TPC-APP benchmark workload show that a classification according to the name of a service operation allows achieving only a (weak) 0.35 correlation between the estimated and the actual response times. As can be seen in figure 6, some operations (“Order Status”, “New Products”) have a high variation of execution time, while operations, like “Change Payment”, exhibit more stable behavior. However, if the data contained in the SOAP messages is also used for classification a (strong) the actual response times is achieved [fig. 7]. The reason for the strong correlation is an effect of using a database centric scenario in which the costs of a service call relate to the number of updates/writes performed on the database. 5.2 Admission Control and Scheduling To obtain the data of a SOAP message is necessary to perform a partial request parsing. In order to minimize scheduling overhead and avoid double parsing we located the
Transparent Admission Control and Scheduling of e-Commerce Web Services
131
proxy component into the Tomcat framework (proxy resides on the same machine as app-server). Figures 9 and 10 show the observed throughput and response time in relation to used policy PS, FIFO, SJF and number of simultaneous clients. The first observation is that even with admission control (FIFO) and SJF, a decline in throughput can be observed. The decline is a result of the thrashing due to overload of their parsing parts. The overload situation the admission control and scheduler experience and can be easily solved/eased by hosting these components on a more resource rich host. In the current setting, the proxy that performs the scheduling/admission control is residing on a host that is shared with the application server 8.
Fig. 8. Topology of the scheduled environment
The second observation is the significant performance boost FIFO and SJF achieve in overload situations. While PS degrades as expected (at ca. 41 clients) SJF and FIFO begin to degrade only at around 76 clients. Until the load reached its peak values (41 users) the buffer queue created by the admission control is mostly empty therefore all scheduling policies behave in the same manner. After the load exceeded its peak value the excess requests were buffered and the throughput was preserved. Further load growth created a situation in which there were more elements in the queue to be scheduled so SJF and FIFO began to behave different. At higher loads SJF started outperforming
Mean response time, ms
9000 8000 7000 6000 5000
SJF
4000
PS
3000
FIFO
2000 1000
1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76
0
Number of clients
Fig. 9. Throughput of the service governed by various scheduling policies
132
D. Dyachuk and R. Deters
Throughput,iteractions/s
80 70 60 50
SJF
40
PS
30
FIFO
20 10
1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76
0
Number of clients
Fig. 10. Response times of the service governed by various scheduling policies
FIFO by 20% in average. Table 2 shows that scheduling did not affect the consumption of the resources like, memory or network. Meanwhile the CPU utilization went down from 63% to 37-40% due a lesser amount of parallel jobs. In general, low values of CPU utilizations are the result of using HTTP 1.0 since establishing a new socket causes a significant delay (up to 1 sec) and does not require the processor resources. It can also be seen that the database server’s CPU load and amount of reads/writes had highest values in case of SJF [table 3]. Thus even these loads were not fully utilizing the database. The under-utilization of the database can be explained by the low frequency of database queries issued by application server. Consequently, the application server is the bottleneck. In the current setup scheduling of the service requests mostly affected the performance of the bottleneck element - the application server. Table 2. Application server loads with 36 clients Policy Original FCFS SJF CPU, % 63 40 37 RAM, MB 557 557 558 Network Bytes/sec 271220 261629 73750
As a result of scheduling, the application server was able to produce database queries faster. Thus the database server utilization (CPU, read/write operations) increased by 10% [table 3]. Applying scheduling caused an interesting effect of “resource load
Transparent Admission Control and Scheduling of e-Commerce Web Services
133
equalization” in which the usage of overloaded resources decreased, while the utilization of lesser used resources increased. Nevertheless, this is an observed effect that requires research. It is important to note, that the skew of the response time distribution was small [fig. 11] and that the difference between the slowest and the fastest request reached only the ratio of one to six. Therefore the possibility of optimizations was lower and the bigger jobs were penalized only lightly. Figures 12 describe an example of a light [fig 12.a] and a heavy request [fig. 12].b in a medium load case (11 clients). SJF to FCFS heavy jobs experienced the 2% increase of the response time, while the response time of smaller jobs shrunk up to 45%. In any case, each of these policies significantly outperformed default PS.
Fig. 11. The distribution of the service response times (without network overhead)
350 300 250 PS
200
FIFO
150
SJF
100 50 0 Change Item
(a)
1800 1600 1400 1200 1000 800 600 400 200 0
PS FIFO SJF
Create Order
(b)
Fig. 12. “Light” and “Heavy” Operations
6 Related Work Elnikety et al. [10], present a method for admission control and request scheduling for multi-tired e-commerce Web-sites with a dynamic content in. The authors use a transparent proxy, as the intermediate scheduling component between application server and database server. The method enabled overload protection and preferring scheduling SJF augmented with aging method.
134
D. Dyachuk and R. Deters
SJF scheduling has been successfully used for http requests by Cherkassova in [8]. This work proposes the use of Alpha-scheduling to improving the overall performance. Alpha-scheduling, is a modified SJF algorithm and prevents the starvation of big requests, in way similar to aging methods [10]. The job size in [10] is estimated by classifying all http requests according a name of the requested servlet. In the domain of web services admission control research has focussed on supporting QoS differentiation (providing different levels of QoS). Siddaharta et al. [15] introduce the concept of using a Smartware platform as managing infrastructure for Web Services. The Smartware platform consists of three core components namely interceptor, scheduler and dispatcher. The interceptor intercepts the incoming requests, parses them and classifies them according to user and client (device) type. Once classified the requests are subjected to a scheduling policy the scheduler determines the order of execution. Finally the dispatcher routes the request to the service endpoint. Smartware implements a static scheduling by assigning priorities to service calls according the classification of the interceptor. As a main scheduling means the Smartware authors chose a randomized probabilistic scheduling policy namely lottery scheduling [17]. Sharma et al. [14] achieve QoS differentiation using predefined business policies. The authors examined static scheduling, in which the request class is determined by the application type, device type and client (user) level. The overall priority is calculated on the base of component priorities, for example it can be a sum or a product. The problem of starvation was eliminated by applying probabilistic scheduling. However, static scheduling does not appear to be a capable for ensuring differentiated throughput in case of fluctuating loads [14]. The authors suggest augmenting the static scheduling by altering priorities at runtime using the observed throughput. The adjustments to the required throughput are done using a correcting factor that is calculated as a penalization function of a hyperbolic nature, which speeds up slow responses and slows down fast responses in compliance with each request class.
7 Conclusions This paper presents the idea of transparent scheduling of Web Services requests as a means for achieving better performance. Using the TPC-APP, a two-tier B2B application, as a benchmark we evaluated the performance gains of SJF (Shortest Job First) compared to FIFO admission control and standard PS. By simply adding a proxy between consumer and provider, it was possible to achieve a transparent scheduling and admission control that lead to significant performance improvements in overload cases. In addition, the experimental evaluation showed that even in the absence of a priori knowledge, a SJF scheduler that uses observed runtime behaviour can lead to schedules that outperform FIFO and PS, making it an attractive approach for boosting Web Services performance. The results of the experimentation indicate that transparent scheduling can be applied to Web Services as an effective and easy to implement approach for boosting performance and avoiding service provider trashing. And while SJF does penalize larger jobs, this doesn’t seem to exceed 10% which seems acceptable given the overall gains.
Transparent Admission Control and Scheduling of e-Commerce Web Services
135
8 Future Work While the results of applying scheduling are very promising it is important to note that the current work only focused on a very simplified SOA architecture. Future work in transparent scheduling of Web Services will overcome this by addressing the following issues. 8.1 Document-Style In the current work we focused on RPC-style Web Services, that exhibit the basic request/response MEP (Message Exchange Pattern). Document-style interaction supports more complex MEPs and raises new question in regards to scheduling. 8.2 Composite Services Composite Services orchestrate the functionality provided by the other services thus creating complex environment with significantly more complex behaviour. We hope that by applying scheduling to this environment it should be possible to increase services dependability, performance, etc. 8.3 Service-Level Agreements (SLA) SLAs are an increasingly important aspect of SOA. Scheduling can be used as a means for achieving this by minimizing penalties and supporting QoS contracts in critical situations. Acknowledgements. This research has been supported by NSERC grants and the equipment provided by the Canadian Foundation for Innovation (CFI).
References 1. Apache Axis, http://ws.apache.org/axis/ 2. Eclipse, http://www.eclipse.org/ 3. TPC-App - Application Server and Web Service Benchmark, http://www.tpc.org/tpc-app/ 4. Transaction Processing Performance Council (TPC), http://www.tpc.org/ 5. Visual Studio Home, http://msdn.microsoft.com/vstudio/ 6. Web Services @ W3C, http://www.w3.org/2002/ws/ 7. Box, D.: Four tenets of service orientation. Technical report, Microsoft (2003), http:// msdn.microsoft.com/msdnmag/issues/04/01/Indigo/default.aspx 8. Cherkasova, L.: Scheduling strategy to improve response time for web applications. In: Bubak, M., Hertzberger, B., Sloot, P.M.A. (eds.) HPCN-Europe 1998. LNCS, vol. 1401, pp. 305–314. Springer, Heidelberg (1998) 9. Conway, R.W.: Theory of scheduling. Addison-Wesley Pub. Co., Reading (1967) 10. Elnikety, S., Nahum, E., Tracey, J., Zwaenepoel, W.: A method for transparent admission control and request scheduling in e-commerce web sites. In: WWW 2004: Proceedings of the 13th international conference on World Wide Web, pp. 276–286. ACM Press, New York (2004)
136
D. Dyachuk and R. Deters
11. Graham, S., Davis, D., Simeonov, S., Boubez, T., Neyama, R., Nakamura, Y.: Building Web Services with Java. Sams Publishing, Indianapolis (2004) 12. Heiss, H.-U., Wagner, R.: Adaptive load control in transaction processing systems. In: VLDB 1991: Proceedings of the 17th International Conference on Very Large Data Bases, pp. 47– 54. Morgan Kaufmann, San Francisco (1991) 13. Lenstra, J., Rinnooy Kan, A., Brucker, P.: Complexity of machine scheduling problems. Ann. of Discrete Math. 1, 343–362 (1977) 14. Sharma, A., Adarkar, H., Sengupta, S.: Managing qos through prioritization in web services. WISEW 00, 140–148 (2003) 15. Siddhartha, P., Ganesan, R., Sengupta, S.: Smartware - a management infrastructure for web services. In: Proc. of the 1st Workshop on Web Services: Modeling, Architecture and Infrastructure (WSMAI 2003), Angers, France, April 2003, pp. 42–49. ICEIS Press (2003) 16. Smith, W.E.: Various optimizers for single-state production. Naval Research Logistics Quarterly (1956) 17. Waldspurger, C.A., Weihl, W.E.: Lottery scheduling: Flexible proportional-share resource management. In: Operating Systems Design and Implementation, pp. 1–11 (1994)
Part II
Web Interfaces and Applications
Life Cases: A Kernel Element for Web Information Systems Engineering Klaus-Dieter Schewe1 and Bernhard Thalheim2 1
2
Massey University, Information Science Research Centre Private Bag 11 222, Palmerston North, New Zealand [email protected] http://isrc.massey.ac.nz/ Christian Albrechts University Kiel, Department of Computer Science Olshausenstr. 40, D-24098 Kiel, Germany [email protected] http://www.is.informatik.uni-kiel.de/˜thalheim
Abstract. On a high level of abstraction a Web Information System (WIS) can be described by a storyboard, which in an abstract way specifies who will be using the system, in which way and for which goals. While syntax and semantics of storyboarding has been well explored, its pragmatics has not. This paper contributes the first step towards closing this gap. For this we present life cases, which capture observations of user behaviour in reality. We discuss the facets of life cases and present a semi-formal way for their documentation. Life cases can be used in a pragmatic way to specify a story space, which is an important component of a storyboard. Keywords: Storyboarding, pragmatics, life case, application domain description.
1 Introduction Web information systems (WIS) augment classical information systems by modern web technologies. They aim in supporting a wide variety of users with a large diversity of utilisation stories, within different environments and with desires for personal webenhanced work spaces. The development of WIS is therefore adding the user, story and interaction dimension to information systems development. So far information systems development could concentrate on the development of sophisticated structuring and functionality. Distributed work has already partially been supported. WIS have been oriented towards a support for work. Therefore, WIS require at the same time a careful development and support for the interaction or story spaces beside the classical support for the working space of users. These dimensions complicate the system development process. On a high level of abstraction a WIS can be described by a storyboard, which in an abstract way specifies who will be using the system, in which way and for which goals. In a nutshell, a storyboard consists of three parts: – a story space, which itself consists of a hierarchy of labelled directed graphs called scenarios, one of which is the main scenario, whereas the others define the details of J. Filipe and J. Cordeiro (Eds.): WEBIST 2007, LNBIP 8, pp. 139–156, 2008. c Springer-Verlag Berlin Heidelberg 2008
140
K.-D. Schewe and B. Thalheim
scenes, i.e. nodes in a higher scenario, and a plot that is specified by an assignmentfree process, in which the basic actions correspond to the labels of edges in the scenarios, – a set of actors, i.e. abstractions of user groups that are defined by roles, which determine obligations and rights, and user profiles, which determine user preferences, – and a set of tasks that are associated with goals the users may have. In addition, there are many constraints comprising static, dynamic and deontic constraints for pre- and postconditions, triggering and enabling events, rights and obligations of roles, preference rules for user types, and other dependencies on the plot. While syntax and semantics of storyboarding has been well explored, its pragmatics apart from the use of metaphors [21] has not. Pragmatics is part of semiotics, which is concerned with the relationship between signs, semantic concepts and things of reality. Main branches of semiotics are syntactics, which is concerned with the syntax, i.e. the construction of the language, semantics, which is concerned with the interpretation of the words of the language, and pragmatics, which is concerned with the current use of utterances by the user and context of words for the user. Pragmatics permits the use of a variety of semantics depending on the user, the application and the technical environment. This paper contributes the first step towards pragmatics of storyboarding. For this we present life cases, which capture observations of user behaviour in reality. The idea is that these observations give rise to both a concrete scenario for the usage of the WIS including roles and user profiles. We then derive the candidate scenarios by abstraction. Further integration and refinement of these scenarios gives rise to the story space. We deal with the problem of extracting user models in a separate article. The third component of the storyboard, the tasks, has already been subject of strategic modelling. The life cases permit the refinement and further decomposition of these tasks.
2 Related Work Most methods for web application engineering such as OOHDM [19], WebML [3], HERA [9], WSDM [6] and variants of UML [4,11] pay little or even no attention to high-level modelling. The rationale underlying our work is that this is far too little to capture strategic and usage issues of WISs. The integration of goals and soft goals into the information systems development process has been proposed by [13,7]. The detailed distinction of intentions as targets, objects, objectives and aims is based on linguistics. The integration of the temporal dimension is necessary because of the information systems context. The extension by the representational dimensions has been made in order to specify aspects of WISs. Life cases extend and formalise scenarios used in software engineering and user modelling [5]. Early work like already discovered that human computer interfaces can be developed in techniques of movie writing or play development [22]. Life cases have already been envisioned in [2]. Scenarios as defined in software engineering are examples of interaction sessions. This kind of scenario starts with an outline of the interaction, and during requirements elicitation, details are added to create a complete description of interaction. They are specified by a general description of the start and
Life Cases: A Kernel Element for Web Information Systems Engineering
141
termination states of the system, a description of the normal flow of events, and a description of concurrent events [20]. Later scenario research has been integrated into agile programming. Customer involvement is one of the requirement of agile methods. Customers should be integrated as early as only possible into the development process [1]. Both approaches are to much oriented towards systems and do not consider life cases as we do. Scenario development has been applied to development of websites in [16]. In [8] life cases were integrated into the entire software engineering process. Recently, software engineering research developed an engineering approach to simple life case. Business use cases [15] generalise the requirements shell of [14] that is based on verbal description of event/use cases, requirement types, the rationale, originator, fit criterion, customer (dis)satisfaction, supporting materials, and the history. Our life case specification is not as complex as these business use cases. They combine intention, users, actors, constraints/assumptions/scope or context, tasks, migration, risks, costs, documentation, testing, and requirements for functions, data, look and feel, usability, humanity, operating, maintenance, support, security, culture, politics, legal and performance. We concentrate on the requirements the user has as such.
3 Facets of Intention The description of intention is based on a clear understanding of aims and targets of the WIS including a specification of long-range and short-term targets, expected visitors, characteristics of this audience, tasks performed by the users, necessary and unnecessary content, and finally an understanding of restrictions of usage. Utilisation scenarios are developed on the basis of intentions. An intention specifies what one purposes to accomplish or do, i.e. one has in mind to do or bring about. It has four facets that are based on the general characteristics of WISs: Purpose Facet: The purpose of stakeholders specifies an anticipated outcome that is intended or that guides the planned actions. It may be based on two additional pieces of information: The aims specify what is intended to be attained and what is believed to be attainable. The objectives are more general than the aims. They specify something toward which an effort is directed, e.g. goals or ends of an action. The purpose is already specified at the strategic layer. It depends on the audience intended for the WIS, and is influenced by the mission that is determined by the WIS provider. Intent Facet: The intent suggests a clearer formulation or greater deliberateness. It distinguishes between the following two aspects: The targets of stakeholders specify the steps of a plan and its termination or satisfaction conditions. The object of stakeholders is related to wishes or needs of users, and specifies an effort or activity of a user in order to satisfy the need. The intent facet is related to tasks the user wants to accomplish. The intent may be ordered into major intents specifying the main uses of the system and minor intents that may be only partially of interest. Intents may be generalized to themes that represent classes of intents.
142
K.-D. Schewe and B. Thalheim
Time Facet: The time facet is used for specification of the general time restrictions such as the design, which implies a more carefully calculated plan, and the end, which stresses the intended effect of an action, often in distinction or contrast to the action or means as such. Time facets may be very general. In this case, we use occasions to represent an entire class of time frames. Representation Facet: Intentions are described or annotated through utterances, words or pictures. The word representation is related to word fields used in the strategic model. The icon representation is based on metaphors. Word fields may be specialised to concept fields discussed later in this chapter. Representation is deeply dependent on the cultural environment and the community that uses the WIS. Intentions can be supported by providing stimuli that rouse or incite activity. Intentions may be restricted by a scope. The scope allows to concentrate on the main aspects. Typical restrictions are the cultural environment, education or other profile properties of potential users, or specific time facets for utilisation of the WIS. The first two facets of intention have a general form (objective, object) and a more concrete form (aim, target). The purpose facet depends on the mission and the audience, whereas the intent facet depends on the tasks. Therefore, the first facet specifies the ‘what’, the second the ‘how’, and the third facet the ‘when’ of an intention. We may either concentrate on a more general specification or a more concrete one. The different facets may be considered separately or altogether in a condensed form. The detailed consideration is necessary, whenever a fastidious audience requires sophisticated content. High demands come into consideration due to the content provided, the community that must be satisfied, the sophisticated functionality required for the WIS, or the high attention the WIS gains. Example 1. The purpose facet is often considered to cover ‘soft intentions’ or by ‘business rules’. In an information service a user may be interested or becoming more interested in some content. The intent facet is different and mainly driven by tasks the user tries to solve. Similarly, we may derive a number of intentions a user has in mind whenever s/he visits an edutainment site: Aim: An aim of an edutainment user may be to obtain a certificate that proofs the success of learning. Aims of providers of edutainment sites might also be binding the learner to some products such as the software of a supporter of the website. Typical general aims within an edutainment site are to grant equal rights to everybody; nobody should have higher rights than other learners. Objectives: Typical objective of using an edutainment site are greater ease of learning, greater pleasure or satisfaction during leraning, and more fun. Another general objective is security and privacy. Edutainment applications also host confidential information, e.g. information on progress and errors. The learners need to be sure of the security and privacy of their data. They should be informed of the security precautions. The audience of an edutainment site may be rather unspecific or more specific such as students of a college or analysts of a bank.
Life Cases: A Kernel Element for Web Information Systems Engineering
143
The mission of an edutainment site is to support learning by providing easy-to-grasp knowledge. At the same time, the provider may be interested in increased visitorto-customer conversion rate, increased number of returning customers, and increased revenue. The intent is related to the more concrete tasks a user has in mind while visiting a WIS. Example 2. Intents come directly from analysing the needs and the demands of users. Some possible tasks a user may want to accomplish in an edutainment site are the following: – A user realises that certain knowledge or information is necessary for solving a task. For instance, an analyst of a bank wants to know what the changes are in the customer community that led to a rise of faulty credits. – A student in a college needs to know methods for analysing data. The student is interested in a data mining course that provides material on his/her educational level, permits learning the content interactively, and is comparable to the material the student is currently working with in the college. Possible intents in an edutainment site are the following: Objects: There are a number of objects such as faster task completion, successful completion of more tasks, or commission of fewer errors. Targets: In the bank analyst case, the analyst needs to know data mining methods, to capture the achievements and the disadvantages and pitfalls of such methods. The themes in the bank analyst case could be the interest in learning association methods, preprocessing of data, and prediction methods. The time facet represents general time restrictions such as the time or the interval for visits of the WIS, the time and the interval of repeated visits, or the temporality that may force a user to leave the site. The representation facet represents the flavour or atmosphere of a site, which is largely determined by the other three facets. The set of intentions we should consider may be rather large. Moreover, the faceted representation may become too difficult to manage and to satisfy. The order we may use depends on the audience and on the provider. The audience is oriented towards solving tasks. The provider has a ‘mission’. We can use these two restrictions to figure out which kind of website supports this profile, to harmonize our understanding with the corporate identity, to order the target realisation (short-term, long-term), and finally to develop an understanding how the site will look like in two years from now on. Example 3. The intention prepare for a visit is based on a number of intents such as addressing visitors beforehand for some purpose, use, or activity. It includes the aim to put the visitor of the WIS in a proper state of mind. A task associated with preparation may be to work out the details of a plan in advance. The task may be extended by putting pieces of information together into a compound form or preparing a report. The time facet may range from ‘now’ until ‘next opportunity’. Typical metaphors supporting this intention are baskets, descriptions, and cultural journals. The intention prepare for a cinema visit extends the intention of prepare for a visit by a certain content, a refinement of the time facet to the next possible visit, and a refinement of the purpose facet now targeting at entertainment.
144
K.-D. Schewe and B. Thalheim
We can combine the description of intentions in a semi-formal way as follows. The items in the list are optional except the first one. Intention space: Purpose: Aims: Objectives: Intents: Targets: Objects: Themes: Time: Design: End: Occasion: Presentation: Atmosphere: Metaphors: Based On:
intention name outcome description list of aims list of objectives outcome description list of weighted targets list of weighted objects class of intents outcome description general flow effects, termination conditions list of objectives general style guide general description of atmosphere list of metaphors tasks, audience, mission, goal
4 Life Cases Life cases allow to overcome the information overload and lost in the hyperspace syndromes typically observed for WISs. For completion of tasks users need the right kind of data at the right moment of time, in the right dose, of the right form, in the complete extent, and within the restrictions agreed upon in advance. Moreover, users are limited in their abilities for verbalisation and digestion of data, and by their habits, practices, and cultural environment. These limitations may cause intellectual overburdening of users. Most systems that require sophisticated learning courses for their exploration and utilization did not consider these limitations and did not cope with real life situations. The approach we use for avoiding overload is based on observation of real applications before developing the system. We may extract life cases from observations in reality, which are very useful source, whenever a WIS is going to be developed from scratch or is required to provide a ‘natural’ behaviour. In the latter case users are not required to learn the behaviour of the WIS. Instead, the user can continue using a ‘classical’ behavioural pattern. Example 4. As a motivating example let us consider the life case relocation of a person, which consists of – the change of basic relocation data including the possible removal of data on the old location, – the change of official documents such as the passport, – the optional change of relation enhancements such as the registration of pets, relocation of cars,
Life Cases: A Kernel Element for Web Information Systems Engineering
145
– the change of personal specific data such as family enhancements, or relationships to religious bodies, – the change of data for additional relocation announcements such as tax, insurance changes, and – specific additional tasks such as applications for housing allowances. The person acts in the role of an issuer. We observe that relocation is enhanced by the profile of the issuer, by the specific tasks related to the relocation of the issuer, by specific laws and regulations, and by advanced functionality required for associating the life case with other life cases. from
proof of moving out
to
proof of moving in
address citzen office town clerk`s office factory inspectorate
issuer
passport
public authorities
income tax card associated documents
civil servant
special names
contracters
ministery tax office statistics agenca police aliens department TV/radio
automated contracting directory companies schools religious organizations parties
documentation agency
agencies
data protection official
special support
official bodies
recipients
pseudonyms fratemity
basic changes passport
identification
birth certificate degrees
necessary documents
certificaters
certificate of authority
actors
companies
for authorizing others pension approval certificate
proofs
special documents
employment documents marriage certificate
citizenship support of organization
driving licence child identification card
relocation address tax housing benefit house owner benefit
partners
housing allowance housing programme
children social support
housing eligibility associated parties special parking permission
forwarding period forwarding address insurance agencies health insurance tv/radio
insurance bank collect charges
potential changes potential changes
move life case
forwarding mail
pets
pet tax pets registration
supply
energy,gas water,sewage phone
address school
employment employer employment office
associated life cases contracting
vehicle documents car ownership
parking card special
documents from handicaped
housing house number
private people parties, organisation
house registration
special contracts
application for registration
obligation of secrecy registration directory accepted restrictions overruled restrictions
disclosure provided reasons
foreign resident
restrictions special/exceptional
foreign temporary second home
additional taxation
rent level
Fig. 1. Facets of the relocation life case
The life case relocation in Figure 1 consists of steps such as change of address data, change of data for associated people, change of registration data for cars, pets, etc., change of specific data, e.g. data for public authority responsible for aliens, change of data for social aid, etc. These steps are bundled together due to their relationship to one person and to one life case. The associations may be represented by adhesion of different steps, e.g. representing the association of steps by a hypergraph. The life case in the example above has overall more than 150 variations in Germany. The general target is the same everywhere. The treatment is very different. We can group these variations in around 2 dozen solutions. About half of them are already supported by WIS. The integration or matching of these WIS is far from being trivial. Moreover, the life case is associated to more than a dozen other life cases. Some of those cases are still only managed through paper exchange. This life case is one of the simplest among the life cases for proliferation and syndication of sovereign rights.
146
K.-D. Schewe and B. Thalheim
4.1 The Concept of Life Case Life cases are characterized by: Observations: We are interested in the collection and assessment of behaviour relating to the specific application. This would typically involve an observation of behaviour of users in real environments, including a background check that relates usage to intentions, goals or tasks. Processes: This involves arranging all the actions observed in an application into a main logical and coherent pattern. In some case, deviations of the main pattern must be modelled through exceptions. In most cases, we can use parallel execution of independent actions. Assessment: This involves the reconstruction of the sequence of actions and specific behaviour of users. This will aid in understanding the role each individual has within the story. It assists in developing the user profile. Individual Profiles: A list of background including physical, and behavioural characteristics of individuals is conducted. This list can also be used for deriving the most appropriate interview technique we discuss below. Interpersonal Coherence: A variation in the activity will relate to variations of other life cases. Significance of time and place: The choices made also depend on mobility, surrounding, and schedules of events of interest. Characteristics of the Life Case: Individuals using a service may be grouped by characteristics. Based on this grouping a similar behaviour is observed. Experience and Skills: Individuals may have their own experience with services provided in real life and thus use different behavioural pattern of service employment. In general, life case studies are designed to produce a picture of service employment that is as accurate as possible. Determining why, how, where, when and why a service is called using what data provides a better picture for utilisation scenario. As life cases are used for quality control, we must carefully record our observations. We either use a natural language specification or a semi-formal one as described later. Example 5. Let us consider the support of hotel search within an information service. In this case we may observe the behaviour of individuals in travel agencies while seeking for hotels. We observe that in most cases search based on associations is preferred over search by main properties such as name, address or facilities. Associations may be of a large variety, e.g. convenience to reach a hotel, location in certain maps, places of interest, or events that have caused the search. Other search criteria may be bargain or bundled offers. At the same time we observe that hotel search is combined with other intentions of users such as visiting cultural institutions. It may depend on results for search of other individuals. Another example of a life case study is the booking of train tickets depending on individuals, offers of railway companies, circumstances of individuals are using trains, etc.
Life Cases: A Kernel Element for Web Information Systems Engineering
147
4.2 Life Case Development Life cases may be developed and interpreted step by step: 1. The first step during life case collection is the survey of possible application cases we might consider. The observations could have more than one meaning and may follow a variety of user-related goals. In this case we consider the most likely meaning. 2. The second step involves a deep assessment of the life cases. We extract the different execution orders, the relationship among actions, and the roles individuals play these actions. 3. The third step extracts those life case characteristics that are distinguishing features and are relevant for time and location. At the same time we search for similarities within these life cases. 4. The final step is concerned with the characterization of potential users, their behavioural patterns, and their roles in various stages of the life case. Collectively, this information will produce a picture of the life case we are intending to support by a WIS. This may produce further life cases, or may aid in reducing the amount of development. It may result in a prioritisation of life cases under consideration, assist in the linkage of potentially related life cases, assist in assessing the potential of the WIS development, provide the WIS developers with relevant leads and strategies, and keep the WIS development on track and undistracted. These life case are mapped to scenarios in the sequel. The integration of scenarios can also be based on life cases. Example 6. Let us consider life cases of an information service for a city or a region: – Attracting visitors: The information we are providing is based on some information need visitors may have. The life case follows those that we observe during marketing activities. – Inhabitants information: The life case is based on selected information chunks that may be of interest to individuals, an ordering of the corresponding available information, and a derivable personal newspaper. – Informing tourists: The life case is similar to those observed in city information centres. – Providing official city information to inhabitants: This life case follows the message board metaphor that is used for newspaper information. During the development of city information services such as the service www.cottbus.de we have made a life case analysis for city information services and detected around 30 different life cases. We can categorize these life cases by the content and information demand of individuals. We distinguish: – life cases related to tourist content for inhabitants permanently living in the city, for inhabitants temporarily living in the city, for short-term tourists and business people on short term trip, for vacationists planning a trip for more than 5 days, for teenagers with uncertain wishes, for seniors with particular interests, for week-end visitors, for visitors of highlights, festivals, etc.;
148
K.-D. Schewe and B. Thalheim
– life cases related to official content for inhabitants on search through directory of public authority, for inhabitants on search for emergency cases, for inhabitants orienting themselves before applying at public offices, for inhabitants interested in problems of city management, etc.; – life cases related to business content, e.g. for investors considering engagement in the area. For the collection and development of life cases it is normally a good idea to interview the personnel currently providing the service. Life case should be checked against sufficy. As they may have a variety of traces, we add two stages to life case analysis: Exploration: The actual life case is provided to WIS customers and incorporated into their processes. Life case exploration can be conducted in normal environments of the user such as home or work. Then users can show the actions while they are explaining their aims and targets. Apprehension: Finally, cross checking of life cases and utilisation scenario is used for correction, extension and affirmation of the developed utilisation scenarios. During life case elicitation and specification users are confronted with sketches of scenarios. We ask what users think about these conceptualisations. The feedback is used for the refinement of life cases. We may use affinity diagramming, in which case we arrange the individual conceptions we discover on a blackboard, associate them, and discuss their relationship to the general characteristics of WISs. Another technique is based on card sorting similar to the model-view-control cards used in software engineering. In this case cards represent simple life situations. Their associations are then used for organising the conceptions. These sketches can also be used for discussing deviations from the normal case and for search of exceptional life cases. Instead of directing users to specific life cases we can show them two or more alternatives for the problem solution. Life case analysis goes beyond task analysis. It is simple and easy to carry out, lightweight and straightforward to understand, and yet quite accurate in identifying the real demands users of a WIS might have. While observing real life cases and mapping them to life cases that must be supported by the WIS we carry out a user-centered development. Additionally, we may identify and resolve conflicting or competing intentions. These conflicts can be resolved on the basis of life cases by accommodating both intentions, i.e. fulfilling the stakeholder intentions and supporting the demand of users. Besides observation of real life situations life case detection can also been based on interview techniques. Research on artificial intelligence has also resulted in a number of interview techniques: – Unstructured interviews can be considered as an informal and explorative conversation on goals a user is following and on tasks a user has to complete. Unstructured interviews observe the rules that are applied to brainstorming. All interruptions should be avoided. The questions asked must be short and simple to answer and should be open-ended questions. Interview partners should share their thoughts and
Life Cases: A Kernel Element for Web Information Systems Engineering
–
–
–
–
149
experiences. There is no judgement, confrontation or condescension. Users should not be lead toward a certain scenario. Unstructured interviews should only be interrupted if clarification is required. They need consecutive feedback in order to give the interviewees the impression that the interviewer is listening. While the interviewees are talking, everything that seems to be important is written down and used for later questions. Structured interviews are based on query plans. These plans can be based on the general characteristics of WISs. We start with easy questions on data and main functions and continue with the life cases of their utilization. Context and intentions are more difficult to capture. Structured interviews are also based on features, which can be shown to users for a rating of their importance. Life cases can also be detected by observing and critically analysing products of competitors. Elicitation of life cases from existing solutions is based on specifications, results from interviews with business users, excerpts from documents and spreadsheets, analyzed messaging and transactions, knowledge on the solutions that are currently used, meta-data and context information on the utilization within the current framework, and third party information. As reasons for restrictions cannot be captured, the copying of already existing solutions can only be used in exceptional situations. Users may also use protocols of “loud thinking”, in which case they are provided with a real-life problem of a kind that they deal with during their working life, and asked to solve it. They imagine that they are solving the problem presented to them. As they do so, they are required to describe each step and the reasons for doing what they do. The transcript of their verbal account is the protocol. In this case, they work and explain their current work. These protocols can be the basis for capturing a scenario. This interview technique should be combined with other interview techniques. There are other interview techniques that might be useful for life case elicitation such as the laddering or grid techniques. In these cases, a hierarchical structure of the application domain is formed by asking questions designed to move up, down, and across the hierarchy.
4.3 Semi-formal Representation of Life Cases Although it is sufficient for life cases to be stated in natural language, we may use a semi-formal representation instead using the following template: Life case: Characterisation: Tasks: Problems: Background: Objectives: Milestones: Content consumed: Content produced: Themes:
life case name outcome description list of user tasks list of problems general characterisation list of objectives Life case flow: general graphical graph of milestones consumed content items produced content items class of intents
150
K.-D. Schewe and B. Thalheim
Figures: actors list Expected profile: general profile description Collaboration: general collaboration description Context: general context description Time: temporality limitations Place: assignment of places Legacy: names of documents WIS: general WIS context Representation: general behavior Approaches: general description of approaches This template captures the following components of life cases: Characterisation: Life cases are characterised on the basis of strategic issues, the problem statement, background and objectives for the life case, the methodology that is used for solving the life case, and by describing the basic modules that are used for solving the life case. The characterisation is harmonized with intentions, tasks and goals. Life case flow: The life case flow combines the observations we made, the processes involved, and the data which are consumed or produced. The life case flow is mapped to a scenario. Figures: We develop profiles, especially those of individuals, as part of the WIS utilization portfolio, and interpersonal coherence specifications. Context: Time and location is explicitly described for life cases. The applicability of life cases is restricted by regulations, laws, and orders. These restrictions are seldom given in an explicit form. In addition, the context of life cases is given by the provider, the intended audience, the utilization history, and the availability of data due to the technical environment. We also use this information for the context specification. Requirements: Life cases are restricted by habits, general approaches, good practices, and boundary conditions for their application. They presuppose experiences and skills of the users involved. Note that life case we intent to support by a WIS can be completely different in real life. Sometimes we need a complete reorganisation of the business activities. In this case we should not map the real life case to a suite of associated scenarios, but rather envision a better organisation of the tasks and goals and then map these to a new envisioned hypothetical life case. Let us now outline briefly how life cases can be used in a pragmatic way to specify a story space. More precisely, we just want to derive a prototypical scenario. Other scenarios may result from other life cases, and the scenarios collected this way can be integrated and then be subject to further refinement. Let us concentrate on the life case relocation again. Example 7. In the case of the relocation of a person the steps identified in Example 4 give immediately rise to scenes in a scenario, i.e. in addition to an entry scene, say start, we obtain scenes for change of address data, change of data for associated persons,
Life Cases: A Kernel Element for Web Information Systems Engineering
151
change of registration data, change of specific data, change of data for social aid. For a start we only have a single actor citizen. In principle, the visit of any of these scenes is optional, which gives rise to a classification of actors into citizens with children, citizens with pets, foreign residents, etc. The life case of a citizen with children gives rise to a scenario for the change of data for associated persons, while the life case of a citizen with pets gives rise to a scenario for the change of specific data, etc.
5 Transforming Life Cases to Requirements 5.1 Mapping Life Cases to Business Use Cases Life cases are the major element of the application domain model. They characterise something that occurs, happens or takes place in an application or actions or instances of occurring. They may apply to a happening without intent, volition, or plan. They may denote events, incidents, episodes, and circumstances within an application. They are associated with the context description, the characterisation of the kind of the system, and the user characterisation that consists of profiles of users and of description of their intention. The description framework depicted in Figure 2 supports an elegant and sophisticated elicitation of requirements and the derivation of the presentation system specification. Application Domain Description Profiles H Users
- Intentions
H H j H -
Life case
Context Scope of the system
XXX XXX XX X z 9 ? Business use Stories Portfolio cases
Fig. 2. Elements to be specified for the application domain description
The life case study combines cognitive techniques with contextual approaches, WIS utilisation analysis and traditional approaches. This combination allows to avoid the known disadvantages of the more specific elicitation approaches. Users are characterised [18] through their profile and intentions. The profile consists of the working profile, the knowledge and abilities profile and of the psychological profile. The context specification allows to characterise the entire environment of the life case. Additionally, the solution may be restricted to certain systems and platforms. The concept of business use cases [12,15] generalises the use case concept and reflects cases or case of basic tasks in the reality. Context can be specified using the solution discussed in [10]. The storyboard describes the story space that consists of all possible stories for the given application [17]. Portfolio have been formally described in [18] and consist of tasks and the supporting instruments that are necessary. In the
152
K.-D. Schewe and B. Thalheim
sequel we decompose the life cases into business use cases and extract the portfolio. The decomposition must be invertible by a composition operation. This composition operation is the kernel of basic stories that consist of the business use cases and their story flow. We may condense this application domain description by an abstraction of users into groups depending on their general profile and their main intentions. A group is called actor. Actors can be represented by persona, e.g. ‘Jack-of-all-trade’ as a representation of a business man. The persona is characterized by an expressive name, their kind of profession, their purposes and intents, their technical equipment, their behaviour, skills and profile, disabilities , and specific properties such as hobbies and habits. Context and scope my be condensed to environment. We thus might use the elements in Figure 3. Application Domain Description - Life cases Actor
Environment
XXXX XXX XX z 9 ? Business use Stories Portfolio cases
Fig. 3. The abstractions of users, profile, intention, context and scope to actors and environment
5.2 The Algebra for Life Cases Life cases in an application domain are typically interrelated with each other, partially depend on each other, partially exclude each other, and may be (hierarchically) ordered. For instance, matters for inhabitants in a city office are relocation, registration, notice of departure, registration card, residence certificate, primary place of residence, conscript, income tax card, tariff rate for services, register of residents, archive of residents, clearance certificate and other certificates such as qualification certificate, inspection certificate, and social benefit certificate. These life cases are associated with each other. At the same time time, we may consider some of them to be basic life cases and some of them constructed from basic life cases. Whether a life case is considered to be basic depends on the application. We assume that basic life cases cannot be decomposed into separate life cases. This construction mechanism should be supported by an algebra of construction operations. We generally may assume that each life case consists of a number of components. The construction results then in a combination, replacement or removal of some of these components. Components may be essential (not removable) or auxiliary (optional). Life cases may also be related with each other. We may consider a positive relationship and a negative relationship. For instance, relocation and notice of departure are not compatible. The registration card is only issued for those that are registrable. The income tax card is based on the primary place of residence. We thus use predicates that relate life cases with each other.
Life Cases: A Kernel Element for Web Information Systems Engineering
153
We prefer binary predicates or relationship among life cases. Since life cases consist of components we do not need n-ary relationship if we wish to compare life cases with each other. The life case world (A , O , P ) consists of – basic life cases A , – algebraic operations O for computing complex cases such as combination of cases, abstraction of cases by projections, quotient of cases, renaming ρ of cases, union of cases, intersection of cases, full negation ¬ of cases, and minimal negation of cases within a given context, and – predicates P stating associations among cases such as the sub-case relation , a ∃
statement whether cases can be potentially associated with each other, a state∃
ment whether cases cannot be potentially associated with each other, a statement ∃
∃
whether cases are potentially compatible with each other, and a statement whether cases are incompatible with each other. The combination of two life cases results in a life case that has all components of the two life cases. The abstraction is used for a reduction of components of a life case. The quotient allows to concentrate on those components of the first life that do not appear in the second life case. The union takes all components of the two life cases and does not combine common components into one component. The full negation allows to generate all those components that do not appear in the life case. The minimal negation restricts this negation to some given context. We require that the sub-case relation is not transitively reflexive. The compatibility and incompatability predicate are not contradicting. The potential association and its negation must not conflict. The predicates should not span all possible associations among the life cases but only those that are meaningful in a given application area. We may assume that two life cases are either potentially associated or cannot be associated with each other. The same restriction can be made for compatibility. This life case world is very general and allows to derive more advanced operations and predicates. If we assume completeness of compatibility and association predicates we may use expressions defined by the operations and derived predicates: ∃
∃
The predicate := ∧ is used for diverging cases. ∃
∃
The predicate := ∧ is used for isolated from each other cases. ∃
∃
∃
∃
The predicate := ∧ ∧ ∧ , and is used for homogenizable cases. ∃
The predicate ⊕:= ∧ is used for heterogeneous cases. Life cases are either basic or constructed cases. We use a number of representations for visualisation of the life cases within the application domain: Life case maps are graphs consisting of nodes representing life cases and edges representing the sub-case relation and the combination of cases to complex cases. Life case hierarchies are forests representing cases and their sub-case relation.
154
K.-D. Schewe and B. Thalheim
Full life case maps are graphs consisting of nodes representing life cases and edges representing the sub-case relation and the combination, abstraction, union, intersection, quotient, full negation and minimum negation of cases if these cases are used within the application. The operations introduced above form the basis for the mapping of life cases to business use cases, to stories and to portfolio. The extraction of business use case from life cases is supported by five operations that are easing the task of efficiently acquiring relevant requirement prescriptions and of documenting them: 1. Life case projection narrows or scopes the life case to those parts (entities or concepts, axioms or invariants relating entities, functions, events, and behaviours) that are of concern for the business use cases. 2. Life case instantiation lifts the general cases to those that are of interest within the solution and instantiates variables by values which are fixed for the given system. 3. Life case determination is used for selecting those traces or solutions to the problem under inspection that are the most perspective or best fitting for the system envisioned. The determination typically results in a small number of scenarios for the life cases that are going to be supported. 4. Life case extension is used for adding those facets that are not given by the life case but are given by the environment or by the platforms which might be chosen or that might be used for simplification or support of the life case (e.g., additional data, auxiliary functionality). 5. Life case are often associated, adjacent, interact or fit with each other. Life case join is used to combine life cases into more complex and combined cases that describe a complex solution. Business use cases prescribe the “operating” part of the life cases for actors and the supporting system. Typically they are representing only one task or problem. Therefore, we determine which part of the life cases is going to be left out, which parts are emulated, and which solution is going to be preferred. The stories combine these business use cases to flows of activities. The combination is determined though the application of the five operations. The application of these operations allows to extract which subcases, which functionality, which events and which behaviour is shared among the business use cases. These shared facilities provide cross-cutting concerns among all business use cases and the exchange activities or scenes in the stories. They also hint on possible architectures of information systems and on separation into candidate components. For instance, entity sharing describe which information flow and development can be observed in the application. Functionality may be either shared by several components or may be a part of the user interface. The profiles are specifying tasks and obligations or permissions of users of the system. These profiles can be extracted from the life cases by application of these operations. The interaction between the users and the system can directly be mapped to message or life sequence charts. Additionally, we can extract quality properties. This information is later enhanced by computer requirements such as performance, dependability, maintenance, platform, and documentation requirements.
Life Cases: A Kernel Element for Web Information Systems Engineering
155
5.3 Mapping Life Cases to Requirements We can now map the three results of application domain description to properties typically used for requirements prescription. Software engineering has divided properties into functional and non-functional properties, restrictions and pseudo-properties. This separation can be understood as a separation into essential properties and non-essential ones. Functional and non-functional properties require prescription of main data structures, of the main functionality, of control facilities, and of collaboration of components of a system envisioned. Figure 4 depicts the mappings of elements of the application domain description to elements used in requirements engineering. The storyboard can be understood as a step-wise or scene-wise prescription of how a particular actor interacts with the system. Stories
Business use cases
Portfolio
Non-functional requirements
Functional requirements
XX XX XXX XXX X X XXX XXX 9 9 z X z? X ? ?
Requirements Prescription
Storyboard
Fig. 4. The use of application domain information for requirements elicitation and analysis
6 Conclusions In this paper we introduced the concept of life cases as a contribution to supporting pragmatics of storyboarding, which closes a gap in our development methodology. The general idea is to start from real life observations and to characterize them in a semiformal way. This gives rise to prototypical scenarios, tasks, roles and user profiles by abstraction. Several of such prototypes can then be integrated and refined to obtain the desired storyboard. The concept of life cases is new and original. It has already successfully been applied in our web information system projects, e.g. for information services (e.g., www. cottbus.de), for edutainment systems (e.g., DaMiT), and for community services (e.g., SeSAM). Use cases in UML [4] and business use cases [15] share some similar intentions, but are far too simplistic to capture the same information as the novel life cases. However, life cases still contribute only a part of storyboard pragmatics. They have to be complemented by user models, which should give rise to a deeper understanding of actors, and contexts. Both topics will be addressed next and published in due time.
References 1. Ambler, S.W.: Agile Modeling: Effective Practices for eXtreme Programming and the Unified Process. John Wiley & Sons, Chichester (2002) 2. Carroll, J.M. (ed.): Designing Interaction: Psychology at the Human-Computer Interface. Cambridge University Press, Cambridge (1991)
156
K.-D. Schewe and B. Thalheim
3. Ceri, S., Fraternali, P., Bongio, A., Brambilla, M., Comai, S., Matera, M.: Designing DataIntensive Web Applications. Morgan Kaufmann, San Francisco (2003) 4. Conallen, J.: Building Web Applications with UML. Addison-Wesley, Boston (2003) 5. Courage, C., Baxter, K.: Understanding your users: a practical guide to user requriements methods, tools & techniques. Morgan Kaufman, Boston (2005) 6. De Troyer, O., Leune, C.: WSDM: A user-centered design method for web sites. In: Computer Networks and ISDN Systems – Proceedings of the 7th International WWW Conference, pp. 85–94. Elsevier, Amsterdam (1998) 7. Giorgini, P., Mylopoulos, J., Nicchiarelli, E., Sebastiani, R.: Reasoning with goal models. In: Spaccapietra, S., March, S.T., Kambayashi, Y. (eds.) ER 2002. LNCS, vol. 2503, pp. 167–181. Springer, Heidelberg (2002) 8. Harel, D., Marelly, R.: Come, Let’s play: Scenario-based programming using LSCs and the play-engine. Springer, Berlin (2003) 9. Houben, G.-J., Barna, P., Frasincar, F., Vdovjak, R.: HERA: Development of semantic web information systems. In: Cueva Lovelle, J.M., Rodr´ıguez, B.M.G., Gayo, J.E.L., del Pueto Paule Ruiz, M., Aguilar, L.J. (eds.) ICWE 2003. LNCS, vol. 2722, pp. 529–538. Springer, Heidelberg (2003) 10. Kaschek, R., Schewe, K.-D., Thalheim, B., Zhang, L.: Integrating context in conceptual modelling for web information systems, web services, e-business, and the semantic web. In: Bussler, C.J., Fensel, D., Orlowska, M.E., Yang, J. (eds.) WES 2003. LNCS, vol. 3095, pp. 77–88. Springer, Heidelberg (2004) 11. Lowe, D., Henderson-Sellers, B., Gu, A.: Web extensions to UML: Using the MVC triad. In Conceptual Modeling – ER 2002. In: Spaccapietra, S., March, S.T., Kambayashi, Y. (eds.) ER 2002. LNCS, vol. 2503, pp. 105–119. Springer, Heidelberg (2002) 12. Maciaszek, L.: Requirements analysis and design. Addison-Wesley, Harlow, Essex (2001) 13. Mylopoulos, J., Fuxman, A., Giorgini, P.: From entities and relationships to social actors and dependencies. In: Laender, A.H.F., Liddle, S.W., Storey, V.C. (eds.) ER 2000. LNCS, vol. 1920, pp. 27–36. Springer, Heidelberg (2000) 14. Robertson, J., Robertson, S.: Mastering the Requirements Process. Addison-Wesley, Reading (1999) 15. Robertson, J., Robertson, S.: Requirements-Led Project Process. Addison-Wesley, Reading (2006) 16. Rosenfeld, L., Morville, P.: Information Architecture. O’Reilly, Cambridge (1998) 17. Schewe, K.-D., Thalheim, B.: Conceptual modelling of web information systems. Data and Knowledge Engineering 54, 147–188 (2005) 18. Schewe, K.-D., Thalheim, B.: Pragmatics of storyboarding for web information systems: Usage analysis. Int. Journal Web and Grid Services 3(2), 128–169 (2007) 19. Schwabe, D., Rossi, G.: An object oriented approach to web-based application design. TAPOS 4(4), 207–225 (1998) 20. Sommerville, I.: Software Enginering, 7th edn. Addison Wesley, San Francisco (2004) 21. Thalheim, B., D¨usterh¨oft, A.: The use of metaphorical structures for internet sites. Data & Knowledge Engineering 35, 161–180 (2000) 22. Vale, E.: The technique of screen and television writing. Simon and Schuster, New York (1982)
Finding Virtual Neighbors in 3D Blog with 3D Viewpoint Similarity Rieko Kadobayashi National Institute of Information and Communications Technology 3-5 Hikaridai, Seika-cho, Kyoto 619-0289, Japan [email protected]
Abstract. We propose a new mechanism for viewing entries of the 3D Blog system that enables users to find “virtual neighbors” who may have similar interests. The 3D Blog system allows a user to interact with a 3D model, annotate any spot on the model, and publish the annotations in a blog so that the user can easily communicate with other people and share interests and idea about 3D models of interest to the user. Since each annotation is saved as a blog entry with 3D viewpoint information, i.e., the user’s viewpoint and view direction as well as 3D information on annotated points, the system can detect similar entries using similarity of the 3D viewpoint information. Users can easily and intuitively compare the similar entries using the “playback” function which renders the 3D scene in the similar entries. Keywords: 3D Blog, 3D viewpoint information, Community support, Blog search, TrackBack, Viewpoint similarity.
1 Introduction The number of 3D Computer Graphics (CG) models available on the Internet is rapidly increasing due both to strong demand for 3D models in such domains as online catalogues and cultural heritage sites and to the advancement of digitizing and CG technology. For example, in the cultural heritage sector, there is an increasing demand to be able to record and visualize cultural heritage items in 3D CG models because these models have many advantages over traditional recording and presentation media such as 2D drawings. In online catalogues, 3D models can provide consumers with more detailed information and interactivity than photographs can. This means that users can change points of view as freely as if they were looking at an object in a real space, and change the texture or color of a 3D model as many times as they want to test how its appearance changes, which may actually be more difficult in real space. The number of blog users (bloggers) is also rapidly increasing. Blogs are now a very important and widespread medium for people to express and exchange ideas. One of the key features of a blog is that it allows readers to comment on blog entries, meaning that the owner of a blog site and its readers can easily communicate with each other. This bidirectional communication is facilitated by “TrackBack” mechanism which some blog systems use. J. Filipe and J. Cordeiro (Eds.): WEBIST 2007, LNBIP 8, pp. 157–168, 2008. c Springer-Verlag Berlin Heidelberg 2008
158
R. Kadobayashi
However, the use of 3D models is still restricted. Many 3D model applications provide users with some interactivity but do not provide users with a communication function for using the models. Assume that you are browsing an online catalogue and you like the design of a particular object. You may need to write a long sentence that explains why you think the design is excellent or take a snapshot of your computer display of the 3D model of the object if you want to recommend it to your friend. If the online catalogue supports communication using 3D models, the task is much easier, and you may only need to send a link to a 3D view of the object in a particular orientation and a comment to your friend. Another example in the cultural heritage sector arises when people visit a museum. Visitors can usually share the experience with their family, friends, or schoolmates, watching how they interact with exhibits. This observational or social learning is thought to be very important [1]. Sharing knowledge, interests, impressions, and experiences can deepen each person’s understanding and appreciation of the exhibits. In addition, there are a lot of hands-on exhibits in real museums that enable visitors to interact, which facilitates their understanding. However, existing 3D applications support only limited sharing and interactivity. This is especially true of applications for the general public. Experts can interact with 3D models using proprietary or expensive commercial software, but the general public is allowed only to view them passively, e.g. to see still images or animation in most cases. The experts may convey their interpretation of cultural heritage via 3D visualization, but the general public does not have any such opportunity. Of course, there are 3D applications targeting the general public such as virtual 3D museums. For example virtual Leonardo provides on-line chat functions so that users can communicate with each other [2]. The main problem in virtual 3D museums, from the viewpoint of communication and collaboration, is that users need to be in a virtual museum at the same time. Practically, this may be an overly restrictive requirement, and other means of collaboration should be provided [3]. To overcome these problems, the 3D Blog system has been proposed. This system allows a user to interact with a 3D model, annotate any spot on the model, and publish the annotations in a blog so that the user can easily communicate with other people and share interests, ideas, and questions about 3D models of interest to the user [4]. One of the main features of a 3D blog is that each annotation is saved as a blog entry with 3D information on the user’s viewpoint. Readers of a 3D Blog can reconstruct the exact same 3D scene the blogger saw when reading the blog entry, perhaps helping the readers to understand it better. In this paper, we propose a new mechanism for viewing 3D Blog entries that enables users to find similar entries easily and that intuitively calls their attention to the similarities among entries by using 3D models. It also enables them to find people who may have similar interests and to promote emergence of communities of people who share interests. The rest of this paper is organized as follows. Section 2 introduces the 3D Blog system and describes 3D viewpoint information. In Section 3, the algorithm for finding similar entries and the mechanism that links them are described. In Section 4 the algorithm for playing back similar entries is explained. Section 5 discusses the features of
Finding Virtual Neighbors in 3D Blog with 3D Viewpoint Similarity
159
Entry Snapshot
List of 3D models
Fig. 1. Top page of a 3D Blog. Annotations to 3D models are listed as blog entries in reverse chronological order. Each entry has a snapshot of the 3D scene the user saw when annotating.
the 3D Blog system that make it a very useful communication tool. Section 6 concludes the paper.
2 The 3D Blog System 2.1 Features 3D Blog enables users to interact with a 3D model, annotate any spot on the 3D model, and publish the annotations as a blog. It does not require any expensive proprietary software to make annotations or read blog entries. It only requires a 3D model viewer that runs as a browser plug-in for interacting with 3D models. The current 3D Blog system uses lily [5] as a blogging tool and Cortona VRML Client [6] as a viewer for the 3D models. Lily is a Ruby implementation of blosxom [7] written in Perl. Users need to download and install the plug-in, i.e., Cortona VRML Client, to their web browser (we are currently using Internet Explorer only) before they start 3D blogging. The main feature of 3D Blog is that it uses the 3D viewpoint information of the 3D scene that the user (blogger) sees when adding an annotation. The 3D viewpoint information is composed of viewpoint and view direction, which are expressed as 3D coordinates, and a 3D vector. The 3D viewpoint information, as well as the position of an annotated spot, is automatically added to the annotation by the system and stored in the database. It is then used to reconstruct the same 3D scene the blogger saw when a
160
R. Kadobayashi
reader reads entries. This is thought to greatly help readers to understand the content of the entries. The viewpoint is the point from which a user looks at the point being annotated. The view vector is a vector from the viewpoint to the point being annotated. The view direction is derived by normalizing the view vector. In implementation, these items are altered by the parameters specific to a computer graphics software. For example, in VRML, which we used in our system, the viewpoint and the view direction are expressed as parameters “translation” and “rotation” of a camera. This system is based on the assumption that impressions may change if the user’s point of view changes, even if the user looks at the same thing. The resulting design may enable users to convey their feelings more accurately if they can show the object as it looked when they saw it. The 3D viewpoint information is also useful for searching [8,9] and in the 3D Blog system is used to find similar entries. When a new entry is posted, the system detects similar entries based on a threshold set by users and sends links from the new entry to similar entries. We call this mechanism “automatic TrackBack.” It helps users find other users with similar interests and create communities and facilitates communication among users. To summarize, by embedding 3D viewpoint information into blog entries, the system enables users to easily understand the idea of blog entries by playing back a viewpoint of the annotated point, finding similar entries, and communicating with similar users, which facilitates forming communities. 2.2 Writing 3D Blog Entries When a 3D Blog user logs in, the top page of the user’s blogsite will open and show recent entries in reverse chronological order, as in ordinary blogsites (Figure 1). By selecting a 3D model from the list on the left-hand side of the window and clicking on the model’s name, the 3D model viewer will appear at the top of the list of entries enabling the user to browse and interact with the model in the walk-through, fly-through, or study mode and to add notes about the model (Figure 2). When the user clicks on a spot that he/she wants to annotate, a small ball, i.e. a marker, appears and covers the spot in the 3D model viewer. The user types an appropriate title and a comment in the dialogue box below the 3D model viewer and then clicks the ‘post’ button. Information, such as the position of the annotated area, the viewpoint of the virtual camera which is treated as that of the user, the date and time, the snapshot of the 3D model with the marker, and the annotation (i.e., the title and comments) are converted into a blog entry and stored in the blog database. The annotation is then added to the top of the entries of the user’s blog site as the latest entry. Users can annotate a 3D model as often as they desire. All annotations are automatically added to the blog site as separate entries. The user can change 3D models by clicking on a model’s name and continuing to blog on a different model.
Finding Virtual Neighbors in 3D Blog with 3D Viewpoint Similarity
161
3D model viewer
Text input box Entries on this 3D model
Fig. 2. Typical page layout for interacting with and annotating 3D models. 3D model viewer is displayed on top of entry list.
2.3 Reading 3D Blog At first glance, the top page of a 3D blogsite looks similar to the top pages of an ordinary blog, but there is a big difference. Every entry in the 3D blog includes information about the 3D viewpoint, i.e., the viewpoint from which the 3D model was annotated and about the spot to which the annotation was added. The snapshot included in each entry is an anchor to the 3D scene and, once it is clicked by a reader, the 3D model viewer pops up to render the 3D scene using the 3D viewpoint information while at the same time enabling the reader to read the entry (Figure 3). When rendering the 3D scene, the system first shows the 3D model in a default orientation and then changes the orientation of the model so that the 3D scene is exactly the same as the blogger saw. This animation brings to the reader’s attention how the blogger changed viewpoints when annotating. This is a great help to readers in understanding the content of entries, because it enables them to visually confirm it. Readers can add comments to an entry by clicking the ‘comment’ button at the bottom of the entry to open a dialogue box. Other basic functions such as viewing old entries and searching entries are also available.
3 Finding Similar Entries The 3D Blog system uses the 3D viewpoint information to find similar entries and to render 3D scenes and calls user attention to these similar entries by automatically
162
R. Kadobayashi
3D model viewer
Anchor to the 3D scene
Fig. 3. 3D model viewer pops up when a snapshot of an entry is clicked since a javascript is embedded as an anchor to the 3D scene. The user can read the entry and confirm how the 3D model looks.
(1) Targets are similar.
(2) Viewpoints are similar.
(3) Directions are similar.
(4) Gaze is similar.
Fig. 4. Similarity measures
Finding Virtual Neighbors in 3D Blog with 3D Viewpoint Similarity
163
setting TrackBack links. There is a blog service that connects entries that share the same keywords [10], i.e., semantic similarities. In contrast, the 3D Blog system uses the physical perspective similarities. In the current implementation, four measures are used to detect similar entries: target, viewpoint, direction, and gaze. Target is the distance between annotated points while viewpoint is the distance between viewpoints. Similarity in direction is obtained by cosine similarity of view direction vectors. Since the value of cosine theta is not linear, it is not reasonable to expect users to set the threshold. As a result, the degree of the angle between two vectors is used as a measure and users set the threshold from 0 to 180 degrees. Gaze is a combined measure of target and viewpoint. If target is less than or equal to the threshold and viewpoint is less than or equal to the threshold, the entries are similar in terms of gaze. Figure 4 depicts examples that are similar for each measure. Usually the TrackBack link is used by a blogger who writes an entry referring to another blogger’s entry to explicitly notify the other blogger of the reference. In the 3D Blog system, the system also automatically sends TrackBack to entries in the database if it finds similar entries as new entries are written. The system checks the similarities between each new entry and the stored entries whenever a new entry is posted. Then it makes the TrackBack links from the new entry to the stored entries. This means the latest entry by any user of the 3D Blog system has no automatic TrackBack links. The 3D Blog system includes the information about the similarity measures, i.e., the similarity score, as well as such information as the title and URL of the blog, which is normally included in the TrackBack ping [11]. The blogger’s own entries are excluded from the similarity check since it is unremarkable that entries by the same blogger tend to be similar. This automatic TrackBack mechanism helps 3D Blog users find similar entries and others who may have similar interests or perspectives. Users can set the threshold according to their preference. For example, some people may want to find people who have similar interests and set a low threshold for the distance between annotated spots. Others may want to find people who have a similar (physical) perspective and set that distance threshold low. 3.1 Finding Similar Entries: A Scenario Assume that user A has written an entry about a 3D model and the entry is stored in the database (see (1) in Figure 5). After user B writes an entry about the same 3D model (see (2) in Figure 5), the system checks for similar entries in the database. In this case, the system finds that an entry written by user A is similar to the entry that user B has just posted because the distance between the annotated spots of the two entries is very small and is less than the threshold specified by user A. It then sends TrackBack to user A’s entry so that user A will notice user B’s entry (see (3) in Figure 5).
4 Automatic Playback of 3D Blog Entries The 3D Blog system enables a user to see similar entries in playback mode as well as in list mode. The list mode, which is shown in Figure 6, looks similar to many blog pages.
164
R. Kadobayashi
User A
(1)
(2)
User B
(3) TrackBack
Fig. 5. Automatic TrackBack is conducted by the system when similar entries are found. Users may easily find other users with similar interests.
Similarity measures Target Viewpoint Direction Gaze
Fig. 6. List display of similar entries. Entries are sorted in ascending order of the score of the measure selected by the user. In this case entries are sorted using the score of “target” measure.
Finding Virtual Neighbors in 3D Blog with 3D Viewpoint Similarity
165
Markers
Selected user name
Fig. 7. All markers placed by a selected user in similar entries are shown when “show markers” button is clicked. Users can easily grasp the distribution of annotated points.
The user’s entry is shown on top of the blog page and similar entries written by other bloggers are shown below the user’s entry. Users can choose the sort key from the four measures. For example, if target is chosen, the entries are sorted in ascending order of the target score. In the playback mode, by contrast, the user can playback the 3D scene in each similar entry so that she or he will notice which spots on a 3D model are of interest to other users and which of those spots are annotated by other users. First, the user chooses “user name” from the pull down menu located just below the 3D model viewer. This menu shows the names of those who have made entries similar to the user’s. Second, the user chooses a function by checking either “show markers” or “play views.” The “show markers” function shows all the markers from similar entries written by the selected user, as shown in Figure 7. Playback of the 3D Blog entries in “show markers” mode thus helps users to visually grasp the distribution of other user’s interests. If the user selects the entries of another user from the pull down menu, the user can graphically see which spots she or he annotated. Alternatively, the “play views” function renders all the 3D scenes of similar entries one by one, smoothly changing the viewpoint as shown in Figure 8. This helps users see which other users saw and annotated the 3D model and how the viewpoint changed spatially. The algorithm for “play views” is as follows: First the system collects all of a particular user’s similar entries for a particular 3D model and then categorizes them by writer. Finally the system sorts the entries in reverse chronological order by writer.
166
R. Kadobayashi
Initial scene of a selected 3D model
3D scenes of similar entries
Marker
Playback button User list
Fig. 8. Playback of 3D Blog entries. Choose a user from pull down menu and then click the “play views” button. The system automatically shows the 3D scenes of the entries in sequence in the 3D model viewer.
User B's Blog
B1 User A's Blog
B2 A1 B3 A2
B4
A3
B5
A4
User C's Blog
C1 A5 C2 C3 C4 C5
Fig. 9. Relation of similar entries. In this case, user A can notice that there are links between A’s entries and B’s entries or between A’s entries and C’s entries.
Finding Virtual Neighbors in 3D Blog with 3D Viewpoint Similarity
167
Figure 9 shows an example of a relation among entries. User A wrote five entries about a particular 3D model, and three of them have automatic TrackBack links from users B and C’s entries. Although entry B2 has two links to user A’s blog, it is treated as one link. In this case, user A can choose user B or user C in addition to himself from the pull down menu, and playback user B’s entries B1, B2, B4 or user C’s entries C2, C3, C5 or his or her own entries A1, A2, A3, A3, A5 in the 3D model viewer.
5 Discussion The most essential feature of the 3D Blog system is that it uses 3D viewpoint information of a user. In other words, the system places importance on from where the user observes the 3D model as well as on to which the user has an interest. Considering a way of asking in our daily life, it is quite often that we ask a question like “what is that?” when we do not know the name of what we are looking at. Those who are talking with you can understand what is of interest to you since they can share your viewpoint and follow your glance. It is, therefore, 3D viewpoint-based search is very natural way to find things of interest. In contrast, information about what a user looking at in the 3D Blog system is similar to a keyword in keyword-based search. If a user uses keyword-based search system, the user needs to know keywords or choose keywords from a list to formulate appropriate queries. This requires the user to have knowledge a priori about what he/she will search while 3D viewpoint-based search does not require such assumption. The automatic playback function enables users to take advantage of the 3D viewpoint information easily and intuitively. As the system automatically changes the 3D scene according to the viewpoint of entries, users do not need to struggle with the 3D model to recreate the 3D scene and therefore they can concentrate on understanding similarities among entries written by other users. Although the 3D Blog system has many advantages, there are many topics that need further research and development to make it fully functional. Development of an excellent user interface for 3D models, especially for annotating and comparing 3D models, is crucial. It is difficult, intuitively and otherwise, to select an arbitrary part of a 3D model, but this is essential when adding annotation. In addition, easy-to-use interfaces that compare 3D models in a single viewer should be provided since comparing objects is fundamental to research and discussion. Functions for analysis are also important in terms of community support tools. The “show markers” function can provide an overview of a spatial distribution of annotated points. Users can know which areas are of interest to community members while they cannot know why the areas are of interest. That is functions to support semantic analysis as well as physical analysis are needed.
6 Conclusions We discussed the automatic playback function for 3D Blog entries. This function is useful especially when viewing similar entries that the 3D Blog system automatically finds
168
R. Kadobayashi
based on 3D viewpoint information since users can easily and intuitively understand the similarities among entries by viewing 3D scenes. The 3D Blog system uses 3D viewpoint information to display blog entries and to find similar entries and users. With this tool, people can easily discuss a particular topic using 3D models. The automatic playback function expands this advantage. Although further investigation is needed on additional functions of the system to support user communication and knowledge sharing via 3D blogging, we believe 3D blogging with the automatic playback function has great potential and can be applied to many domains.
References 1. Falk, J.H., Dierking, L.D.: Learning from Museums: Visitor Experiences and the Making of Meaning. AltaMira Press (2000) 2. Paolini, P., Barbieri, T., Loiudice, P., Alonzo, F., Gaia, G., Zanti, M.: Visiting a Museum together: How to share a visit to a virtual world. In: Proceedings of Museums and the Web 1999. Archives & Museum Informatics, pp. 27–35 (1999) 3. Barbieri, T., Paolini, P.: Cooperative Visits for Museum WWW Sites a Year Later: Evaluating the Effect. In: Proceedings of Museums and the Web 2000. Archives & Museum Informatics, pp. 173–178 (2000) 4. Kadobayashi, R.: Automatic 3D blogging to support the collaborative experience of 3D digital archives. In: Fox, E.A., Neuhold, E.J., Premsmit, P., Wuwongse, V. (eds.) ICADL 2005. LNCS, vol. 3815, pp. 109–118. Springer, Heidelberg (2005) 5. Lily, http://lily.sourceforge.jp/ 6. ParallelGraphics. Cortona VRML Client, http://www.parallelgraphics.com/products/cortona/ 7. Blosxom, The zen of blogging (2006), http://www.blosxom.com/ 8. Kadobayashi, R., Tanaka, K.: 3D viewpoint-based photo search and information browsing. In: Proceedings of the 28th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pp. 621–622. ACM, New York (2005) 9. Fujita, H., Arikawa, M.: A ubiquitous photo mapping considering users’ lines of sight. In: Proceedings of International Workshop on Ubiquitous Data Management (UDM 2005), pp. 71–77. IEEE Press, Los Alamitos (2005) 10. hatena, Hatena diary keyword (2007), http://d.hatena.ne.jp/keyword/ 11. SixApart. Trackback technical specification, http://www.sixapart.com/pronet/docs/trackback spec
Model Driven Formal Development of Digital Libraries Esther Guerra1, Juan de Lara2 , and Alessio Malizia1 1 2
Abstract. This paper shows our model-driven approach for the formal construction and validation of Digital Libraries (DLs). We have defined a Domain Specific Visual Language (DSVL) called VisMODLE, which allows the description of a DL using five different viewpoints: services, behaviour, collections, structure and society. From a meta-model based description of the different viewpoints, we have generated a modelling environment for VisMODLE. We have provided the environment with a code generator that produces XUL code for the DL’s user interface and composes the application using predefined components that implement the different services. Moreover, we have also added validation and simulation capabilities to the environment. Using the behavioural models (statemachine based), we can visually animate the system. In addition, the combined behaviour of actors and services can be transformed into a Petri net for further analysis. Keywords: Digital Libraries, Model Driven Development, Formal Methods, Meta-modelling, Graph Transformation.
1 Introduction The concept of Digital Library (DL) seems hard to be completely understood and evades definitional consensus. Licklider [10] visualized a collection of digital versions of the worldwide corpus of published literature and its availability through interconnected computers. More recently, a Delphi study [9] of DLs coalesced a broad definition: organized collection of resources, mechanisms for browsing and searching, distributed networked environments, and sets of services objectified to meet users’ needs. The Presidents Information Technology Advisory Committee (PITAC) Panel on DLs treats them as the networked collections of digital texts, documents, images, sounds, scientific data and software, that make up the core of today’s Internet and tomorrow’s universally accessible digital repositories of human knowledge [14]. Underlying these definitions there is the consensus agreement that DLs are fundamentally complex due to its inherently interdisciplinary nature. They are usually built from scratch using specialized architectures that do not benefit from previous DLs and software design experiences. The lack of formal models leads to branching efforts and has made interoperability (in both metadata and software levels) a crucial problem in the DL field. From the beginning of computer science, software engineers have sought methods to increase the productivity and quality of applications. A means to achieve such goal J. Filipe and J. Cordeiro (Eds.): WEBIST 2007, LNBIP 8, pp. 169–183, 2008. c Springer-Verlag Berlin Heidelberg 2008
170
E. Guerra, J. de Lara, and A. Malizia
is to increase the level of abstraction of system descriptions. In Model-Driven Software Development (MDSD) [15], models are the primary asset, from which the code of the application is generated. The idea is to capitalize the knowledge in a certain application domain by providing developers with DSVLs describing the domain concepts. DSVLs are less error-prone than other general-purpose languages and easier to learn because the semantic gap between the user’s mental model and the real model is smaller. Thus, from high-level, possibly visual models, a number of platform artifacts are generated which sometimes account for the 100% of the final application. Since the code generation process is automated from the models, there is a strong need to validate and verify them. In this respect, techniques for simulating the models (i.e. performing an animation showing their operational semantics) as well as to transform them into formal semantic domains for further analysis are of great interest. Our work applies MDSD techniques to the DL domain. In [11] we presented a DSVL called VisMODLE oriented to the description of DLs, and built a modelling tool for it with an integrated code generator. VisMODLE is visual, which makes easier for people to learn and interpret; it is domain-specific, leaving less room for misunderstanding; and is formal, making possible the simulation and validation of the models. It allows the specification of the different aspects of a DL using four dimensions (or diagrams): services, collections, structure and society (i.e. interactions of services and actors). In this work, VisMODLE is extended with a new type of diagram to specify behaviour (based on state machines). The operational semantics (i.e. simulator) of the new diagram type is formalized by using graph transformation [5]. This simulator allows a visual animation of the models (i.e. seeing messages being produced/consumed by actors and services) in order to validate the DL design and understand its behaviour. We have also designed a transformation from VisMODLE to Petri nets [13], a formal notation useful to specify concurrent and distributed systems. Its analysis techniques allow investigating system properties such as deadlocks, state reachability and invariants. Paper Organization. Section 2 introduces VisMODLE. Section 3 shows our MDSD approach for generating DLs. Sections 4 and 5 describe the simulation and analysis techniques we propose, and its application to VisMODLE. Section 6 presents related work, and section 7 ends with the conclusions and future work. As a running example, we develop a simple DL for a university library.
2 VisMODLE, a DSVL for Digital Libraries VisMODLE is a DSVL for the Visual MOdelling of Digital Library Environments. A preliminary version of the language was presented at [11], but we have added a new diagram type for expressing behaviour. In this way, the specification of a DL in VisMODLE encompasses now five complementary dimensions: multimedia information supported by the DL (Collection Model); how that information is structured and organized (Structural Model); the services of the DL (Service Model); the societies of actors and services that interact to carry out the DL behaviour (Societal Model); and the individual behaviour of actors and services (Behavioural Model). Figure 1 shows the complete meta-model.
Model Driven Formal Development of Digital Libraries
Collections are sets of static (e.g. text) and dynamic (e.g. video) elements. Our running example includes a collection model (not shown for space constraints) with a collection (called Library) of two documents: long1.pdf and long2.pdf. The only relevant entity for this kind of diagram in the meta-model is class Collection. The Structural diagram specifies how parts of a whole are arranged or organized. Structures can represent hypertexts, taxonomies, system connections, user relationships and containment. The window to the right in Figure 2 shows the structural model for the running example. Thus, collection Library is made of documents structured with Publication, Author and Title metadata information. The meta-model specifies that metadata entities (class Struct) can be connected together with the node relation (organized as a tree) and linked to a collection by a link type relation. Services describe activities, tasks, and operations (the functionality). Human information needs, and the process of satisfying them in the context of DLs, are well suited
Fig. 2. Modelling Environment for VisMODLE
172
E. Guerra, J. de Lara, and A. Malizia
to description with services, including fact-finding, learning, gathering and exploring. In this diagram type, we describe the kind of messages services can produce and consume, and their synchronization. The only relevant entity from the meta-model for this diagram is class Service. In our example, we make two services available: FrontDesk and DoSearch. The former is responsible for managing communications between actors and is asynchronous (sync attribute is set to nowait), while the latter executes queries on the DL and is synchronous. A Society is a set of entities and their relations. The entities include actors as well as hardware and software components, which either use or support services. Our metamodel is CSCW (Computer Supported Cooperative Work) oriented and thus includes the explicit description of the communication between entities and services. In complex systems such as DLs it is important to consider also the actors involved in the usage and so we introduced the society concept. A society is the highest-level view of a DL, which exists to serve the information needs of its entities and to describe the context of its use. Society diagrams are similar to UML collaboration diagrams. Figure 3 shows the society model for our example involving actors Student and Librarian. The scenario represents a Student borrowing a paper from the Library; he interacts with the FrontDesk service requesting the paper and obtaining a response message about its availability. The FrontDesk service forwards the borrow request to the Librarian actor. Then it sends a doc request message to the DoSearch service, which queries the document collection (get operation) using metadata information provided by the borrow request, and waits the result to send back the response. The service returns an is available boolean message which is propagated as a response to the Librarian and eventually to the Student.
Fig. 3. A Societal Model
Behavioural diagrams are used to specify the individual high-level activity of both services and actors by means of state machines. The transitions of the state machines are fired whenever a certain event occurs, which corresponds to the arrival of a message to the actor or service. In the transition, an action can also be specified, which is the sending of another message. Figure 4 shows the behavioural diagram for the Librarian (initially in state active). If he receives an event borrow request, he sends a doc request message to service DoSearch and changes his state to borrow request.
Model Driven Formal Development of Digital Libraries
173
Then, if he receives an event is available, he becomes active again and sends the availability to service FrontDesk. For the current DL example, we have defined four different behavioural models, for actors Student and Librarian as well as for services FrontDesk and DoSearch.
Fig. 4. A Behavioural Model
3 Model Driven Approach to Digital Libraries The overall architecture of our MDSD approach for DLs is shown in Figure 5. The upper part depicts the process of designing the DSVL VisMODLE with the help of experts in the field of DLs, and its implementation in the meta-modelling tool AToM3 . The environment for VisMODLE was defined by using the meta-modelling capabilities of AToM3 . Figure 6 shows a step in its definition. The window at the background (labelled “1”) partially shows the complete VisMODLE meta-model. The tool allows splitting the meta-model into different diagram types (called viewpoints). The five VisMODLE viewpoints are shown in the window labelled “2” in the figure. Window “3” contains the portion of the meta-model belonging to the Structural viewpoint. A special viewpoint named repository DigitalLibrary contains the complete meta-model. In this way, in the generated environment, the user builds instances of the different viewpoints, and a repository model is created in the background with the gluing of the different diagrams the user has built. Consistency relations (shown as arrows in window “2”) specify how the different diagram elements are copied into the repository, how the different diagrams are kept consistent and how changes are propagated to the other diagrams if necessary. These arrows contain triple graph grammar (TGG) rules, which are automatically generated by AToM3 [8]. In addition, a semantic view has been defined that allows expressing the semantics of the repository by using a transformation into Petri nets (see section 5). This is also performed by using TGG rules. With the information described before, AToM3 generated a customized modelling environment for VisMODLE (shown at the background of Figure 2). DL designers can use this environment to build VisMODLE models (layer “DSVL use” in Figure 5). The environment also integrates a simulator that allows the visual animation of the models. The simulator helps to validate and understand the DL design by observing
174
E. Guerra, J. de Lara, and A. Malizia
Fig. 5. The MDSD Architecture for DLs
Fig. 6. Building the Environment for VisMODLE
the message flow between services and actors (see section 4). As stated before, some analysis capabilities have been also integrated based on model transformation into Petri nets (see section 5). Finally, we have provided the environment with a code generator (called “LibGen” in Figure 5) able to produce XUL code for the user interface (UI) of the DL. The generator also selects predefined service components to implement the required functionalities of
Model Driven Formal Development of Digital Libraries
175
the VisMODLE models [11]. Figure 7 shows the generated interface for the example (for the Librarian). The generated UI is built upon a set of XUL template files that are automatically specialized depending on the attributes and relationships designed in the modelling phase. The layout template for the UI is divided into two columns. The left part manages the collections of documents and its metadata information. The right part manages visualization and multimedia information obtained from documents. The basic features provided with the UI templates are document loading and visualization, metadata organization and management.
Fig. 7. Generated GUI for the Example DL
For the composition of the services specified in the VisMODLE models, we have used XDoclet [17]. This is an open-source code generation library which enables Attribute-Oriented Programming for Java via insertion of special tags. It includes a library of predefined tags, which simplify coding for various technologies such as web services.
4 Simulation of DL Models In order to visualize and better understand the behaviour of a DL, we have built a simulator for VisMODLE. The simulation is performed in the repository, since it needs all the dynamic information expressed in VisMODLE (i.e. the societal and behavioural models). The simulation consists of the animation of the state machines that describe the behaviour of actors and services. In the simulation, it is checked that the events and actions specified in the state machines are coherent with the message interchange specified in the society. Note how simulation is a useful tool for the early detection of design errors.
176
E. Guerra, J. de Lara, and A. Malizia
The repository has been enriched with special elements used only for simulation purposes, but not for system specification. These elements are not part of VisMODLE; therefore, they do not appear in any diagram type. Thus, behavioural elements (i.e. actors and services) can receive messages through a special element called input, which in addition has a pointer to the current state of their state machine. Messages are depicted as blue rectangles with its name inside, while input elements are shown as black small rectangles. In the simulator, for a transition to be executed, a message matching the event specified in the transition has to be found in the behavioural element’s input. The other possibility is that the transition does not require any event. On the other hand, when a transition that defines an action with form “target element.message” is executed, then such message is created in the target element’s input. Figure 8 shows the DL example being simulated. A student has asked for a book, so he is in state student borrow (i.e. the input for actor Student has a pointer to the student borrow state). That means that the transition going from state student active to the current state has been executed, so a message borrow has been sent to service FrontDesk, as the action of such transition indicates. Indeed, the state machine for service FrontDesk (upper left corner) has a message borrow as input. The next simulation step would execute the transition going from the service’s current state to state front desk borrow, since the transition requires an event borrow, available in the input.
Fig. 8. A Step in the Simulation of the Example DL
This simulator has been built by using the graph transformation capabilities of AToM3 . Graph transformation [5] is an abstract, declarative, visual, formal and high-level means to express computations on graphs. Roughly, graph grammars are composed of rules, with graphs in their left and right hand sides (LHS and RHS respectively). In order to apply a rule to a graph (called host graph), first a matching (an occurrence) of the LHS has to be found on it. Then, the rule is applied by substituting
Model Driven Formal Development of Digital Libraries
177
the match in the host graph by the rule’s RHS. In addition, rules can define application conditions that restrict their applicability. One of the most used are the so-called Negative Application Conditions (NACs). These are graphs that must not be present in the host graph for the rule to be applied. On the contrary, Positive Application Conditions (PACs) are graphs that must be present in the host graph in order to apply the rule. Finally, we can combine meta-modelling and graph transformation allowing abstract nodes to appear in rules [5]. In this way, nodes can be matched to instances of any subclass, greatly improving the expressive power of rules. Our simulator is made of five rules, three of them shown in Figures 9, 10 and 11. Rule in Figure 9 assigns an input element to each behavioural element, that is, to each actor and service in the repository (i.e. each possible concrete subclass of class BehaviouralElement). Initially, the input does not contain any message, and points to the initial state of the behavioural element’s state machine. If the behavioural element already has an associated input element (NAC), the rule is not applied. In this way, we are sure there is only one input for service and actor.
Fig. 9. Simulation Rule create input
The other four simulation rules perform a simulation step (i.e. a state transition if the required event was produced) for different cases. For example, the rule in Figure 10 considers transitions where no action is defined. In this case, if a behavioural element receives a message whose label matches some of its outgoing transitions (LHS), then the message is processed and the current state is changed (RHS). The attribute condition in the LHS checks that the message name in the input is equal to the event specified in the outgoing transition of the current state.
Fig. 10. Simulation Rule process event
178
E. Guerra, J. de Lara, and A. Malizia
Similarly, rule in Figure 11 considers transitions with an action that sends a message to another behavioural element (i.e. the action has the form “target element.message”). Here, in addition to the previous conditions for a behavioural element to change its state, it is necessary that the target element accepts messages of that kind. This is given by the PAC below, which checks that the corresponding link has been specified in the societal model. Indeed, one of these three positive conditions has to be fulfilled in order to be able to apply the rule: either the source element is an actor that sends a message to a service (first case), or it is a service that sends a message either to an actor (second case) or to another service (third case). The use of abstract nodes (labels 8 and 11) together with the three PACs allows using a single rule for the three cases. This rule deletes the message from the input, changes the current state and creates a message in the target element input. The action below the RHS assigns the message specified in the action’s transition as the name of the newly created message.
Fig. 11. Simulation Rule process event with action
Finally, two similar rules (not shown in the paper) consider transitions where either an action is performed without the necessity of an event, or a transition occurs with neither event nor action associated.
5 Analysis of DL Models We have provided the VisMODLE environment with some analysis mechanisms to validate the dynamics of a DL specification. More in detail, given the set of models that conform a DL design, we can:
Model Driven Formal Development of Digital Libraries
179
1. check if some behavioural element reaches a deadlock state (i.e. local deadlock). In the case of services, we normally want to keep them always available, so they should not define final states. Final states are shown highlighted in the model as result of the analysis, and they are presented textually in a dialog window as well. 2. check whether a given behavioural element reaches certain state. The sequence of state transitions leading to the requested state is shown highlighted as a result, and also textually. 3. ask if a given message is always sent in any possible execution flow. The answer (true or false) is shown in a dialog window. 4. check if the system execution always finishes (i.e. global deadlock). 5. detect if there is some state for a behavioural element that is never reached in the given society. This may be considered a design error, since we should not define states that are not possible. 6. check if some behavioural element can receive an unbounded number of messages at some point, which could lead to an overflow. In order to provide these analysis mechanisms, we have expressed the operational semantics of the models by means of Petri nets [13] (using a semantic view, see Figure 6). Thus, we have defined a triple graph transformation system (TGTS) [7] that transforms the repository into the equivalent Place/Transition Petri net. Once the net is obtained, we internally use analysis techniques based on the reachability/coverability graph, as well as model-checking [3]. Similarly to graph grammars, TGTSs are made of rules that contain, in this case, triple graphs (instead of simple ones). Triple graphs are made of three different graphs: source, target and correspondence. In our case, the source graph is the repository, the target graph is the Petri net resulting from the transformation, while the correspondence graph contains elements relating the elements in the other two graphs. Figures 12 and 13 show some rules of the TGTS that build the Place/Transition Petri net from the VisMODLE repository.
Fig. 12. Triple Rule SM state to PN place
The idea of the transformation is transforming the state machines of each behavioural element into what we call Petri net modules. With this purpose, states are translated into places (triple rule shown in Figure 12) and transitions between states are translated into
180
E. Guerra, J. de Lara, and A. Malizia
Fig. 13. Triple Rule SM transition to PN transition
transitions between places (triple rule in Figure 13). If a state is initial, then the associated place will contain one token. In other case, the place will be empty. In addition, an extra place is created for each possible message invocation (i.e. for each possible event and action specified in each state machine). These places are the interface of the Petri net module of the behavioural element. If a transition needs an event for being executed, then an arc is created from the place corresponding to the event message, to the Petri net transition corresponding to the state machine transition. In that case, the Petri net transition can be executed only if an event of that type is received (i.e. a token is in the right interface place), and in addition, the behavioural element is in the right state (i.e. a token exists in the place corresponding to the state source of the transition). Similarly, if a transition specifies an action, an arc is created from the Petri net transition to the right place of the behavioural element’s interface specified in the action. In this way, executing the transition implies creating a token in the right interface place, and results in the interconnection of the different modules. Figure 14 shows the resulting Petri net after applying the TGTS to the DL example. Each analysis provided to the VisMODLE modelling environment implies evaluating a CTL logical expression on the net’s coverability graph [3]. With this purpose, we use a function that calculates the coverability graph of the net, as well as a model checker
Fig. 14. Place/Transition Petri Net for the DL Example
Model Driven Formal Development of Digital Libraries
181
to evaluate the expression, both implemented in AToM3 . In order to use these analysis techniques, users of the VisMODLE environment do not have to know Petri nets at all. All the analysis process is hidden. CTL expressions are defined when the modelling environment is generated, and hidden to the final user. The results of analysing the Petri net are back-annotated and shown to the user in the VisMODLE notation, which is the notation that he knows. This is possible since we maintain in the correspondence graph the relations between the elements in the DL and the ones in the Petri net. AToM3 provides declarative back-annotation mechanisms that allow the specification of the elements to highlight in the source model as a result of the analysis.
Fig. 15. Execution of Analysis Method Call
For example, Figure 15 shows the repository interface of the generated VisMODLE environment. The repository interface includes one button for each one of the defined analysis method calls. In order to check a property, the end-user has to open the repository and click on one of these buttons. Internally, the tool performs the transformation to the semantic domain, evaluates the CTL formula that corresponds to the property on the Petri net’s coverability graph, and returns the result according to the defined back-annotation mechanism. In the figure, the analysis called reachability for state front desk borrow is shown. The window to the right allows navigating through the different solutions (i.e. the different path executions) that lead to such state. The VisMODLE elements that conform a single solution are shown in the navigator and highlighted in the model. In this way, the analysis result is shown in terms of the original notation to the user, who does not need to have knowledge about any formal method. For the example DL, the net showed neither global nor local deadlocks, all the states were reachable, and the tool signalled the possibility of unbounded number of borrow and borrow request messages.
182
E. Guerra, J. de Lara, and A. Malizia
6 Related Work Formal models for DLs are rarely found, likely due to the complexity of the field. The approach of [16] defines DLs as a combination of a special-purpose database and a hypermedia-based UI, and formalizes them by using the Z language. In [2] a multidimensional query language for DLs is used that rely on the notions of views and versions, metadata formats and specifications, and a first-order logic based language. [6] presents a formal foundation theory on DLs, called 5S, based on streams, data structures, spaces, scenarios and societies. We were inspired on it in order to design VisMODLE. However, these formal approaches do not formally specify how to derive the DL implementation from the DL model. Other declarative approaches, such as the Digital Library Definition Language [12], the METIS framework [1] and the FEDORAs structoid approach [4], are not supported by a strict underlying formal theory. To the best of our knowledge, none of these approaches provide a customized environment supporting code generation as well as validation (animation) and analysis techniques. By using MDSD techniques, we help DL experts to cope with the complexity of DL designs without dealing with coding. In addition, the use of a DSVL makes it easier for people to learn the domain concepts and interpret the models, and allows modelling interactions among DL systems and users (as proposed in the HCI field). Finally, our approach generalizes some metadata schemas such as DC, in the sense that DC data structures can be modelled by using our structure entities and their relationships.
7 Conclusions In this paper, we have presented VisMODLE, a DSVL for building DLs in a modeldriven way. The language is made of five diagram types (services, collections, structure, society and behaviour) to describe the aspects of a DL. We have generated a modelling environment for it with the AToM3 tool. The environment integrates a code generator in order to produce the DL. Although it is in its alpha version, we have already used it to build prototypes. In addition, we have built a visual simulator using graph transformation to graphically validate the DL behaviour at the model level. In order to perform further analysis, a transformation into Petri nets has been designed, which allows checking model properties, such as reachability or deadlocks. This analysis is made by internally performing model-checking of the coverability graph by using predefined temporal logic formulae. In the future, we intend to take into account scalability issues, as well as to generate code for some orchestration language to reflect the overall behaviour of the societal model. Acknowledgements. Work sponsored by projects MODUWEB (TIN2006-09678) and MOSAIC (TIC2005-08225-C07-06) of the Spanish Ministry of Science and Education.
Model Driven Formal Development of Digital Libraries
183
References 1. Anderson, K.M., Andersen, A., Wadhwani, N., Bartolo, L.M.: Metis: Lightweight, flexible, and web-based workflow services for digital libraries. In: JCDL, pp. 98–109 (2003) 2. Castelli, D., Meghini, C., Pagano, P.: Foundations of a multidimensional query language for digital libraries. In: Agosti, M., Thanos, C. (eds.) ECDL 2002. LNCS, vol. 2458, Springer, Heidelberg (2002) 3. de Lara, J., Guerra, E., Vangheluwe, H.: Metamodelling, graph transformation and model checking for the analysis of hybrid systems. In: Pfaltz, J.L., Nagl, M., B¨ohlen, B. (eds.) AGTIVE 2003. LNCS, vol. 3062, pp. 292–298. Springer, Heidelberg (2004) 4. Dushay, N.: Using structural metadata to localize experience of digital content. CoRR, cs/0112017 (2001) 5. Ehrig, H., Ehrig, K., Prange, U., Taentzer, G.: Fundamentals of Algebraic Graph Transformation. Springer, Heidelberg (2006) 6. Gonc¸alves, M.A., Fox, E.A., Watson, L.T., Kipp, N.A.: Streams, structures, spaces, scenarios, societies (5s): A formal model for digital libraries. ACM Trans. Inf. Syst. 22(2), 270–312 (2004) 7. Guerra, E., de Lara, J.: Attributed typed triple graph transformation with inheritance in the double pushout approach. In: Technical Report UC3M-TR-CS-06-01, Universidad Carlos III de Madrid (2006) 8. Guerra, E., de Lara, J.: Model view management with triple graph transformation systems. In: Corradini, A., Ehrig, H., Montanari, U., Ribeiro, L., Rozenberg, G. (eds.) ICGT 2006. LNCS, vol. 4178, pp. 351–366. Springer, Heidelberg (2006) 9. Kochtanek, T.R., Hein, K.K.: Delphi study of digital libraries. Inf. Proc. Manag. 35(3), 245– 254 (1999) 10. Licklider, J.C.R.: Libraries of the Future. MIT Press, Cambridge (1965) 11. Malizia, A., Guerra, E., de Lara, J.: Model-driven development of digital libraries: Generating the user interface. In: Proc. MDDAUI 2006 (2006) 12. Maly, K., Zubair, M., Anan, H., Tan, D., Zhang, Y.: Scalable Digital Libraries Based on NCSTRL/Dienst. In: Borbinha, J.L., Baker, T. (eds.) ECDL 2000. LNCS, vol. 1923, Springer, Heidelberg (2000) 13. Murata, T.: Petri nets: Properties, analysis and applications. Proc. of the IEEE 77(4), 541–580 (1989) 14. Reddy, R., Wladawsky-Berger, I.: Digital libraries: Universal access to human knowledge – a report to the president. In: PITAC, Panel on DLs (2001) 15. V¨olter, M., Stahl, T.: Model-Driven Software Development. Willey (2006) 16. Wang, B.: A hybrid system approach for supporting digital libraries. JDL 2(2-3), 91–110 (1999) 17. XDoclet home page, http://xdoclet.sourceforge.net/xdoclet/index.html
Predicting the Influence of Emerging Information and Communication Technologies on Home Life Michele Cornacchia1 , Vittorio Baroncini1 , and Stefano Livi2 1
2
Fondazione Ugo Bordoni Via Baldassarre Castiglione 59, 00142 Rome, Italy [email protected], [email protected] Facolt`a di Psicologia 2, Universit`a degli Studi “La Sapienza” Via dei Marsi 78, 00185 Rome, Italy [email protected]
Abstract. Whether or not ICT represents the most important vehicle to transform the society seems to be out of discussion. The point of interest diverts from how people do really feel with these services and from the way they perceive the advantages as acceptable to improve the quality of life and work. It is matter of fact that the technical innovation is characterized by a certain risk, the problem of how to implement the technology for sure and, ahead of this phase, the problem of predicting its influence on the social, working and private life in view of the high costs effort to produce. This study applies a predictive model for the acceptance to a services integrated home environment properly set-up in a special laboratory. A class of users was selected from the employees of the company which hosted the trial in order to participate at the evaluation sessions. The tasks were designed to point out the main innovative features of the services presented. The questionnaires were suitably designed and submitted to collect the end-users opinions. Statistical analysis was carried out to assess the performance by the side of the real users and to predict their intentions of use. Finally both correlation and regression studies found out how the model predictors may respectively influence the acceptance of the ICT solutions proposed. Keywords: Acceptance, assessment, controlled environment, ease of use, end users, evaluation, home control, home networking, home services, integration of ICT, intention of use, interoperability, laboratory home platform, laboratory trial, perceived usefulness, predictive model of user, services integration, usability.
1 Introduction The study here presented is part of the work carried out in order to investigate the user perception of the ePerSpace [1] personal services for the Home and Everywhere that were set up at the laboratories of a big telephone company, partner in the project. The general aim was to measure the quality of the delivered services, by verifying the usefulness and ease of use as perceived by the real users, then the amount of added value provided by each service, even in a high technology reproduced environment. The basic references given by the Unified Theory of Acceptance and Use of Technology [2] were applied to define a model to forecast the user acceptance (intention as predictor of usage) and arrange the scales (questionnaires) to measure the performance constructs. J. Filipe and J. Cordeiro (Eds.): WEBIST 2007, LNBIP 8, pp. 184–200, 2008. c Springer-Verlag Berlin Heidelberg 2008
Predicting the Influence of Emerging ICT on Home Life
185
1.1 The Evolution of the Personalised Communication in the Home Network The ePerSpace is an Integrated Project under EU 6th Framework Programme dedicated to the development of integrated, personalised communication services in the home area. The main concepts introduced are shortly: the “bubbles”, which contain the user profile information, and the “continuity of services”, the way to follow the user in everyday life scenarios and situations. As for the “home bubbles”, they carry the most important and least variable information, i.e. the basic preferences when the user is at home, at the same time as the “elsewhere bubbles” are more variable and can adapt to all situations when the user is not at home. The main objective of the project is to significantly increase the user acceptance of networked audiovisual systems and applications at home and virtually anywhere by developing innovative interoperable value-added networked services. The existing services such as eLearning, eBanking, eGovernment etc., are limited by several factors. These limitations are removed by making available an Open Access Network: by integrating the home platform and home audiovisual equipment with a variety of user devices; by using the personalisation profiles as a means to adapt contents to specific user, user requirements and communication devices; etc., as well as by improving the user-friendliness of the system. On this main purposes the project developed five scenarios outlining the solutions to support the everyday life. The “At Home” is the reference scenario for this paper. The core of the system is the user profile, which defines the user preferences. It resides on a Residential Gateway, which is both the hardware system and the heart of the digital home life. The home gateway links with all the individual devices and provides them with information needed, even thought all devices have different formats and different form factors (i.e., a mobile phone has a small screen and uses one type of data, a TV has a large screen and receives data in a different way). Data are transferred through three primary networking technologies: broadband, WiFi and Bluetooth. 1.2 The Assessing Environment The way to assess the general architectural concepts and as well as to evaluate such complex technological achievements is arranged in different trials, each one centred on the key themes addressed by the project, i.e. the Home Networking for the services integrated within the home platform. Namely the Home Platform laboratory trial is an approach which gets together all services on a single site where the key architectural features and the basic concepts of the integration could be shown in an sole environment. That laboratory environment (Figure 1) is both targeted towards the demonstration of the overall achievements and comfortable to carry out a user evaluation of the proposed solutions. The demonstration is oriented at showing the intercommunication of the different personal devices with the RG for the purpose of making themselves known in the community of devices available to the users and to present the services provided through them.
186
M. Cornacchia, V. Baroncini, and S. Livi
Fig. 1. Plant of the laboratory reproducing the home environment
2 The Acceptance Model Information Technology represents today a primary way of transforming society but each application is assumed to achieve specific benefits. The new technologies actually can be applied to achieve a wide variety of benefits (e.g. improve quality of the life, of the work, etc.) and have influences in the organisational change (e.g. improve productivity, enhance work, effectiveness, etc.). Because any kind of technical innovation is characterized by a certain risk, there is the problem of how to implement the technology and, ahead of this phase, the problem of predict its influence (in a short: success or failure?) in view of the high costs effort to produce. If we look at some evidence about the success or failure rates of information technology projects, we firstly see that is very difficult to attain data as the high complexity and variability of the whole socio-technical system to consider [3]. The ordinary criticisms are that the technology is being oversold [4] and that it is regularly subject of changes within short periods of time. Nevertheless, there are some studies, named in the following, that give an indication of the scale of the problem and the nature of the possible outcomes. These studies are aiming to support those organisations that accept risky investment decisions for instance in order to get a better competitive position. Many examples of the emerging information technologies have been publicized with consistent investment market projections, but they remain strongly fastened by a broad alone of uncertainty as for their effectiveness. At the last, the most important questions rising up the mind of the decision makers are about which of these technologies will succeed and what the useful applications have to be. In the history the relevant literature describes the development of several models of technology acceptance (by the users) and many extensions to the basic constructs [5], [6], mostly built with the behavioural elements [7] of who is forming an intention to act [8] and the inclusions of some kinds of constraints (limited ability, learning and usage [9], time, environmental, organisational, unconscious habits, and so on) which influence the individuals actions [10], [11].
Predicting the Influence of Emerging ICT on Home Life
187
Information technology acceptance research has applied many competing models, each one with different sets and very often overlapping of the acceptance determinants [12]. In their paper [2] Venkatesh and colleagues compared eight competing models that were applied in order to understand and predict user acceptance: Theory of Reasoned Action (TRA), Technology Acceptance Model (TAM), Motivational Model (MM), Theory of Planned Behavior (TPB), Combined TAM and TPB (C-TAM-TPB), Model of PC Utilization (MPCU), Innovation Diffusion Theory (IDT), Social Cognitive Theory (SCT). Those models were originated from different disciplines mostly connected with the behaviour prediction [13] or specialized for the technology use, from psychology to information system literature. As a result, research on user acceptance appear to be fragmented in different methods and measures [14]. For this reason the authors empirically compared those concepts in order to formulate a Unified Theory of Acceptance and Use of Technology model (UTAUT) with four core determinants of intention and usage, and up to four moderators of key relationships (Figure 2): Performance Expectancy, Effort expectancy, Social Influence and Facilitating Conditions as well as other moderators variables (such as Gender, Age, Experience and Voluntariness of Use).
Fig. 2. Theory of Acceptance and Use of Technology model (UTAUT) from (Venkatesh et al, 2003)
Applied to the tested system, the Performance Expectancy is defined as the believes that using the new services will help him to attain gains in the behavioural objectives. For the ePerSpace aim, this variable will be made operative through the Perceived Usefulness construct, relative advantage of using the innovation compared to its precursor, and outcome expectations. The Effort Expectancy, defined as the degree of ease associated with the use of the new system, has as operative constructs the general perceived ease of use as well as the perceived complexity of the system. The Social Influence, defined as the individual perception of how individual social network believes that he or she should use the new system, has as operatives constructs Subjective norms, Social factors and Social image and Identity similarity. Finally, the Facilitating Conditions,
188
M. Cornacchia, V. Baroncini, and S. Livi
are represented by construct by Perceived Behavioral control, general facilitating conditions (such as objective environment factors) and compatibility with existing values and experience of the potential adopters. One of the basic concept underlying the model of user acceptance states that, in the domain of “consuming the emerging technology”, the actual use of information technology is influenced by the intention to use and by the individual reactions to using. And so, the greater are the positive reactions, the greater is the intention and therefore the possibility to engage in the use.
3 Method 3.1 Measurement Scales The measurements scales of the acceptance applied to the design of the questionnaire instrument are mentioned in the following. All the scales were tested and successfully used (high degree of adaptability, high Cronbach alpha to denote the consistence of the constructs, high variance explained to denote independency of variables) in several contexts or technological environments and for different classes of users. It is nevertheless important to call attention to the fact that the questionnaires were adapted to the tasks and that each subject participating to the evaluation got first confidence with the innovating technology. Said that, each questionnaire was referred to a specific task and there were included, when required, additional lines to purposely measure the usability aspects of some significance [15]. Therefore, besides the central constructs of the acceptance, the questionnaires included also other high reliability scales, either for usability (namely on efficiency, affect and control) either for the identity-similarity or motivations [16]. The results by the submission of such scales in the evaluation provided the empirical evidence of a large effect of personal identity on different behavioural intentions. As for the home services tested, the consumer behaviours may had a symbolic meaning beyond their practical and objective features and consequences. For example, buying a certain equipment/system could have been an associated behaviour with an image of “idealized people” or with the “prototype” of the persons who perform these behaviours. The constructs used in the assessment were: Performance Expectancy (Section A in the questionnaire), Effort Expectancy (Section B), Social Influence (Section C), Facilitating conditions (Section D), Attitudes toward Using Technology (Section E), Attitudes (towards home environment solutions) (Section F), Intentions (Section G), Identity-Similarity (Section H), Usability (SUMI questionnaire). The home services were evaluated by a shorter version from UTAUT model but using the same predictors and items suggested by the authors. In such a way, subjects were asked to state their agreement or disagreement to the items on seven points ranging scale from completely disagree (1) to completely agree (7). Each factor was processed first lonely, by computing the mean on its subdimensions (where items were more than one), then by computing the general mean on components (R=Reverse coded). 1. Perceived Usefulness: Using the system/service at home would enable me to accomplish this activity more quickly; using the system/service would enhance my effectiveness in home activities; using the system/service would make it easier to live better at home; I would find the system/service useful in carrying out home activities.
Predicting the Influence of Emerging ICT on Home Life
189
2. Effort Expectancy (a) Perceived ease of use: My interaction with the system/service would be clear and understandable; I would find the system/service to be flexible to interact with; It would be easy for me to become skilful at using the system/service; I would find the system/service easy to use. (b) Complexity: Using the system/service involves too much time doing mechanical operations (e.g. data input). (R); It takes too long to learn how to use the system/service to make it worth the effort. (R) (c) Ease of use: Learning to operate with the system/service is easy for me. 3. Social Influence (a) Social Factors: In general, my family can support the use of the system/service. (b) Image: Members of my family that could use the system/service could have a high level of knowledge. 4. Facilitating Conditions (a) Perceived behavioural control: I have the resources necessary to use the system/service; I have the knowledge necessary to use the system/service; The system/service is not compatible with other systems/services I use. (b) Facilitating conditions: Guidance was available to me in the selection of the system/service; Specialized instruction concerning the system/service was available to me. (c) Compatibility: Using the system/service fits into my lifestyle at home. 5. Attitudes Towards using Technology (a) Attitudes towards behaviour: I dislike the idea of using the system/service in my daily life. (R) (b) Intrinsic motivation: The actual process of using the system/service is pleasant. (c) Affect toward use: The system/service is okay for my home, but not the kind of home I want. (R) (d) Affect: There are (well identified) aspects in my home for what I would use the system/service. (R); Using the system/service would be frustrating for me. 6. Attitudes Towards EPS Solutions: It states the attitudes by means of four bipolar items with seven steps from 1 to 7: extremely undesirable/extremely desirable, extremely unimportant/ extremely important, extremely unuseful/ extremely useful, extremely disagreeable/ extremely agreeable. At last, the above Attitudes towards EPS Solutions, in the following also mentioned, according to the case, either General Attitudes towards HCP, or General Attitudes towards MMC, or General Attitudes towards PVR, was taken as the main dependent variable. 3.2 The Home Services Evaluated The home environment services evaluated by the users were selected from those that were set-up within the Home Platform Portal. An outlook of the services portal made available is shown in the following Figure 3. Basically the following three sets of services of the Home Platform were evaluated: 1. HOME CONTROL, the automated control for the house domestic devices from the PC in web interface. The user can select the option from a list of services.
190
M. Cornacchia, V. Baroncini, and S. Livi
Fig. 3. The home page to access the ePerSpace services at Home and Everywhere
2. MULTIMEDIA CONTENTS, the management of the personal multimedia files: many possibilities to show, to watch, to write, etc. and the option to firts start playing a video on TV and then to skip TV to continue playing on PDA, or vice versa. 3. PVR, the Personal Video Recording home service of personal files and/or preselected video channels on TV. The HOME CONTROL of the automated home appliances and devices includes the management of: Lonwork actuators and sensors over twisted pair: lights, water valve, blinds, canopies, door lock, fire/gas/water detectors, etc. Lonwork white appliances over power line: oven and washing machine. Actuators and sensors: lights and small motors attached to a demo panel. The MULTIMEDIA CONTENTS is classified into videos, music and photos. It is made available to the service by sharing the folders that contain the contents by means of UPnP (Universal Plug and Play) Media Servers. It is very comfortable to search for the desired content from the PDA and send it to be played at whatever device of the house able to play it. Contents can also be downloaded at the PDA used to manage the service. The PVR is a list of available automatic and/or manual recordings accessible through a web interface. The user of this service may prefer not to do the PRV recording programming by himself (automatic recordings according to user preferences) or may prefer to manage the list of recordings (automatic and manual) that is available from within the home but also in the elsewhere. 3.3 The Users Profile A two steps selection of final 40 users followed some general criteria in order the final class of subjects could be the most homogeneous possible and provide consistent
Predicting the Influence of Emerging ICT on Home Life
191
values of judgements. The subjects selected for this evaluation were chosen among the employees of the company hosting the trial, as close as possible to an ideal type of potential user, open minded enough towards ICT solutions (neither too much enthusiastic neither to much unwilling) and on an average skilled in using electronic digital devices (e.g. the PC or other home familiar devices). The first step in the definition of a class of users was mainly a rough skimming from a initial set of about 70 individuals, preliminarily contacted in order to be certain that: 1. subjects characteristics were close to the home services use requirements; 2. essential skills and basic attitudes towards ICT were not unacceptable; 3. subjects participants were volunteers. The second step was a selection of the final 40 users, among the several items relating about each individual, to check whether: 1. final group were sufficiently represent males and females; 2. age range (at least) were as narrow as significant to applied statistics; 3. availability to effectively participate to the test were an actual statement. The emerging final profile shows: 1. 2. 3. 4. 5. 6. 7.
a proportion of frequencies about 23% for females while 77% for males; evident age crowding in the range 30-39 years, equal occurrence of married and single individuals; medium-high level of education; mostly technicians and engineers (respectively 70% and 15%); high experience accrued with ordinary ICT (PC, Email, Mobile, etc.); proximity to household equipment with communication means.
The users participating to the evaluation went through an introduction of the Project and a training session in order to use the system almost without assistance and to give valuable/not-biased information to the questionnaires. 3.4 Set-Up of the Test Bed All the HAN of the test-bed were connected to the RG (Residential Gateway) which run an OSGi (Open Services Gateway initiative) framework over which the home platform services are managed and activated. A Personal Computer or Laptop or a Personal Digital Assistant (PDA), inside the house, were used either to access the web interface of the services and also to provide the I/O for the tasks planned to be accomplished by the users in the assessment. UPnP Media Servers and Renderers running in the equipment of the test bed. The PDA was connected to the HAN via WiFi in the house for the demonstration. A Set Top Box (STB) was connected to the TV set and could run also home environment services, controlled by the RG Middleware. PVR service started in the RG. STB connected to the TV and DVD-Player. DVD with some recordings from NRK and EPG info loaded in the DVD player. Web interface open in a PC. Basically the automated control for the house domestic devices from the PC in web interface services of the Home Platform were evaluated. The user accessed the
192
M. Cornacchia, V. Baroncini, and S. Livi
Fig. 4. “Home Control” sub-menu
options from a list of services directly through the main page. The procedure to carry out the evaluation followed a prearranged scheme. Each subject was received in front of the house door, informed about the overall session and the services to be evaluated by means of questionnaires. The questionnaires were arranged in a labelled sequence, then submitted to the subject. About 60 questions over the total amount of 193 were answered by each subject in the section of the local services for the home (Figure 4). The evaluation process took place in two weeks in total. The integration work lasted 6 months. The test bed used for the user evaluation gave a configuration and outlook of a real flat (Home). 3.5 The Tasks A set of tasks was properly designed for the class of users profiled for the trial and the services to evaluate. The services were accessed by the user from any PC or PDA wired or wirelessly connected to the LAN of the home. In both cases, the user started the browser of the access terminal to initially authenticate him/herself by username and password. After that, the user was admitted to the home portal and enabled to select from the list of the personal services. In case of being using the web access, a map of the house displayed icons representing the home appliances that can be actuated, as well as its current state (i.e., on/off). At the user click on each icon a menu of the possible actions appeared. For example, in the case of a light, currently on, the user was offered to switch it off and adjust the light intensity. In case of using the PDA access, instead of a map of the house, the user found a list showing the rooms in the house. At the user click on one of the rooms, the list of automated devices to be controlled in that room displayed. The running was similar to the web access, but the graphical interface was more simple to adjust to the limited screen size. Here below the tasks performed by the users in the evaluation trial are briefly described:
Predicting the Influence of Emerging ICT on Home Life
193
1. In the task “Access to the local service portal” the user was asked to authenticate her/himself in the platform (i.e., SIM-card based, RFId iCode card, smart card) and to identify the terminal used to access the local services portal. 2. In the task “Controlling some home appliances from the laptop, wirelessly connected to the LAN of the house” the user did some control actions by pc web interface (raise the blinds, switch off the light of the kitchen, and connect to the camera of the living room to check that everything goes well). 3. In the task “Local handover-UpnP AV streaming” the user, equipped with a WiFi PDA, was asked to start watching a video on the TV in the living room. Then, as soon as she/he moved from the living room to the kitchen, the localization service found her/him in a different room and decided for the most suitable device to continue playing the video from those available in the room (i.e., the WiFi PDA or the laptop placed in the kitchen table). 4. In the task “Private multimedia management” the user interacted with the service to manage the multimedia content stored at home. Multimedia content was classified into videos, music and photos. By means of this service, the user might choose a video, for example, and send it to any of the Media Renderers available, e.g., the STB connected to the TV, the PDA itself, the TV, or a PC. 5. In the task “Use of EPS PVR service” the user interacted with the web interface to manage a list of automatic and manual recordings previously. 3.6 The Data Analysis The questionnaires were coded and the data properly wrapped up to be analysed by means of the SPSS (Statistical Package for Social Science 13.0). Three main groups of analyses were generated as output: 1. Descriptive Statistics for all cases and outliers identification; the data were grouped in the way that the scores of the home services appeared, for each aspect of the Acceptance. 2. Correlation study for the Home services relative to the dependant variable “attitudes towards behavior”. 3. Regression models to identify the factors which have more influence on the UTAUT predictors.
4 Results The descriptive statistics (mean and confidence interval within ± σ) were shortly assembled in the lines of Table 1, as many as the cases were respetively for the Home Control Panel and the home services through it accessed, and for the Multimedia Contents and PVR residential services. 4.1 Descriptive Statistics The analysis of the frequencies gathered pointed out that from a broad point of view the services were well accepted by the users, as innovative for the home and access from
194
M. Cornacchia, V. Baroncini, and S. Livi
Table 1. Descriptive statistics for the Home Services, Multimedia Contents and PVR accessed through the Home Control Panel Home Services Perceived Usefulness Effort Expectancy Perceived ease of use Complexity Ease of use Social Influence Social factors Image Facilitating Conditions Perceived behavioural control Facilitating conditions Compatibility Attitudes towards technology Attitudes towards behaviour Intrinsic motivation Affect towards use Affect General attitudes towards HCP General attitudes towards MMC General attitudes towards PVR
the elsewhere. The variables used to define the constructs of the acceptance model, all showed a definite tendency in positively comparing the ICT solutions presented in the test-bed with the already available personal services that can be seen as a clear advantage and concrete expectation for the home services to improve the life style of its users. This important result was first attained, by showing exactly the same rank order for all the three services HCP, MMC and PVR, by the variables: Effort Expectancy (i.e. very high value of ease-of-use and low perception of complexity), Attitudes towards Technology, Perceived Usefulness, Social Influence (i.e. the family view coherent and close to a doable real use) and Facilitating Conditions. At last, the external conditions tested were compatible with the life style of the subjects and as a matter of fact not opposed to the potential adoption, as well as the wide-ranging attitudes towards the new home solutions. In order to evaluate the differences between Male and Female perception of the Home services, respectively HCP, MMC and PVR, the means of each sample were compared and the Analysis of Variance (ANOVA) was performed to verify if those differences were statistically significant. Results showed that, overall, males and females perceived the user acceptance in the same way for almost all the dimensions explored. The only noteworthy exceptions were pointed out for the “Perceived Usefulness as
Predicting the Influence of Emerging ICT on Home Life
195
concerns the MMC and PVR home services” (so not HCP, the set of appliances closest to the home management), where males, more than females, stated that using MMC and PVR would enhance their job performance at home (ANOVA: F(1,37)=4.31; p<.05 and F(1,37)=4.22; p<.05). The Perceived Usefulness means considered in the ANOVA were as well: males=4.37, females=3.56 for MMC; males=4.52, females=3.53 for PVR. 4.2 Correlation The mean described in previous analysis gives back a clear picture of the general tendency toward this innovative home technology: the EPS solutions are mostly completely accepted despite the system introduces solutions considerably different from those experienced in the everyday life. Nevertheless, the central tendency is not a complete indicator of the sample. As we have seen in previous tables, standard deviations and ranges, that is sample straggling, suggest that overall attitudes may vary depending on subjective perceptions of acceptance. In order to understand which element of the UTAUT model predicts general positive attitude toward the system, Pearson correlations were computed for each of the EPS services, i.e. Home Control Panel (HCP), Multimedia Content (MMC), Personal Video Recording (PVR). Looking at the coefficients concerning Home Control Panel in Table 2, at HCP column, it clearly emerges that all the UTAUT dimensions, such as (in rank order) Attitudes Toward Technology (r=.62; p<.01), Efforts (r=.54; p<.01), Usefulness (r=.53; p<.01), Social Influence (r=.50; p<.01) and Facilitating Conditions (r=.44; p<.01), are strong predictors of a positive evaluation of the HCP. Thus, user positive acceptance helps to have a positive attitude toward the HCP. More interestingly, some subdimensions of the UTAUT reveal a specific bias toward the HCP: for example, the subdimension “image” in the goup “social influence” remarkably decreases the positive evaluation. This means that the family status is not important, but nonetheless the user’s social support through family may positively influence the acceptance. Further about the subdimension “facilitating conditions” in the homonymous group, the perceived guidance availability or the perceived behavioural control result not significant predictors compared to the compatibility of the lifestyle at home, although this result does not affect the general tendency of the factor. Ultimately, the subdimension “general affect” in the group of the “attitudes toward technology” reveals a minimal influence compared to the other subdimensions. The pattern of correlations coefficients between evaluation of the Multimedia Content and Acceptance variables is a little bit different from the previous one. First of all, as shown in Table 2, at MMC column, overall mean of the coefficients is lower compared to the previous service. This is probably due to the fact that the service is perceived as more familiar by some “technological users”. Thus, the main predictors are identified in those dimensions more close to the everyday use of the technology: Attitude toward technology (r=.60; p<.01), but also Usefulness (r=.35; p<.05). Moreover in the analysis of the subdimensions, compatibility, perceived ease of use and social factors remain significant predictors, although the general factors are not significant. The latter result is particular important in confirming how social factors, outmoded technology ambiguity, become an important issue among the others to catch a positive evaluation of the system.
196
M. Cornacchia, V. Baroncini, and S. Livi
The PVR home service assessment shows in Table 2, at PVR column, the strongest coefficients pattern: almost all UTAUT predictors (factors and subdimensions) are significant. In particular usefulness is the strongest predictor (r=.76; p<.01), followed by the facilitating conditions (r=.69; p<.01), attitudes toward technology (r=.67; p<.01) and social influence (r=.42; p<.01). In this case, perceived usefulness and facilitating conditions of the system are the most relevant factors to say that as the utility of the goal is better understood, as the positive evaluation is higher. The most likely reason for this result can be ascribed to an earlier familiarity with the PVR set of devices. Usually the subject attending the test is very impressed from the front of the working functionalities (the part of the technology or the service close to him or under the immediate control), such as a new modality to interact or a new control device. But, whether known devices are used, he applies the previous experience to go ahead in the interaction and in comparing the new items, even though when accessed in a fragmented way due to the condition of being in a laboratory trial. As given in the previous correlation tables, all values in bold and with one/two asterisks indicate the presence of a high/higher degree of correlation between the variables. Having a look to the links in the picture of the predictive model of acceptance, it means that, for the home services, there were good probabilities that the positive attitude of a user-consumer were influenced by those variables. Taken for example the Home Control “Perceived Usefulness: effectiveness in home activities”, it can be said that: either “the Table 2. Correlations between UTAUT variables and the general attitudes towards Home Control Panel, MultimediaC and PVR services UTAUT variables Perceived Usefulness Effort Expectancy Perceived ease of use Complexity Ease of use Social Influence Social factors Image Facilitating Conditions Perceived behavioural control Facilitating conditions Compatibility Attitudes towards technology Attitudes towards behaviour Intrinsic motivation Affect towards use Affect (**) Correlation significant at the degree of 0,01 (2-tails) (*) Correlation significant at the degree of 0,05 (2-tails)
Predicting the Influence of Emerging ICT on Home Life
197
greater is the effectiveness in the home activities the greater is this influence on the “positive attitude towards solutions”, or it can be said “the significant Pearson correlation provides evidence that the variable “perceived usefulness” stimulates the perception of positive attitude towards solutions”. 4.3 Regression Models Although correlation coefficients give a whole picture of the different weights of the UTAUT factors and sub dimensions toward the positive attitude of the EPS solutions, three more regression models are herewith performed. These analyses allow a better understanding of the influence that each factor may have in controlling the weight of the single predictor. Three models were so tested, using as criteria the general evaluation respectively towards Home Control Panel, Multimedia Content and PVR. As shown in Table 3 the first model resulted significant by using as dependent variable the general evaluation toward the Home Control Panel (Adjusted R Square=.44; p<.001). Nevertheless, the only significant predictor is the perceived low efforts expectancies. Thus, when controlling for the other effects, the more subjects perceived low efforts expectancy in using the HCP the more positive is the evaluation (Beta=.36; p<.05). In the second model, as shown in Table 4, the evaluation of the Multimedia Contents used as dependent variable gives a significant result (Adjusted R Square=.28; p<.01). Nevertheless the main predictor in this case is the Attitude toward technology (Beta=.53; p<.01). Finally, also the third model involving the PVR evaluation, as shown in Table 5, explains the highest percentage of variance (Adjusted R Square=.67; p<.01). Two predictors in this case show the highest coefficients: the low efforts expectancies (Beta=.25; p<.05) the same as for the HCP and, above all, the perceived usefulness of the PVR (Beta=.46; p<.01). As it can be seen, the regression models of MMC and PVR are different each other because different variables have influence on the attitudes towards the EPS solutions. It can be reasonably presumed that in both cases subjects applied their own traditional experience in ICT interaction to accomplish the required tasks. So far, in the case of MMC they spent more attention to manage new innovative multimedia objects, even for a known scope, while for the PVR they felt earlier familiar with both media and task scope. The multimedia objects and tasks to perform as well were also presented to the subjects in a fragmented way due to the limited time available to carry out the Table 3. Home Control Panel Regression. Dependent Variable: Attitude toward HCP. Predictor Usefulness Low Effort Social Influence Facilitating Conditions Attitudes towards technology
Beta .12 .36(*) .22 .04 .21
t .71 2.47 1.18 .25 1.06
Sig. .48 .02 .25 .81 .30
198
M. Cornacchia, V. Baroncini, and S. Livi Table 4. Home Control Panel Regression. Dependent Variable: Attitude toward MMC. Predictor Usefulness Low Effort Social Influence Facilitating Conditions Attitudes towards technology
Beta .13 -.11 .01 .13 .54(**)
t .75 -.69 .03 .83 3.04
Sig. .46 .50 .97 .41 .01
Table 5. Home Control Panel Regression. Dependent Variable: Attitude toward PVR. Predictor Usefulness Low Effort Social Influence Facilitating Conditions Attitudes towards technology
Beta .46(**) .25(*) -.02 .15 .15
t 3.44 2.05 -.21 1.00 1.09
Sig. .002 .05 .83 .33 .29
laboratory trial. So the Low Effort significance scored for PVR regression, and not for other variables, is almost certainly due to the intrinsic simplicity of task and skills already owned to execute it.
5 Conclusions The new home services as provided in the trial were on the whole accessed by the subjects with high curiosity and interest. The adoption of a controlled interactive session to present the services and their innovative features was able to give to each single participant the time necessary to understand and quickly build a personal judge about. Then the repeated sequence task-questionnaire to gather data demonstrated to be appropriate in catching the impulsive ideas about the added values given by each service in comparison with the actual home possibilities, as well as the possible adoption in the own life. This is an excellent consequence of the methodology proposed for the evaluation, then proved by the reliable data obtained. As for the acceptance items, this is a composite variable that can be carefully expressed by combinations of different results (and different constructs). Therefore, by considering the “perceived usefulness” (Questionnaire Section A), the users class received a positive feeling from the services. This result is strongly powered by looking at the values of the Questionnaire Section B, which indicates the clear easiness to operate and the low perception of underneath complexity, even not really so (this is and excellent result from the usability point of view, better evident for the Home Control). The social factors (Questionnaire Sections C) confirmed how the subjects view was also shareable with the family and the close neigh borough. The facilitating conditions “having the resources/knowledge necessary to use the system” (Questionnaire Section D) resulted about neutrally considered in relation with the acceptance, while it was very
Predicting the Influence of Emerging ICT on Home Life
199
encouraging the perception of the “compatibility”, actually close to the idea of lifestyle at home. The rest of the items in the questionnaire directly checked the attitudes of the subjects towards both the service idea and the actual usage possibility at home. The resultant Questionnaire Section E scored high positive values, indicating that there were clear intrinsic motivations in the thought of acquiring these services for the home. Finally, the last variable (Questionnaire Section F) of “attitudes towards EPS solutions” in a straight line confirmed that the users should be willing to introduce the new solutions at home, as they were desirable, important, useful and agreeable. The ANOVA applied to find differences of behaviour between males and females, pointed out no statistical differences except that females perceived differently the “usefulness at home” of the services. The analysis on the data gathered of course didn’t investigate the cause of this “social” difference, nonetheless, it is spontaneous to think at the actual different condition of the women at home in different countries and the different perception of “the usual staying at home” they may have with respect to the men. Finally, besides the correlations analysis which offers by itself a complete view of the effects of the UTAUT model predictors, the regression analysis carried out on Home Control Panel services to manage the digital home, the Multimedia Contents and Personal Video Recording services, properly designed to enrich and facilitate the life of people at home (and elsewhere), clearly identifies the main role played by certain variables (i.e., low effort, attitudes towards technology, usefulness) and the special weight that they may have in predicting the acceptance by the users of such emerging new technologies. Acknowledgements. Special thanks are due to Telef´ onica I+D, partner of the ePerSpace Project, without whom the experimental sessions with the users couldn’t have taken place.
References 1. ePerSpace No.IST-506775 Home Page, http://www.ist-eperspace.org/ 2. Venkatesh, V., Morris, M.G., Davis, G.B., Davis, F.D.: User Acceptance of Information Technology: Toward a Unified View. MIS Quarterly 27(3), 425–478 (2003) 3. Eason, K.: Information technology and organisational change. Taylor & Francis, Abington (1988) 4. Cornacchia, M.: Usability, a way to distinguish from the good, the bad, and the irrelevant in the web. In: Cost 269, Conference Proceedings, Media Centre Lume, University of Art and Design, Helsinki (Finland), pp. 159–163 (2003) 5. Malhotra, Y., Galletta, D.F.: Extending the Technology Acceptance Model to Account for Social Influence: Theoretical Bases and Empirical Validation. In: Proceedings of the 32th Hawaii International Conference on System Sciences (IEEE 1999) (1999) 6. Venkatesh, V., Davis, F.D.: A theoretical extension of the technology acceptance model: Four longitudinal field studies. Management Science 46(2), 186–204 (2000) 7. Ajzen, I.: The directive influence of attitudes on behavior. In: Gollwitzer, P., Bargh, J.A. (eds.) Psychology of action, pp. 385–403. Guilford, New York (1996) 8. Bandura, A.: Social Foundation of Thought and Action: A Social Cognitive Theory. Prentice Hall, Englewood Cliffs (1986)
200
M. Cornacchia, V. Baroncini, and S. Livi
9. Bagozzi, R.P., Davis, F.D., Warshaw, P.R.: Development and test of a theory of technological learning and usage. Human Relations 45(7), 660–686 (1992) 10. Compeau, D.R., Higgins, C.A., Huff, S.: Social Cognitive Theory and Individual Reactions to Computing Technology: a Longitudinal Study. MIS Quarterly 23(1), 145–158 (1999) 11. Pierro, A., Mannetti, M., Livi, S.: Self-Identity and the Theory of Planned Behavior in the Prediction of Health Behavior and Leisure Activity. Self and Identity 2, 47–60 (2003) 12. Davis, F.D.: Perceived usefulness, perceived ease of use, and user acceptance of Information Technology. MIS Quarterly 13(3), 319–340 (1989) 13. Ajzen, I., Fishbein, M.: Understanding attitudes and predicting social behaviour. PrenticeHall, Eaglewood Cliffs (1980) 14. Venkatesh, V., Davis, F.D.: Measuring User Acceptance of Emerging Technologies: An Assessment of Possible Method Biases. In: Proceedings of the 28th Annual Hawaii International Conference on System Sciences (IEEE 1995) (1995) 15. SUMI: Software Usability Measurements Inventory (7.38). Human Factors Research Group, Cork, Ireland (1998) 16. Perugini, M., Gallucci, M., Livi, S.: Looking for a Simple Big Five Factorial Structure in the Domain of Adjectives. European Journal of Psychological Assessment 16(2), 87–97 (2000)
RDF Collections Saket Kaushik and Duminda Wijesekera Department of Computer Science, George Mason University Fairfax, VA 22030, U.S.A. {skaushik,dwijesek}@gmu.edu http://www.ise.gmu.edu/∼duminda
Abstract. We add collection types such as sets, bags and lists to the resource description framework (RDF). Keywords: Resource Description Framework, RDF, The Semantic web.
1 Introduction The Semantic Web consisting of a layered collection of languages [1]. Among them is a language referred to as the Resource Description Framework (RDF) [2] for expressing properties of resources. Basic RDF syntax consists of three syntactic categories: viz subjects, objects and properties between subjects and objects. It also has some meta-syntax to declare properties between its own syntactic categories. The most cited semantics of RDF consists of collections of (subject, property, object) triples - referred to as its graph semantics. Although different semantics have been provided for semantic web languages, Lbase [3] has been proposed as a meta-language to translate all of them and thereby provide a uniform semantics through using the Lbase semantics for the target of the translation. Though expressive, as stated in [3], Lbase lacks some desirable features, such as temporal ontologies, version control for ontologies, and higher-order constructs, etc. The same is true for its successor specification – RDF Semantics [4] and predecessors ([5], etc.). Although the cited specifications provide a syntax for collections like bags, lists, alternatives, etc., they do not provide semantics - thereby making their interpretation problematic. Nevertheless collection types are useful in modeling resources. For example, a shopping bag can be modeled as a bag of its items. Consequently any access to or a computation based on the bag - such as computing the total cost of the bag as a sum of its items - depends upon the bag. As a remedy, we add syntax for three collection constructs; namely, sets, bags and lists to the self reference less fragment of RDF, and provide a semantics for the extended syntax that subsumes the RDF semantics of [4] for (soon to be explained) simple and reified graphs. We show how our semantics relates to the graph entailments of [4]. The utility of our syntactic extension is shown by an example detailing a cost computation of the stated shopping bag. The rest of paper is organized as follows. Section 2 introduces RDF, and section 3 specifies the syntax and properties of finite sets, bags and sequences we use. Section 4 J. Filipe and J. Cordeiro (Eds.): WEBIST 2007, LNBIP 8, pp. 201–212, 2008. c Springer-Verlag Berlin Heidelberg 2008
202
S. Kaushik and D. Wijesekera
defines the extended RDF syntax. Semantics is described in section 5 and the relationship to graph entailments described in section 6. Section 7 concludes the paper.
2 RDF RDF consists of statements about resources and their properties. RDF syntax mainly consists of three syntactic categories of objects referred to as objects and subjects and (binary) properties between them. RDF statements are (subject, property object) triples, where (S,P,O) is (infix) notation for subject S having property P with value object O - customarily written as P(s,o) in predicate logic. Thus, subjects and objects are considered atomic, but properties may be extended beyond connectors between atomic objects - for example expressing relationships between properties and atomic objects. This process, carried out recursively, is called reification. RDF Schema (RDFs) is RDFs vocabulary description language with meta-syntax such as rdfs:Class, rdf:type, etc., to describe concepts and resources and relationships between resources through rdf:property. These meta classes are used to specify properties of user defined schema, as detailed in [2]. A (finite) collection of RDF triples form a graph, where if the object of one triple is the subject of another, then the two triples are (informally) visualized as forming a graph. Hayes and McBride [4] give semantics of RDF/RDF(S) as RDF graphs (informally) using so called model theory. They informally state an entailment relationship between graphs. To the best of our understanding, an RDF graph is the collection of triples; i.e., a collection of binary predicate instances - which is how Lbase translates RDF into Lω1 ,ω . They also describe containers such as multisets and lists without further semantics. To formalize, we (recursively) add three collections, viz., sets, bags and sequences as abstract data types [6] over some basic collection of types.
3 Container Data Types Abstract data types (ADTs) sets, bags and lists support a collection of algebraic operations. Following convention, sets do not repeat elements, but bags do, and lists retain a position for each element in the collection. In order to permit reification, our ADTs creation begins over a collection of base types and repeated recursively. We use Set[T ], Bag[T ] and List[T ] for sets, bags and lists constructed from elements of type T . We use the conventional notation of ∈ as the membership relation of sets and bags. The multiplicity of an element in a bag - i.e. the number of its repetitions in the bag [7,8,9,10,11], is modeled as the profile function ⊕, where for bag B with member x, (x ⊕ B) ∈ N+ (the set of positive natural numbers) is the profile of x in B. Size of a bag is given by the number of objects in the bag, counting repetitions. We use, two functions head first(S) (i.e. Car) and tail last(S) (i.e., Cdr) that returns the first element and the remaining sub-list of a list S. Next we provide algebraic specifications for the collection types [6] using the notation x : T for x is of type T. Definition 1 (Finite Sets). Set[T ] is defined in table 1.
RDF Collections Table 1. Typing Set Operators Operation Type Name { x} T → Set[T ] Constructor ins T × Set[T ] → Set[T ] Insert ∈ T × Set[T ] → boolean Membership ⊆ Set[T ]× Set[T ] → boolean Subset == Set[T ]× Set[T ] → boolean Equality rem T × Set[T ] → Set[T ] Remove cnt Set[T ] → N Cardinality
Properties of sets: x ∈ {x}, {x} : Set[T ], x ∈ {}T ← x : T x ∈ a, a : Set[T ] → x : T x ∈ b, b ⊂ a, a : Set[T ], → x : T, b : Set[T ] ins(x, a) : Set[T ] ↔ x : T, a : Set[T ] rem(x, a) : Set[T ] ↔ x : T, a : Set[T ] Properties assumed for operations on sets are as follows: x ∈ {}, x ∈ {x} x ∈ ins(y, z) = if x = y or x ∈ z x ∈ rm(y, z) = if x = y and x ∈ z cnt({ }T ) = 0 cnt({x}) = 1 cnt(ins(x, z)) = cnt(z) if x ∈ z, 1 + cnt(z) otherwise Bag algebra is presented next. Definition 2 (Finite Bags). Bag[T] is defined in table 2.
Table 2. Typing Bag Operations Operation [[x, x, . . . , x]] ins rem ∈ ⊕ cnt
Type Name T → Bag[T ] Constructor T × N× Bag[T ] → Bag[T ] Insert T × N× Bag[T ] → Bag[T ] Remove T ×Bag[T ] → boolean Membership T × Bag[T ] → N Profile Bag[T ]× Bag[T ] → boolean Subbag Bag[T ]× Bag[T ] → boolean Proper Subbag Bag[T ] → N Cardinality
203
204
S. Kaushik and D. Wijesekera
Properties of Bag[T ] are: x ⊕ [[ ]]T = 0 x ∈ [[ ]]T , x ∈ [[ x ]]T ↔ x : T (x ⊕ a) > 0, a : Bag[T ]) → x : T ins(x, a) : Bag[T ] ↔ x : T, a : Bag[T ] rem(x, a) : Bag[T ] ↔ x : T, a : Bag[T ] Given x, y : T, n ∈ N, b, b : Bag[T ], the semantics of bags are as follows: x ∈ [[ x ]], x ∈ [[ ]]T x ∈ ins(y, b) = if x = y or x ∈ b x ⊕ b = 0 if x ∈ b x ⊕ ins(x, b) = 1 + x ⊕ b cnt([[ ]]Ti ) = 0 cnt(ins(x, b)) = 1 + cnt(b) a b if f (x ∈ a → x ∈ b)and (x ⊕ a < x ⊕ b) for all x a b if f (x ∈ a → x ∈ b)and (x ⊕ a ≤ x ⊕ b) for all x y ∈ rem(x, b) if f y = x and y ∈ b) Next we define lists. Following the usual practice, they are built by pre-pending the empty element nil. Definition 3 (Lists). Given base types T and a constant symbol nil of type List[T ], syntax of operations on list type (written List[T ]) are given in table 3.
Table 3. Typing list operations Operation Type Name x Ti → List[T ] Constructor hd List[T ] → T Head tl List[T ] → T Tail ∈ Ti × List[T ] → boolean Membership == List[T ] × List[T ] → boolean Equality
x : List[T ] ↔ x : T, hd(x) : T, tl(x) : List[T ] holds for lists. Definition 4 (Triples). Given types S,O and P (type T denotes type S, type O or type P), syntax of operations on triples is given in table 4.
RDF Collections
205
Table 4. Triple type Operation Type Name (S, P, O) S × P × O → (S, P, O) Constructor sub (S, P, O) → S Subject prp (S, P, O) → P Property obj (S, P, O) → O Object ∈ T × (S, P, O) → boolean Membership == (S , P , O ) × (S, P, O) → boolean Equality
Semantics of operations is given as follows: sub(x, y, z) = x prp(x, y, z) = y obj(x, y, z) = z u ∈ (x, y, z) = true if u = x ∨ u = y ∨ u = z, f alse otherwise a == b = true if sub(a) = sub(b) ∧ prp(a) = prp(b) ∧obj(a) = obj(b), f alse otherwise
4 Formalizing RDF Our formalization of RDF syntax begins with two base types, S and O for subjects and objects and a subtype of S × O for properties. We then build container subjects and objects using the algebraic operations specified in section 3, and typed triples as in definition 5. Definition 5 (Enhanced RDF Syntax) Stage 0: Suppose S, O are type symbols for subjects and objects, and N is the set of natural numbers. Assume that there are an infinite collection of constants and variables of types S, O and N. We also assume that there is a subtype P of the type S × O referred to as the type of properties. For each constant cT and variable −T of type T , add (cT rdf:typeT ) and (−T rdf:typeT ) to T RIPLE 0 . Stage n+1: For every type T that has been created up to stage n, create types SetT , Bag[T ] and List[T ] as described in section 3, with an infinite collection of type variables of types SetT , Bag[T ] and List[T ] and inhabiting objects (recursively) created from inhabiting objects of type T . 1. For every type T and Set[T ] create a new type T → Set[T ], a property ∈T → Set[T ] of type T → Set[T ], a property cnt of type Set[T ] → N, and properties ⊆Set[T ]× Set[T ] and ==Set[T ]× Set[T ] of type Set[T ] × Set[T ]. 2. For every type T and Bag[T ] create a new type T → Bag[T ], a property ∈T → Bag[T ] of type T → Bag[T ], a property cnt of type Bag[T ] → N, a property T × Bag[T ] →N of type T × Bag[T ] → N, and properties Bag[T ]× Bag[T ] and ==Bag[T ]× Bag[T ] of type Bag[T ] × Bag[T ].
206
S. Kaushik and D. Wijesekera
3. For every type T and List[T ], add a property T → List[T ], a property ∈T → List[T ] of type T → List[T ] and a property ==List[T ]× List[T ] of type List[T ] × List[T ]. Add the Cartesian closure and functional closure of types created up to stage n + 1. For each constant cT and variable −T of type T , add (cT rdf:typeT ) and (−T rdf:typeT ) to T RIPLE n+1 . For each types T1 and T2 and property P of types f (T1 , T2 ) created upto stage n + 1, create countable number of variables and constants of those types. RDF graph is defined as a typed collection of RDF triples defined above, where each component is allowed to be a variable or a constant of the appropriate type. We say that a graph s is a subgraph of g, g s, if triples of s are a subset of triples of g. Next we illustrate an example of a shopping bag, discussed earlier, using the additional syntax. Example 1. Consider an ontology for an e-Commerce website that sells several food items online. Customers are allowed collect several items by placing them in a shopping bag. These items are paid for all together whenever the customer is ready to ‘check out’ items. A short sample is presented next. Stage 0: (Milk, rdf:Type, Item) (Yogurt, rdf:Type Item) (Bread, rdf:Type, Item) (Steak, rdf:Type, Item) (price, rdf:Type, N) (Item, hasPrice, price) (Milk, hasPrice, 3) (Yogurt, hasPrice, 1) (Bread, hasPrice, 1) (Steak, hasPrice, 10) Stage 1: (ShoppingBag, rdf:Type, Bag[Item]) (Milk, ∈, ShoppingBag) (Yogurt, ∈, ShoppingBag) (Yogurt, ∈, ShoppingBag) (ShoppingBag[Item], cnt, 3) ([[Milk, Yogurt]],,ShoppingBag) ([[Yogurt, Yogurt]],,ShoppingBag) ([[Milk, Yogurt, Yogurt]],,ShoppingBag) Definition 6 (Algebra for Nodes and Triples). Given a set of atomic types U , a set of boolean values B, integers N and a set of collection constructors – { {}, [[ ]], } over U , collectively called collections and represented by the type C, we inductively define nodes (Ni ) and triples (T ri ) in table 5. Properties of nodes and triples are presented next. x ∈ T ri ⇒ sub(x) ∈ Ni ∧ obj(x) ∈ Ni x ∈ T ri ⇒ prp(x) = sub(x) ∧ prp(x) = obj(x) Table 5. Nodes and Triples Type x : U ∪ x : C[U ] → N0 y : (N0 × U × N0 ) → T r0 x : (C[Ni−1 ] C[T ri−1 ]) ∪ x : C[U ]) → Ni y : (Ni × U × Ni ) → T ri
Name Node 0 Constructor Triple 0 Constructor Node i constructor Triple i Constructor
Definition 7 (Algebra for RDF). Given a universe U, and nodes and triples defined over this universe, we define an RDF graph as a set over U with operations defined in table 6.
Type Name T RIPLE → GRAPH Constructor GRAPH → Set[Si ] subjects GRAPH → Set[Oi ] objects GRAPH → Set[T RIPLE i ] triples GRAPH × GRAPH → GRAPH Merge GRAPH × GRAPH → GRAPH Minus GRAPH × GRAPH → boolean Subgraph GRAPH × GRAPH → boolean Equality
Properties of graphs are presented next. x : GRAPH ⇒ y ∈ sub0 (x) → y ∈ N0 x : GRAPH ⇒ y ∈ subi (x) → y ∈ subi−1 (x) ∨ prpi−1 (x) x : GRAPH ⇒ y ∈ obj0 (x) → y ∈ N0 x : GRAPH ⇒ y ∈ obji (x) → y ∈ obji−1 (x) ∨ prpi−1 (x) x : GRAPH ⇒ y ∈ trpi (x) → y ∈ T ri x : GRAPH ⇒ ∃k : N|T rk (x) = {} z = x ∪ y ⇒ trpi (z) : T ri , trpi (z) ⊆ trpi (x) ∪ trpi (y) The semantics of operations are as follows: subi (a ∪ b) = sub0 (a) ∪ sub0 (b) obji (a ∪ b) = obj0 (a) ∪ obj0 (b) trpi (a ∪ b) = prp0 (a) ∪ trp0 (b) subi (a \ b) = subi (a) \ subi (b) obji (a \ b) = obji (a) \ obji (b) trpi (a \ b) = trpi (a) \ trpi (b) a b = true if trpi (a) ⊆ trpi (b) a =s b = a b ∧ b a Equality relation ‘=s ’, defined above, equates ‘structurally equivalent’ graphs. That is, two graphs are structurally equal if they have exactly the same structure (based on the equality relation of constituent types). RDF also allows name-based equality, which we cover in the next section. 4.1 Named RDF Graphs To enable name-based equality of RDF graphs, we introduce names for graph elements. Definition 8 (Names). Let U be a universe of objects.
208
S. Kaushik and D. Wijesekera
Named Graph: We use a set of names, called N AMES and a naming function N from Graph to N AMES ∪ (⊥ × N) (where (⊥ × N) represents a blank node and its ‘node Id’ [4]) that maps each node and property in an RDF graph to its name. We abuse the notation slightly to say N (x) = ⊥ if a node is a blank node. Ground Graph: A graph that has no nodes mapped to ⊥. Nodes mapped to ⊥ are called variables. We allow variables to be mapped to constants by mapping functions called ground substitutions. Since our graphs are finite, the substitutions are finite as well. Definition 9 (Ground Substitution). A ground substitution, subs, is a mapping that maps variables of type T, i.e., −T , to constants of type T, i.e., cT . Next, we define name-based equality of graphs. With structural equality as a prerequisite, blank nodes at similar structures are assumed to be equal). This is followed by the definition of graph instances. Definition 10 (Equality of Graphs) Structural equality: Two graphs g1 and g2 are said to be structurally equal if ∀i : N ≤ n : N(T RIPLE gi 1 ⊆ T RIPLE gi 2 andT RIPLE gi 2 ⊆ T RIPLE gi 1 ). This relation is represented by ‘=s ’ symbol. Intensional equality: Two graphs g1 and g2 , with name maps N1 and N2 , are said to be intensionally equal, i.e., g1 =n g2 , if g1 =s g2 and for each node n1 and property p1 in g1 , there is a node n2 and property p2 in g2 such that N1 (n1 ) = N2 (n2 ); N1 (p1 ) = N2 (p2 ). Definition 11 (Instance). A graph gi with naming function Ni is called an instance of graph g (naming function N ) if gi =s g and there is a ground substitution subs such that – for all ground nodes (n) and properties (p) in g, there is a ground node (ni ) and a ground property (pi ) in gi with N (n) = N (ni ) and N (p) = N (pi ). – for all variable nodes (n) and properties (p) in g, there are ground nodes (ni ) and properties (pi ) in gi , such that subs(n) = N (ni ) and subs(p) = N (pi ).
5 Semantics In this section we provide limited semantics for finite types introduced earlier. To do so we start with a universe, i.e., a set of URI’s (denoted U RI) and other constants, say B as urelements (i.e., those atomic elements that do not have any further structure to them [12,13]). We do not add a URI to represent a blank node. Hierarchical universe is built as follows: N0 and P0 : N0 = {(u, b) : u ∈ U RI and b ∈ B} P0 = {(u, (b, b )) : u ∈ U RI and b, b ⊆ B}
RDF Collections
209
Nn+1 and Pn+1 : Suppose Nn and Pn have been defined. Let Un = {x : ∃u ∈ U RI(u, x) ∈ Nn }. Then define Nn+1 = {(u, B) : u ∈ U RI and B ⊆ P(Un ) ∪ B(Un ) ∪ S(Un )} Pn+1 = {(u, (B, B )) : u ∈ U RI and B, B ⊆ Un × Un } We use this stratified named universe to interpret RDF graphs and related syntax as follows: Definition 12 (The Semantic Mapping ) Universe: 1. Suppose U is the universe as defined in definition 8. 2. Let u : U → B be the surjective mapping that maps the constants in the syntax to constants in collection of urelements over which the syntax is interpreted. Atomic constants are mapped to themselves, sets over atomic constants of type T (Set[T]) are mapped to sets of atomic constants (Set[T]) in B. Similarly, bags and lists over atomic types are mapped to themselves in B. Collections over next higher types are mapped similarly to collections over constants of the same type and so on. 3. Let v : −U → B be a mapping that maps variables in the syntax to constants in the collection of of urelements over which the syntax is interpreted. Variables over atomic types T (−T ) are mapped to constants of type T in B. Variables of type Set[T], Bag[T] and List[T] are mapped to constants of corresponding types on B, given atomic variables are mapped to constants in B. Similarly, collections over higher types are mapped to collections over constants of same types in B, given all variables of lower order have been mapped to constants of corresponding types. 4. Let name : N AMES → U RI be a mapping of RDF names to URIs. Mapping nodes: Every node n ∈ Nn named a is mapped as (a, n) = ( a name , n ) where n is constructed by replacing every constant cT ∈ U in n with cT u and every variable −T ∈ U in n with −T v . Mapping properties: For every property p ∈ Pn named a is mapped as (a, p) = ( a name , p ) where p is constructed by substituting every constant cT ∈ U in p with cT u and every variable −T ∈ U in p with −T v . Interpreting triples: Triple (s, p, o) is interpreted as (s, p, o) = s ∧ p ∧ o . Interpreting rdf:type: We say that x, rdf:type, y iff x = (a, b), y = (p, q) and b ∈ q . Interpreting rdfs:subClassOf: x, rdfs:subClassOf, y iff x = (a, b), y = (p, q) and b ⊆ q . Interpreting names: The name of an object or a property x is defined as the first coordinate of x . Interpreting intensional equality: x is said to be intensionally equal to y (written x =int y) iff x = y
210
S. Kaushik and D. Wijesekera
Interpreting structural equality of nodes and properties: x is said to be structurally equal to y, written x =str y iff (a, b) = x , (p, q) = y and b = q. Interpreting structural equality of triples: (s, p, o) is said to be structurally equal to (s , p , o ), written (s, p, o) =str (s , p , o ) iff (s, p, o) ∧ (s , p , o ) ∧ s =str s ∧ p =str p ∧ o =str o Interpreting intensional equality of triples: (s, p, o) is said to be intensionally equal to (s , p , o ), written (s, p, o) =int (s , p , o ) iff (s, p, o) = (s , p , o ) Interpreting ground graph: Ground graph {(s, p, o)} interpretation is constructed by interpreting each triple in the graph graph (s, p, o) Interpreting graph: Graph g interpretation is constructed by interpreting each ground triple in the g with the existential closure of all variables in g.
6 Graph Entailments In this section we make precise Hayes and McBride’s [4] informal treatment of graph entailments that include higher order constructs. Definition 13 (Entailment). We say that a graph g entails another graph g if g ⊆ g . We write this as g g . Otherwise we say g does not entail g , written as g g Graph entailments simply state the satisfiability of a graph given a related graph is satisfied. Since we consider finite graphs, compactness is easily shown. Lemma 1 (Graph Entailments). Following results by Hayes and McBride’s [4] hold in our framework: Empty graph entailment Given g:GRAPH, g entails empty graph, i.e., g {} Empty graph Given g:GRAPH and g is not an empty graph, g is not entailed by empty graph, i.e., {} g Subgraph entailment Given g,f:GRAPH and g f , then g f . Merging (L) Given g,f:GRAPH then g ∧ f g ∪ f . Merging (R) Given g,f:GRAPH then g ∪ f g ∧ f . Monotonicity Given g,f,h:GRAPH, g f and f h then g h Proof Sketch: Empty graph entailment Let g:GRAPH and g = {}. Now, ∅ ∈ g . Empty graph Let g:GRAPH, g = {}. By definition, g = ∅. Subgraph entailment Let g,f:GRAPH and g f . By definition, for all i ≤ n : N(T RIPLE fi ⊆ T RIPLE gi . Clearly, whenever g is defined f ⊆ g . Merging L By definition, if g1 , g2 : GRAPH, then ∀i(T RIPLE gi 1 ∪g2 = T RIPLE gi 1 ∪ T RIPLE gi 2 ). Clearly, with N∪ = Ng1 ∪ Ng2 it is easy to show g1 ∧ g2 g1 ∪ g2 . This is easily extended to a set of graphs. Merging R Shown in the similar way as in previous case. Monotonicity Shown similarly using subgraph lemma.
RDF Collections
211
Next we state an entailment relation for collections. That is, given a property for a collection of objects, we can infer a higher order property for the collection with the support for a simple arithmetic (Presburger arithmetic [14]). Lemma 2 (Collection Entailment). If g:GRAPH, x:[T], ∀x ∈ C[T ](∃ triple (x, f unc1 , n : N)), and f unc1 , f unc2 are mathematical functions related to one another by simple arithmetic on collection C[T]. g, x : T, ∀x ∈ C[T ]((x, f unc1 , N)) g ∪ {(C[T ], f unc2, N )} Proof Sketch: Follows from subgraph lemma and basic arithmetic. Collection entailment states that if there is a graph, say g, containing a node, say n, that is a collection (set, bag or a sequence) and each of its members have an arithmetic property, then g entails a graph that expresses a property of a collection of n in addition to all other properties expressed by g. This property is derived by basic arithmetic over the lower order arithmetic properties. We revisit our running example to show collection entailment in action. We omit RDF syntax in our example, however, it is easily checked that a simple translation to RDF syntax exists. Example 2. Consider an RDF graph g with some stage k nodes, xi , i ≤ n : N of type T such that for each xi there is a node yi that is a number (literal), related to xi through a property pk . Also, there is a stage k + 1 node X of type rdf:Bag[T], with each xi as its member. This graph models the shopping bag in example 1, where each member of a bag has a price. Because this property is functional, we can use a function to represent such properties. That is, we express pk as a function pk (x) := x ∈ X → N for all members xi of the bag X. Clearly, it’s easy to compute the total value of all the items stored in X by adding them all together: pk (x)) p(X) = ∀x ∈ X( x
Thus, p is a function whose domain is a shopping bag with number valued items as members and range is N. That is, we can now express a stage k + 1 property for a collection type, expressible in an RDF graph, say pk+1 . More concretely, suppose X is [[milk, yogurt, yogurt]] and graph g has triples (milk, hasPrice, 3) and (yogurt, hasPrice, 1). Then, using basic arithmetic, we can deduce a property (ShoppingBag, total, 3+1+1) or (ShoppingBag, total, 5). In summary, using the collection entailment lemma, we were able to deduce a new graph from g, with an additional triple that identifies a functional relation of the shopping bag to the set of natural numbers.
7 Conclusions RDF syntax allows for ontologists to express higher-order types like sequences, bags and alternatives. These collections are essential for representing many real world facts
212
S. Kaushik and D. Wijesekera
like work flows, shopping bags, payment receipts, etc. Although collections can be expressed, they are not interpreted in RDF. Consequently, Semantic Web agents cannot inferences useful facts about these collections. Authoritative specifications of RDF syntax and semantics [3,4] acknowledge the need for building these inferences into RDF, however, to the best of our knowledge, this deficiency has not yet been addressed in the literature. In this paper, we provide an approach to construct three higher-order types, viz., sets, bags and sequences. We integrate these types into a reified RDF framework with a hierarchical semantics. To incorporate collections into graph entailments, we formalize graph entailments [4] with collections in reified graphs and prove collection entailments for incorporating deductions based higher-order types.
References 1. Burners-Lee, T., Hendler, J., Lassila, O.: The semantic web. Scientific American (2001) 2. Brickley, D., Guha, R.: Resource Description Framework (RDF) Schema Specification 1.0: RDF schema. W3C workding Draft (2003) 3. Guha, R., Hayes, P.: LBase: semantics for languages of the semantic web. W3C workding Draft (2003) 4. Hayes, P., McBride, B.: RDF semantics. W3C Recommendation (2004) 5. Fikes, R., McGuinness, D.L.: An axiomatic semantics for rdf, rdf-s, and daml+oil. Technical Report Note 18, World Wide Web Committee (W3C) (2001) 6. Harrison, R.: Abstract Data Types in Standard ML. John Wiley and Sons, Chichester (1993) 7. Grumbach, S., Milo, T.: Towards tractable algebras for bags. In: Twelfth ACM SIGACTSIGMOD-SIGART Symposium on Principles of Database Systems, pp. 49–58 (1993) 8. Albert, J.: Algebraic properties of bag data types. In: 17th International conference on Very Large Databases, pp. 211–219 (1991) 9. Dayal, U., Goodman, N., Katz, R.H.: An extended relational algebra with control over duplicate elimination. In: 1st ACM SIGACT-SIGMOD symposium on Principles of database systems, pp. 117–123 (1982) 10. Kent, W.: Profile functions and bag theory. Technical Report HPL-SAL-89-19, HewlettPackard Laboratories, Palo Alto, CA (1989) 11. Mumick, I., FinkelStein, S., Pirahesh, H., Ramakrishnan, R.: Magic is relevant. In: ACM SIGMOD international conference on Management of data, pp. 247–258 (1990) 12. Aczel, P.: Non-well-founded sets. Lecture Notes, vol. 14. CSLI (1988) 13. Kunen, K.J.: Set Theory. Reprint edn. North Holland, Amsterdam (1983) 14. Presburger, M.: Ueber die vollstaendigkeit eines gewissen systems der arithmetik ganzer zahlen, in welchem die addition als einzige operation hervortritt. In: Comptes Rendus du I congr´es de Math´ematiciens des Pays Slaves, Warsaw, Poland, pp. 92–101 (1929)
Logging and Analyzing User’s Interactions in Web Portals Gennaro Costagliola, Filomena Ferrucci, Vittorio Fuccella, and Luigi Zurolo Dipartimento di Matematica e Informatica, Università di Salerno Via Ponte Don Melillo, I-84084 Fisciano (SA) {gcostagliola,fferrucci,vfuccella}@unisa.it, [email protected]
Abstract. Content Management Systems and Web Portal Frameworks are more and more widely adopted in Web development. Those kinds of software often produce Web pages whose layout is divided in sections called, in the case of Web Portals, “portlets”. Portlets can be produced by different sources and then aggregated in the same page by the portal. For Web portals, traditional Web metrics based on page visits can be inadequate for fully understanding user’s interest, due to the heterogeneity of content and the variety of sources. This paper proposes a system for evaluating the Web traffic at a deeper level than the page visit one: the level of the sections, or of the portlets. The interest of the user in the sections of the page is gauged through implicit interest indicators, such as, section visibility, mouse movements and other client-side interactions. Our system is composed of two different products: a framework that, opportunely instantiated in a Web portal, allows the production of a log, and a log analyzer. The possible uses and benefits gained by research in the fields of Web traffic analysis, portal design and usability are investigated in depth. Keywords: Web, www, metrics, portal, portlet, JSR 168, WSRP, logging, log, query, log4p.
concern of the portal). The development of standards and specifications, such as JSR 168 [17] and WSRP [24], has helped to this extent. A noteworthy example of a Web site, whose layout demonstrates the use of a portal framework in its development, is that of Yahoo. A screenshot of its home page is shown in figure 1.
Fig. 1. The Yahoo home page and its sections
As a result of the aggregation, a portal page can contain a highly heterogeneous content, taken from various portlet producers. For Web sites developed with CMS and portal framework technology, traditional metrics based on page visits can be inadequate to fully understand user’s interest: new forms of metrics are needed. What we need are metrics which can give information at a deeper level than that of page visits: the level of the sections, or of the portlets. Unfortunately, at present there is a lack of these metrics. This paper presents some tools which make use of new metrics suitable for describing the behaviour of the visitors on the portal pages. The gathering of such information is carried out through the use of a framework directly instantiated in the portal. The framework produces an XML log, which includes raw data, such as the implicit interest indicators (i. e. any interaction of the user with the portlets) and the visibility of the portlets in the pages, captured when the users browse the portal pages. The logs are analyzed through a suitable log analyzer that obtains some higher level information by performing several queries on the logs, such as:
An estimation of the visibility of the portlets in the page. The interactivity level of the portlets.
From this information we can obtain an estimation of the interest shown by the users in the portlets. The analysis can be made available per single user visit and session, across multiple visits of the same user or for all the users.
Logging and Analyzing User’s Interactions in Web Portals
215
Once obtained, these data can be used for multiple purposes. The possible uses and benefits gained by research in fields of Web traffic analysis, portal design and Web usability are investigated in depth. Furthermore, the same principle of logging user’s interactions with a Web-based interface and analyzing the gathered data in order to obtain interesting information about user’s behavior is applicable in other circumstances: our approach has been exploited for analyzing the strategies used by the learners for executing on-line educational tests in an e-learning system. The rest of the paper is organized as follows: section 2 gives the knowledge background necessary to understand some concepts on which the system is based. The system is presented in section 3: the section is composed in two sub-sections, the first to describe the logging framework and the second for the log analyzer. Lastly, in section 4, we briefly discuss possible uses and benefits of our system. Several final remarks and a discussion on future work conclude the paper.
2 Background A portal is a Web site which constitutes a starting point, a gate to a consistent group of resources and services on the Internet or in an intranet. Most of the portals were born as Internet directories (as Yahoo) and/or as search engines (as Excite, Lycos, etc.). The offer of services has spread in order to increment the number of users and the time they spend browsing the site. These services, which often require user registration, include free email, chat rooms, and personalization procedures. In the history of portals, many authors identify two generations. Second generation Web portals distinguish themselves from first generation ones for their architecture, which is component-oriented. In particular, the basic component constituting them, is often referred to as portlet. The portal is responsible for aggregating information coming from different sources, local or remote, available in the form of mark-up fragments. Each of these fragments is produced by a portlet. In the context of Web portals, the possibility of deploying a portlet in any portal is particularly significant. To this extent, that is, to achieve interoperability among portals, it has been necessary to define a standard way to develop and deploy portlets. Two main standards have been defined and widely adopted by producers: the Web Services for Remote Portlets (WSRP) and the Java Portlet Specification and API (JSR 168). The former is more oriented to the definition of rules about the use of remote portlets, the latter is focused on the definition of interfaces for the development of portlets which can run in Java-based portals. WSRP defines a Web service interface through which portals can interact with the remote producer’s portlets. The WSRP 1.0 specification was approved as an OASIS standard in August, 2003. Being based on Web services, several interfaces to adopt the standard have been developed for the most used technologies (e.g. J2EE, .NET, and so on). Most of the Java technologies, part of the Java 2 Enterprise Edition, the platform for the development and deployment of distributed enterprise applications, follow a consolidated architectural model, called container/component architecture. This model offers the chance to develop components and deploy them on different containers. Both component and containers compliant to specifications can be developed independently and commercialized by different software vendors, thus creating a market economy on Java software. Furthermore, several good-quality Open Source products compete with them. The JSR 168 follows the container/component
216
G. Costagliola et al.
Fig. 2. JSR 168 compliant portal architecture
model and its adoption has grown until it has become an important reference-point which cannot be excluded from the projects aimed at the development of Web portals. Among other things, it defines the architectural model of conformant portals. Its main constituents are the portlet, which produces content mark-up, the portal, which aggregates the mark-up, and the portlet container, which manages the portlet lifecycle and provides an API to the portlets for accessing to a set of services. The typical architecture for a JSR 168 conformant Web portal is shown in figure 2.
3 The System Our system is aimed at obtaining detailed statistics in order to have a deep analysis of the user’s interest in the Web portal and, in particular, in the sections which compose its pages. Our system is composed of two different pieces of software: 1
2
A logging framework, called Log4p, to be used by Web portal developers, which, instantiated in the Web portal, is in charge of capturing information about user’s behavior during the navigation of the portal and storing it in an XML-formatted log. A log analyzer, to be used by Web portal administrators, developed as a stand-alone application, which is responsible for analyzing the data gathered by the logger.
3.1 Log4p: The Logging Framework The purpose of the Logging Framework is to gather all of the user actions during the browsing of the portal and to store raw information in a set of log files in an XML format. The framework is composed of a server-side and a client-side module. The clientside module is responsible for “being aware” of the behavior of the user while he/she is browsing the portal pages. The server-side module receives the data from the client and creates and stores log files on the disk. Despite the required interactivity level, due to the availability of AJAX (Asynchronous JavaScript and XML), the new technology for increasing the interactivity of Web content, it has been possible to implement the client-side module of our framework without developing plug-in or external modules for Web browsers. Javascript has been used on the client to capture user interactions and the text-based communication between the client and the server has been implemented through AJAX method calls. The client-side scripts can be added to the portal pages with a light effort by the programmer.
Logging and Analyzing User’s Interactions in Web Portals
217
Fig. 3. The information model for log data
The events captured by the framework are the following:
Actions undertaken on the browser window (opening, closing, resizing) Actions undertaken in the browser client area (key pressed, scrolling, mouse movements)
The event data is gathered on the browser and sent to the server at regular intervals. It is worth noting that the event capturing does not prevent other scripts present in the page to run properly. The server-side module has been implemented as a Java servlet which receives the data from the client and prepares an XML document in memory. At the end of the user session the XML document is written to the disk. To reduce the size of log files, a new file is used every day. The information model used for the log data is shown in figure 3. The model is organized per user session. At this level, an identifier (if available) and the IP of the user are logged as well as agent information (browser type, version and operating system). A session is composed of page visits data. For every page, the referrer is logged and a list of events is present. The data about the user interactions are the following:
Event type HTML source object involved in the event Portlet containing the HTML object and its position in the page (coordinates of the corners)
218
G. Costagliola et al.
Mouse coordinates Timing information (timestamp of the event) More information specific of the event
An important concern in Web metrics is the log size. In very crowded Web sites, even simple HTTP request Web logs can reach big sizes. A configuration system, including the following configuration settings has been conceived above all to reduce log sizes:
List of events to capture; List of portlets to monitor; Time interval between two data transmissions from the client to the server; Sensitivity for mouse movements; Sampling factor (logging is activated only for a random user out of n users).
The configuration is read by the server-side module but affects the generation of the javascript modules run on the client-side. The architecture of the framework is graphically represented in figure 4.
Fig. 4. Logging framework architecture
On the client machine, everything can be done in the Web Browser. The Javascript modules for event capturing, dynamically generated on the server, are downloaded and run in the browser interpreter. Data is sent to the server through an AJAX request. On the server-side, a module called RequestHandler receives it. Once received, a module called LoggerHandler organizes the XML document in memory and flushes it to the disk every time a user session finishes. 3.2 The Log Analyzer The next phase of the data gathering is the data analysis. In our system this is done through a Web-based stand-alone application, optionally hosted on the same server of the logging framework, which takes in input the log files. The analysis phase consists of a series of analysis on the behavior of the user, starting from the data stored in the log. The analysis can have several aims. Among them we can cite:
Giving a better organization to the portal; The choice of the contents more suitable to the user or to group of users; Usability analysis of the portal.
A deeper analysis on the uses that our system can offer is contained in the next section.
Logging and Analyzing User’s Interactions in Web Portals
219
The analyzer performs queries on the logs to obtain the desired data and then calculates statistical indicators, shown in the form of charts and tables. Since the log files are in XML, the query engine has been developed to understand XQuery [25] language, and has been carried out by using an implementation of the JSR 225 [16]. In the next sub-sections we will show some useful statistics we can obtain using the system. The module for drawing charts has been developed using a free Java library, named jCharts [13]. PageCounts. A generic analysis is given by the simple count of page visits. Even though such a task is easily performed by a lot of already existing tools, page visits count is an important statistic for our system, since it allows us to understand on which page the interest of the user is mostly concentrated. Starting from this data, we could focalize our attention on a subset of the portal pages and perform a deeper, portlet-based, analysis on them. This statistic is easily obtainable with our analyzer through a count query on the page elements of our logs. A sample of page visits chart, drawn using our analyzer, is shown in figure 5.
Fig. 5. Page count chart sample
Portlet Visibility Time. A more specific analysis can be obtained by calculating the visibility time for each level of portlet visibility (total, partial, invisible). Portal layouts are usually organized per columns. A commonly used layout organizes portlets in two columns of the same size (50%, 50%). Another common layout is composed of three columns (i.e. 25%, 50%, 25% in size). The portal page can contain some other elements, such as, a header, a footer and an horizontal or a vertical menu or both of them. If the number and size of portlets is such that the portal page exceeds the size of the browser window, only a part of the page is shown, while some other parts are hidden and can be shown through scrolling. Thus, at any time some portlets can have full visibility, some others partial visibility (only a percentage of the portlet area is visible), while the remaining are completely hidden to the user. Every time a scrolling event occurs, our logger records its timestamp and the position of the portlet in the page. The availability of this data allows us to precisely
calculate the amount of time the portlets were fully visible, partially visible or completely invisible during the visit. This information, in the context of a portal, is very useful, since, after knowing which page has attracted the user more, it let us know his/her interest in the single content units of the page. Figure 6 contains a chart showing the visibility percentage of the portlet. When a portlet is partially visible, the chart of figure 6 does not exactly tell us the extent of the visible and hidden areas of the portlet. Thus, in order to complete the visibility analysis, we considered it opportune to show another chart summarizing the visibility percentage of the portlet across users’ page views. This indicator is calculated as the weighted mean of all the visibility times, using the following formula: V=
∑ (t
i
i
* vi )
(1)
T
V is the total visibility indicator for the portlet, ti and vi are, respectively, its time and percentage of visibility in the i-th interval, T is the total time of the page visit.
Logging and Analyzing User’s Interactions in Web Portals
221
A sample of the chart is shown in figure 7. For the sake of readability, the bars are shown in green for high visibility, in yellow for low visibility and in red for scarce visibility. The elaboration and the queries performed on the log for obtaining the parameters in (1) , have been reported in appendix. Portlet Interactions. Some portlets can be more informative while others can be more interactive. For example, an article of an on-line news magazine is supposed to be informative, while a section containing a form should be more interactive, that is, it should receive more user interactions than the former. Many people use the mouse as a pointer while reading on-line news.
Fig. 8. Portlet interactivity level chart sample
Fig. 9. Portlet mouse pointer focus
With our tool, we can obtain an information about portlet interaction from a bar chart. An example is shown in figure 8. In the chart, each bar is composed of three sections of different color. They represent three different types of interactions: window, mouse and keyboard events.
222
G. Costagliola et al.
Another interest indicator to be considered is the total time a portlet has the mouse pointer in it. An eye tracking study [7] shows that there is a significant correlation between the eye movements and the mouse movements: tracking the trajectory drawn by the mouse pointer could be useful for obtaining the probable trajectory of user’s eyes, that is, what the user is interested in. While eye tracking cannot be performed, if not in ad hoc equipped laboratories, mouse tracking can be easily performed by our tool. All of the mouse movements can be reconstructed from the log analysis. With a great number of users, the reconstruction of all mouse movements can be too onerous. A similar interest indicator can be obtained by just calculating the amount of mouse movements and the amount of time spent by the user with the mouse pointer inside a portlet. The number of movements has already been described and charted in figure 8. As for the amount of time the portlet has the mouse pointer in it, a pie chart, showing the times of presence of the mouse pointer inside the portlets of the same page, appears as the most appropriate choice. A sample is shown in figure 9.
4 Related Work, Uses and Applications The effectiveness of implicit interest indicators is witnessed by several studies in literature [5; 20]. These works demonstrate the correlation between the implicit indicators and the actual interest of the user. Furthermore, such studies produce a list of the most used interest indicators. Some works propose the development of tools for determining user’s interests. For example, in [1] a proxy server based system is used to capture client-side interactions. Other portal framework, such as Websphere Portal Server [22], also analyze the user behavior, but their analysis is based on classical clickstream metrics. Some works are aimed at understanding the structure of Web pages in order to determine page sections. Many algorithms have been presented to this extent. The purposes of determining page sections include the detection of similarities among pages, the adaptation of the pages to small screens, the detection of the most significant content of the page, etc. Wenyn et al. [23] divide pages in sections to detect similarities between two pages, in order to prevent phishing. Chen et al. [6] do the same thing in order to better view Web pages on small screen devices. Blocks in the pages are detected for identifying the informative sections of the page to reduce storing sizes for search engines [10] or to eliminate redundant information for Web mining [21]. It is clear that, if efforts have been made to divide pages in sections, where the pages are explicitly divided in sections, we can pursue the above discussed objectives and do much more. Once determined, the indication of user interest, calculated from the log analysis, can be used for several purposes. The next sub-sections analyze the possible practical uses of our system in several Web research fields. 4.1 Integrating Web Metrics Web metrics tell us how the users are using a Web site. E-commerce sites need to know this information in order to improve their selling capacity. Some of the most commonly used Web metrics are: the number of page visits, the number of banner or link clicks, the percentage of users who complete an action, etc.
Logging and Analyzing User’s Interactions in Web Portals
223
Unfortunately, clickstream analysis metrics have the following limitations, as remarked by Weischedel and Huizingh [22]:
They report activity on the server and not user activity; They can overestimate the actual use of Web sites due to spiders activity; They do not include the real time spent on the page by the user.
Our system overcomes these problems, in fact, it captures client side activity, can easily recognize spiders from the absence of mouse movements and records times, in such a way that it is easy to detect inactivity time due to user absence from the screen. 4.2 Portal Design The location of the portlets in the portal page has a great importance, since some portlets can have more visibility than others. An eye tracking study [12], analyzing the behaviour of users in some browsing tasks, has shown that user’s interest is more concentrated, at least in the initial phases of page browsing, in the portlets placed on the top of the first column on the left. In the same study, a complete classification of the places which are candidates to gain more user interest has been performed. It is advisable that, if the portal holder wants to emphasize the content of a portlet more than another, he/she should put these portlets in those places. Our tool can help in determining the portlets which attract user’s interest more and, on the basis of this data, it can help portal administrator in placing the portlets in the pages. 4.3 Personalization of the Portal Web portal customization is often used to tailor the services of the portal to a single user or to a group of users. In some cases, the user has the freedom of choosing his/her favourite portlets to place in his/her home page. In other cases, the interest of the user can be inferred from the logs, and the pages of the portal constructed in order to give more visibility to content which matches user’s interests. Our system can help in the latter case. Moreover, in the case of groups of users, groups can be obtained through clustering procedures. Our system can be useful for gathering data to obtain cluster of users. 4.4 Portal and Portlet Usability Since portlets can be considered small Web applications, the definition of usability can be extended for portlet usability. Diaz et al. [11] define it as the capability of a portlet to be understood, learned or used under specified conditions. The implicit interest indicators can be used to facilitate the task of usability evaluators. Atterer et al. [1] show many situations in which this is true, for example by using true users instead of volunteers in the lab. Furthermore, a study [19] states the possibility of performing a choice among different portlets with similar features, choosing on the basis of their usability. Our tool can be useful to this extent, in order to isolate the interactions relative to a given portlet thus evaluating its usability. The position of the portlets affects the usability of the portal. For example, let us suppose that a task can be performed by interacting with more than one portlet. The position of the portlets involved in the task can affect the amount of time necessary to
224
G. Costagliola et al.
perform the task itself. A usability study can be aimed at finding the best location for each of these portlets. Our system allows the usability evaluators to record and evaluate the action of the users in all of the different portlet arrangements. Lastly, our tool captures key press events. In the case of portlets with forms, the data can be used to understand if the user had problems in filling the form. This is valid for any kind of Web site and not only for portals. 4.5 Logging and Analyzing Learner Behavior in On-Line Testing An approach similar to the one described so far has been used in on-line testing in order to log and analyze learner behavior in on-line testing. On-Line Testing, also known as Computer Assisted Assessment (CAA), is a sector of e-learning aimed at assessing learner’s knowledge through e-learning means. Tracking learner’s interactions during the execution of a test can be useful for understanding the strategy used by the learner to complete the test and for giving him/her advise on how to perform better in future tests. Several experiments have been performed to this aim [2; 15; 18] in the past. Our approach for on-line testing is rather similar to the one used for Web portals: learner’s interactions during tests based on multiple choice questions have been logged and stored in XML files, then the information gathered is analyzed and visualized in a suitable chart. The interactions have been logged by instantiating a slightly modified version of Log4p in an on-line testing system, called eWorkbook [8]. This system presents the questions, one at a time, in a Web page. A screenshot of the test interface is shown in figure10. The learner can browse the test questions by clicking on the next and previous buttons in the bottom of the page. The page is composed of different sections (in analogy to portlets) showing, respectively, the stem and the options of the question. The framework records the interactions of the user with the above described interface, including response and browsing events and mouse events.
Fig. 10. A Screenshot of the Test Execution
Logging and Analyzing User’s Interactions in Web Portals
225
A chronological review of the test has been made available through a chart, obtained by showing the salient points of a test execution, synthesized in the interactions recorded in the log file. This chart shows, at any time, the item browsed by the learner, the mouse position (intended as the presence of the mouse pointer on the stem or on one of the options) and the presence of response type interactions, correct or incorrect. The chart is two-dimensional: the horizontal axis reports a continuous measure, the time, while the vertical axis displays categories, the progressive number of the item currently viewed by the learner. The test execution is represented through a broken line. The view of an item for a determined duration, is shown through a segment drawn from the point corresponding to the start time of the view to the one corresponding to its end. Consequently, the length of the segment is proportional to the duration of the visualization of the corresponding item. A vertical segment represents a browsing event. A segment oriented towards the bottom of the chart represents a backward event, that is, the learner has pressed the button to view the previous item. A segment oriented towards the top is a forward event.
Fig. 11. Graphical Chronological Review of a Sample Test
The responses given by a learner on an item are represented through circles. The progressive number of the chosen option is printed inside the circle. The indication of correct/incorrect response is given by the filling color of the circle: a blue circle represents a correct response, while an incorrect response is represented through a red circle. The color is also used for representing the position of the mouse pointer during the item view. The presence of the mouse pointer in the stem area is represented through a black color for the line. As for the options areas, the red, yellow, green, blue and purple colors have been used, respectively, for 1 to 5 numbered options. More than 5 options are not supported at present. Lastly, grey is used to report the presence of the mouse pointer in a neutral zone. The graphical chronological review of a sample test is shown in figure 11. By analyzing the charts obtained in an experiment carried out during a laboratory exam involving approximately 80 learners, we can conclude that learners often follow
226
G. Costagliola et al.
common strategies for completing on-line tests. In our experiment we have identified the following three different strategies: o
o
o
Single Phase. This strategy is composed of just one phase (a part of the test execution needed by the learner for sequentially browsing all of the questions). The time available to complete the test is organized by the learner in order to browse all the questions just once. The learner tries to reason upon a question for an adequate time and then gives a response in almost all cases, since he/she knows that there will not be a revision for the questions. Eventual phases subsequent to the first one have a negligible duration and no responses. Active Revising. This strategy is composed of two or more phases. The learner intentionally browses all the questions in a shorter time than the time available, in order to leave some time for revising phases. The questions whose answer is uncertain are skipped and the response is left to subsequent phases. As a general rule, the first phase lasts a longer time and the subsequent phases have decreasing durations. Passive Revising. This strategy is composed of two or more phases. The learner browses and answers all the questions as fast as possible. The remaining time is used for one or more revising phases. As a general rule, the first phase lasts a longer time and the subsequent phases have decreasing durations.
For both the definition of the strategies and the classification of test instances, the charts have been visually analyzed by a human operator. The above tasks are rather difficult to perform automatically, while a trained human operator can establish the strategy used by the learner from a visual inspection of the charts of the test instances and giving advice to the learners on how to perform better next time. Other uses of the above described method are the detection of correlation among questions and the detection of cheating during tests. The reader can refer to a separate paper [9] for obtaining a more detailed description on the educational results of our experiment.
5 Conclusions In this paper we have presented a system aimed at obtaining and analyzing the data about the behaviour of the users of Web portals. The system overcomes the limitation of the simple page visit-based metrics, giving more valuable information related to the portlets, such as their visibility and interactivity and, consequently, the interest of the user in them. The system is composed of two components: a framework for obtaining XMLbased log files and an application for log analysis. Several charts, drawn using the analyzed data have been shown. Referencing some recent work in literature, we have argued that our system can be useful for numerous purposes, such as, integrating Web metrics, optimizing portlet layout both for all users and for personalization, studying usability of portals and, in a
Logging and Analyzing User’s Interactions in Web Portals
227
slightly modified version, analyzing learner behavior in on-line testing. Future work is aimed at further demonstrating the use of our system in some of these fields. Finally, an aspect that has been considered, but not yet put into practice, is the availability of the log data both to the portal and to the portlet producer. At present some architecture and secure schemas have been taken into account, as the one proposed by Blundo and Cimato [4], applied for determining banner clicks in advertising campaigns. The system has been tested on a portal developed with Apache Jetspeed II [14] Portal framework.
References 1. Atterer, R., Wnuk, M., Schmidt, A.: Knowing the User’s Every Move – User Activity Tracking for Website Usability Evaluation and Implicit Interaction. In: Proceedings of the 15th international conference on World Wide Web WWW 2006, ACM Press, New York (2006) 2. Bath, J.A.: Answer-changing Behaviour on objective examinations. The Journal of Educational Research 61, 105–107 (1967) 3. Bellas, F.: Standards for Second-Generation Portals. IEEE Internet Computing 8(2), 54–60 (2004) 4. Blundo, C., Cimato, S.: A Software Infrastructure for Authenticated Web Metering. IEEE Computer (2004) 5. Claypool, M., Le, P., Wased, M., Brown, D.: Implicit interest indicators. In: Proceedings of the 6th international conference on Intelligent user interfaces, ACM Press, New York (2001) 6. Chen, Y., Xie, X., Ma, W.Y., Zhang, H.J.: Adapting Web pages for small-screen devices. IEEE Internet Computing (2005) 7. Chen, M.C., Anderson, J.R., Sohn Moore, M.H.: What can a mouse cursor tell us more?: correlation of eye/mouse movements on Web browsing. In: CHI 2001 extended abstracts on Human factors in comp. syst. (2001) 8. Costagliola, G., Ferrucci, F., Fuccella, V., Oliveto, R.: eWorkbook: a Computer Aided Assessment System. International Journal of Distance Education Technology 5(3), 24–41 (2007) 9. Costagliola, G., Fuccella, V., Giordano, M., Polese, G.: A Web-Based E-Testing System Supporting Test Quality Improvement. In: Proceedings of The 6th International Conference on Web-based Learning, pp. 272–279 (2007) 10. Debnath, S., Mitra, P., Pal, N., Giles, C.L.: Automatic identification of informative sections of Web pages. In IEEE Transactions on Knowledge and Data Engineering (2005) 11. Diaz, O., Calero, C., Piattini, M., Irastorza, A.: Portlet usability model. IBM Research Report. RA221(W0411-084). ICSOC 2004. pp. 11–15 (2004) 12. Goldberg, J.H., Stimson, M.J., Lewenstein, M., Scott, N., Wichansky, A.M.: Eye tracking in Web search tasks: design implications. In: Proceedings of the 2002 symposium on Eye tracking research & applications (2002) 13. jCharts, Krysalis Community Project – jCharts (2006), http://jcharts.sourceforge. net/ 14. Jetspeed 2, Apache Group. Jetspeed 2 Enterprise Portal (2006), http://portals. apache.org/jetspeed-2/ 15. Johnston, J.J.: Exam Taking speed and grades. Teaching of Psychology 4, 148–149 (1977)
228
G. Costagliola et al.
16. JSR 225, JSR 225: XQuery API for JavaTM (XQJ) (2006), http://jcp.org/en/ jsr/detail?id=225 17. JSR 168, JSR-000168 Portlet Specification (2003), http://jcp.org/aboutJava/ communityprocess/review/jsr168/ 18. McClain, L.: Behavior during examinations: A comparison of “A”, “C” and “F” students. Teaching of Psychology 10(2) (1983) 19. Moraga, M.A., Calero, C., Piattini, M.: Ontology driven definition of a usability model for second generation portals. In: Workshop proceedings of the sixth int. conference on Web engineering, ICWE 2006 (2006) 20. Shapira, B., Taieb-Maimon, M., Moskowitz, A.: Study of the usefulness of known and new implicit indicators and their optimal combination for accurate inference of users interests. In: Biham, E., Youssef, A.M. (eds.) SAC 2006. LNCS, vol. 4356, Springer, Heidelberg (2007) 21. Taib, S.M., Yeom, S.J., Kang, B.H.: Elimination of Redundant Information for Web Data Mining. In: Proceedings of ITCC 2005, Int. Conf. on Information Technology: Coding and Computing, vol. 1 (2005) 22. Websphere. IBM WebSphere Portal Server, http://www-306.ibm.com/software/ genservers/portal/server/index.html?S_TACT=103BEW01&S_CMP=camp aign 23. Weischedel, B., Huizingh, E.K.R.E.: Website Optimization with Web Metrics: A Case Study. In: Proceedings of ICEC 2006, the 8th international conference on Electronic commerce, ACM Press, New York (2006) 24. Wenyin, L., Huang, G., Xiaoyue, L., Deng, X., Min, Z.: Phishing Web page detection. In: Proceedings of Eighth International Conference on Document Analysis and Recognition (2005) 25. WSRP, OASIS Web Services for Remote Portlets (2003), http://www.oasis-open. org/committees/tc_home.php?wg_abbrev=wsrp 26. XQuery, XQuery 1.0: An XML Query Language W3C Candidate Recommendation (2006), http://www.w3.org/TR/xquery/
Appendix As an example, the pseudo-code procedure used by the log analyzer for obtaining the portlet visibility percentage chart, follows. Each bar in the chart represents the percentage of visibility of a portlet across all page views. Those values are calculated using (1). The procedure assume, simplistically, that all of the analyzed pages contain the same portlets. Their number is passed as a parameter to the procedure (line 1), and is used to associate the correct event timestamp (lines 16-18) to portlet data (name and coordinates). Every time a user scrolls the page, the percentage of visibility of a portlet changes, since part or all of its area can fall inside/outside the browser’s client area. On the initial page load event, and on every scroll event, our log records the coordinates of each portlet (through the portlet element of the information model, see figure 3). Through our sample code, for each event element and for each portlet element, portlet names, coordinates and event timestamps are obtained by querying the log and by storing the results in the portletNames, coordinates and timestamps vectors, rispectively (lines 3-7).
Logging and Analyzing User’s Interactions in Web Portals
229
1 procedure showVisibilityChart(portletNum) 2 3 portletNames = let $x := //portlet return $x/@name; 4 coordinates = let $x := //portlet return 5 $x/@coordinates; 6 timestamps = for $x in //event where $x/@type="load" 7 or $x/@type="scroll" return $x/@timestamp; 8 9 old_time = timestamps[0]; 10 T = 0; 11 12 for(j=0; j < portletNames.length; j++) 13 14 v = calculateVisibilityPercentage(coordinates[j]); 15 16 if (j % portletNum == 0) 17 i = j/portletNum + 1; 18 time = timestamps[i]; 19 t = time - old_time; 20 T += t; 21 old_time = time; 22 23 sum{portletNames[j]} += v * t; 24 25 for each (portlet in sum) 26 visibility{portlet} = sum{portlet} / T; 27 28 createChart(visibility);
Once obtained event timestamps and portlet coordinates, the numerator in (1) is calculated through the iteration of lines 12-23. The partial sum is kept by the sum associative array (line 23), whose keys are portlet names. The calculation of the visibility percentage in the i-th time interval vi is delegated, as shown in line 13, to the calculateVisibilityPercentage sub-routine. The time intervals ti can be easily calculated by subtracting the (i+1)-th and the i-th timestamps (line 19). Those time intervals are summed in line 20 to obtain the total time T. The final results are put in the visibility associative array, as shown in line 25. Those results are obtained by dividing the partial sums by T. Lastly, visibility is passed to the createChart sub-routine, which is responsible for drawing the bar chart (line 27).
A Semantics-Based Automatic Web Content Adaptation Framework for Mobile Devices Chichang Jou Department of Information Management, Tamkang University 151 Ying-Chuan Road, Tamsui, Taipei 25137, Taiwan [email protected]
Abstract. With the rapid development of wireless communication technology, many users are accessing the internet from mobile appliances, such as notebooks, PDAs, and cellular phones. These devices are miscellaneously limited in computing resources, like CPU speed, memory, temporary storage, power supply, installed software, and communication bandwidth. The web accesses in these mobile devices thus encounter distorted user interface, broken images, slow responses, etc. To overcome this problem, we design and implement a semantics-based web content adaptation framework to provide automatically adapted contents to miscellaneous mobile devices. RDF Semantics of client device CC/PP configurations and web page tag structures are extracted to determine the proper parameters for the format and layout of web contents. Heuristic transcoding rules in the Jena Inference System are then applied to transform the web contents for each particular device. The functionality of this framework is illustrated by its capability of adjusting the layout of the main page of a shopping portal and of adjusting the parameters of images in the page. Keywords: Content adaptation, semantic web, CC/PP, Jena.
A Semantics-Based Automatic Web Content Adaptation Framework
231
Profiles) [12] client device configuration and web contents. These semantics properties will be stored inside the Jena Inference System [6] as knowledge facts to infer proper transcoding parameters for each particular device. The adapted web pages with proper layout modification and images with proper rendering parameters will be constructed and delivered. Thus, this technology is suitable for resource-limited devices to balance information loss and information availability. For the rest of this paper, we will review related work in Section 2. In Section 3, we describe the system architecture of this framework. Semantics extraction and knowledge base construction for the Jena Inference System are discussed in Sections 4 and 5. Section 6 demonstrates the content adaptation function with respect to the generated transcoding parameters. Section 7 explains the system implementation and illustrates through three mobile devices examples of rewriting one web page, and of adjusting the width and height parameters of images in that page. Section 8 concludes this paper.
2 Related Work Many researchers have tackled the incompatibility issue of mobile device in accessing web pages. Bickmore and Girgensohn [2] designed a “Digestor System” which is capable of automatic filtering and re-authoring so that WAP-enabled cellular phones could read HTML contents. Their basic idea is to extract plain texts in the HTML document by discarding all formatting elements and unnecessary information. The result is then divided into a navigation page and several plain text sub-pages. They also utilize transcoding cache to diminish the run-time overhead. Ardon et al. [1] prototyped a proxy-based web transcoding system based on network access control, user preferences, and displaying capability of equipments. Since all transcoding procedures were finished in the content provider’s server, this server-centric framework avoided potential copyright problems. Lum and Lau [14] built a qualityof-service oriented decision engine for content adaptation. They designed flows for content negotiation and processing for multimedia contents. Hua et al. [9] integrated content adaptation algorithm and content caching strategy for serving dynamic web content in a mobile computing environment. They constructed a testbed to investigate the effectiveness of their design in improving web content readability on small displays, decreasing mobile browsing latency, and reducing wireless bandwidth consumption. Buyukkokten et al. [5] used an “accordion summarization” transcoding strategy where an HTML page could be expanded or shrunk like an accordion. The HTML page is restructured as a tree according to the semantic relationships among its textual sections. All textual sections are split into several Semantic Textual Units, which are automatically summarized. Users can check each summary to expand the node for detailed information. However, this framework only works in the browser they designed for digital libraries. Hwang et al. [11] also treated web page layout as a tree according to its tag hierarchy. They defined a grouping function to transform such tree into sub-trees, and introduced a filtering mechanism to modify the sub-trees for adequate display in the target device. They analyzed specific web page layout structure and re-authored,
232
C. Jou
according to heuristics, web pages for several mobile devices. Each of their transcoding method could handle only specified layout structures of web pages and did not consider mobile device characteristics. Huang and Sundaresan [10] tried the semantics approach in transcoding web pages to improve web accessibility for users. Their system was designed to improve the interface of e-business transactions and to extend interoperable web forms to mobile devices. They used XML/DTD to specify the semantic and grammatical relationship among web contents, so that web forms could achieve consistency, simplicity and adaptability. The advantage of this system is its ability to provide concept-oriented content adaptation, but it is hard to extend. DELI [4], an HP Semantic Lab project, adopted simple negotiation algorithms for rewriting web pages based on context information, like user preference and device capabilities. Glover and Davies [8] used heuristic algorithms to find proper pre-defined web page templates according to device attributes. Their focus is in applying XML/XSLT styles to database contents retrieved in dynamic web pages. In our framework, we follow the guideline of the “Device Independent” [3] principle and utilize the device contexts described in CC/PP.
3 System Architecture In this section, we explain the system architecture and the data flows of our framework in Figure 1. The adaptation mechanism of our framework, called Content Adaptation Proxy Server, resides behind web servers. The Proxy Listener is responsible for receiving HTTP requests (message 1) from miscellaneous mobile devices and for dispatching these requests to the Content Fetcher. The client’s device information as well as user’s personal preferences will be embedded inside these requests through CC/PP diff, which is a modified version of predefined CC/PP profile from the hardware manufacturers. CC/PP is a two-layered user preferences and device capabilities description based on XML/RDF [15]. CC/PP consists of the following three categories: hardware platform, software platform, and browser user agent. Figure 2 shows example detailed attributes in the XML/RDF format of one mobile device [12]. In our framework, all predefined CC/PP profiles are stored within a profile directory. A CC/PP diff is manually configured by the user, and is normally used to reflect the user preferences. It is dynamic and can be further modified in later sessions. Many protocols have been proposed to enhance HTTP 1.1 protocol to include CC/PP profile diff. Two of such protocols are CC/PP-ex [16] and W-HTTP [17]. We adopt CC/PP-ex in this framework. When Proxy Listener accepts a request from a client, it will spawn a working thread in the Web Content Fetcher to handle the request (message 2). Web Content Fetcher performs the standard task of proxy servers. If the requested web content is already in the cache, it will be fetched from Cached Web Content (messages 3.3 and 3.4). If not, then it will be fetched from the source through the internet (messages 3.1 and 3.2), and then be saved in the Cached Web Content. The working thread is also responsible for resolving the CC/PP profile diff.
A Semantics-Based Automatic Web Content Adaptation Framework
233
Content Adaptation Proxy Server
Client 1. HTTP Request
Proxy Listener Internet 3.1 Content Request
2. Dispatched Request
Web Fetcher
Content
3.2 Web content
3.3 Web content request
Cached web content
3.4 Cached web content request 4. Web Content and CC/PP diff
Semantics Extractor
5.1 Device info
Device Info DB
5.2 Semantics properties
5.3 Transcoding parameters
6. Web Content and Transcoding Service Request
Jena Inference System
7. HTTP response
Transcoder
Fig. 1. System Architecture of Content Adaptation Proxy Server
The web content and CC/PP diff will then be sent (message 4) to the Semantics Extractor to acquire the implicit RDF semantic information within the CC/PP diff, HTML web pages and parameters of image files. RDF describes resource information in the internet by the three components: resource, property, and statement. A RDF statement is represented by a triple of (subject, predicate, object). These semantics information will be sent (messages 5.2) to the Jena Inference System as basic facts. Jena Inference Engine will combine these facts with the knowledge base of CC/PP UAProf RDFS model [20], Transcoding Rules and Web Page Auxiliary Vocabulary to determine the proper transcoding parameters in the format of sequential RDF predicates for the requests. The web content and transcoding parameters will then be passed (message 6) to the Transcoder. Besides the layout rewriting mechanisms, the Transcoder is equipped with transcoding toolkits, such as ImageMagic [19] for image resolution adjustment. The results of the Transcoder processing consist of web pages with modified layout as well as images with proper resolution and parameters for the requesting mobile device. They will be returned to the client (message 7) and displayed in its browser.
234
C. Jou
Fig. 2. Example CC/PP for a mobile device
4 Semantics Extractor The Semantics Extractor is responsible of collecting information of the fetched web content by checking its file type, information of analyzing the structure of web pages, and information of constructing facts about the requested web contents and device characteristics. Since most HTML pages are not well-formed, it is hard to extract semantics from them directly. This module first transforms HTML pages into the well-formed XHTML format through the Tidy toolkits [18]. The following two file types of the requested URL will first be handled in this framework: (1) XHTML files: Their metadata are about layouts of the document, possibly with hyperlinks to external textual or binary files. (2) Image files: These are binary files with adjustable parameters, like color depths and resolution. Currently, the system could handle JPEG, PNG, and GIF images. For files encoded in indestructible formats, like Java applets, since they could not be adjusted, the system directly forwards them to the Transcoder for delivery. To extend this web adaptation framework to new file types, we just need to specify the metadata about the new file type, and have to build the semantics extraction component and transcoding rules for web contents of the new file type. For XHTML files, the semantic extractor module collects the following schema information: the identification of each XHTML element, and the layout of the XHTML page. We apply XHTML DOM (Document Object Model) [13] tree nodes scanning to extract node information and relationships among XHTML elements. We solve the element identification problem in an XHTML page by XPath [7] so that each node in the DOM tree could be specified accurately. Statistical or inferred semantics data for the following Web Page Auxiliary Vocabulary will be extracted for each XHTML DOM tree node:
A Semantics-Based Automatic Web Content Adaptation Framework
235
1. NumberOfWords: This data indicates number of words in a paragraph. It is used to determine whether the paragraph corresponding to the XHTML node should be split. 2. NumberOfImages: This data indicates number of the tags in a specific XHTML node. It is used to decide whether a tabular cell is an advertisement banner. 3. AverageLink: This is the quotient of the number of links and the number of words within a XHTML node. In web contents with useful information, this value tends to be very high, and all contents in the node should be preserved. 4. Title: For XHTML nodes with the
or
tag, or with texts surrounded by pairs of the or <STRONG> tags and followed by immediately, the collected content is treated as a title. This could be used as the title of the sub-page corresponding to this node. 5. Layout: This information indicates whether the node is used for layout composition. For example, to determine whether a
is a layout element or an actual tabular cell, we calculate the number of words for the element. If its number of words exceeds a specific threshold, we mark such a
element as a layout element. Consider the following simplified XHTML page:
Gentoo Linux is a totally new linux distribution.
We can describe the
tag in the above page with RDF, XPath and Web Page Auxiliary Vocabulary as follows: 8false0TD0
Semantics of device characteristics will be collected through CC/PP diff. The CC/PP semantics of device configurations are already RDF compatible. They could be sent to the Jena Inference System as facts. The total customized device description can be translated into a graph model within the Jena Inference System.
5 Jena Inference System We use Jena [6], a semantic web toolkit of “Device Independent” [3] ideal, to determine the transcoding parameters, which are represented as sequential RDF
236
C. Jou
Fig. 3. Jena Inference System
predicates. The Jena Inference System, displayed in Figure 3, has three main components. These components are utilized in our framework as follows: Knowledge Base. It contains the acquired knowledge and rules in deciding the content adaptation parameters. XHTML schema is derived by mapping from XHTML XML schema to RDF/RDFS as one knowledge base. Transcoding Rules contain rules using web content ontology and device characteristics ontology for transcoding. Auxiliary Vocabulary for transcoding parameter decision are also described by RDF/RDFS and serialized into Jena knowledge base. All RDF knowledge is serialized in the XML format to provide more flexibility and interoperability in content adaptation. Jena Inference Engine. This is the decision engine to inference and to generate transcoding parameters. We make use of the engine without any modification. Facts. These are facts supported in the form of instantiated predicates. In our framework, they are the semantic data collected by the Semantic Extractor. We apply heuristics to design the transcoding rules in the “IF…THEN…” format. If the precedent parts of a rule are all true, then the consequent part of the rule would be added as a statement of transcoding parameters into the knowledge base. We provide rules regarding device characteristics by defining restriction rules using first order predicate logic. Rules are categorized to back up each other. For example, if rules for HP 6530 PDA are not sufficient, then rules for PPC Pocket PC could be used. In the rules, names prefixed with ‘?’ are variables. The following example rule is for image resizing: [ScaleImageByWidth: (system:Content content:Width ?image_width), (system:HardwarePlatform ccpp:DisplayWidth ?display_width), lessThan( ?display_width, ?image_width ) -> (system:Content content:ScaleImageByWidth ?display_width)]
A Semantics-Based Automatic Web Content Adaptation Framework
237
The rule is named ScaleImageByWidth. The three namespaces “system”, “ccpp”, and “Content” point to the knowledge bases regarding the content adaptation proxy server, the standard UAProf Schema [20], and web content under transcoding, respectively. The above rule means that if the width of the device (?dispaly_width) is less than the width of the image (?image_width), then set the width of the image as the width of the device, and set the height of the image proportionally with respect to the adjustment ratio of the width.
6 Transcoder The Transcoder is composed of several transcoding modules corresponding to file types of XHTML, JPEG, etc. It could be extended to handle other file types. According to the transcoding parameters, it performs content adjustment and filtering. For image files, currently the system not only transforms image files into the same format with different parameters, but also transforms image files into different formats. The Transcoder dispatches the web contents according to their file type to the corresponding adaptation component. To perform the required content transcoding operation, it will query the inferred RDF model by the RDF query language RDQL to obtain the required transcoding parameters. The use of RDQL could prevent tight coupling of the transcoding components. An example RDQL query is demonstrated as follows: SELECT ?predicate, ?object WHERE ( system:Content, ?predicate, ?object) USING system FOR http://www.im.abc.edu/~def/proxy.rdfs#
In RDQL, USING is to specify the name space. This query could obtain all RDF statements with subject system:Content, where system is the name space and Content represents web content currently under transcoding. Transcoding query results are represented as instantiated predicates. For example, if the transcoding predicate for an image file is ScaleImageByWidth, then the Transcoder would adjust the image width and height proportionally. After all RDQL query results are handled, the resulting web content would be returned to the client.
7 System Implementation We implement the content adaptation proxy server in the Fedora Linux 2.6.18 operating system by the Java Language J2SDK 1.5.0. We use the package org.w3c.dom as the class for handling XHTML DOM trees for web pages. The JPEG, PNG, and GIF image files are handled by the Java class java.awt.image.BufferedImage. We make use of the inheritance mechanism, so that the Content Fetcher interface only specifies one method: public Object getContent( ).
The Semantics Extractor is implemented by a Java interface, a factory for semantic extraction, and one class for each supported file type. Some of the transcoding rules for images are listed in Table 1, and some of the transcoding rules for web pages are listed in Table 2.
If the browser in the mobile device does not support Flash, then filter the Flash content.
A Semantics-Based Automatic Web Content Adaptation Framework
239
Note that ?li is a predicate for literal, and noValue is a Jena function to test whether its argument is in the knowledge base. To demonstrate the functionalities of this framework, we tested three client mobile devices: HP iPAQ hx2400, Symbian S80 Simulator, and Panasonic EB-X700. We would like to show the effect of the following two CC/PP parameters: supported file types and display size. The goal of the adaptation is to avoid the use of horizontal scroll bar, so as to increase the readability of transcoded pages and images. The related specifications and restrictions of these devices are listed in Table 3. Table 3. Specifications and restrictions of mobile devices
Device
HP iPAQ hx2400
Symbian S80
Panasonic EB-X700
Category
PDA
Smart Phone
Smart Phone
Operating System
Windows Mobile 5.0
Symbian Series80
Symbian Series60
IE Mobile
Built in
Built in
text/html text/css image/jpeg image/png image/gif 480 x 320 (pixels) Bluetooth
text/xhtml text/css image/jpeg image/png image/gif 220 x 640 (pixels) WLAN
text/chtml image/jpeg
Browser Supported file types
Display size Connection
176 x 148 (pixels) GPRS
Figure 4 shows the upper part of the tested web page (http://www.amazon.com) in a Microsoft IE 6.0 browser in a desktop computer. The upper parts of the transcoded pages in the built-in browser for the three tested mobile devices are displayed by two screen shots in Figures 5 to 7. All resulting transcoded web pages satisfy the goal of avoiding the use of horizontal scroll bar by adjusting the page layout, image size, and image resolution. Unsupported CSS, Javascripts, flashes, div’s and tables are filtered out.
Fig. 4. Test web page in a desktop computer
240
C. Jou
Fig. 5. Results of test page in HP iPAQ hx2400
Fig. 6. Results of test page in Symbian Series80
Fig. 7. Results of test page in Panasonic EB-X700
A Semantics-Based Automatic Web Content Adaptation Framework
241
8 Conclusions We designed and implemented a flexible and robust infrastructure for web content adaptation using RDF semantics from CC/PP device characteristics, XHTML web pages, and JPEG, PNG, and GIF image files. Past researches in this area either did not take device characteristics into consideration, or were not a general purpose solution for miscellaneous mobile devices. We made use of the Jena Inference System to obtain the transcoding parameters through the fact and knowledge base built from the collected semantics. In this framework, a single copy of the web pages could serve many different mobile devices. Previous tedious web page rewriting labour for mobile devices could be saved. Our framework could be easily extended to new file types by importing related semantics and transcoding modules. We plan to incorporate support of style sheet and web form specifications, such as CSS and XForm, into this semantics-based content adaptation framework. By supporting these dynamic web pages, the increased user interactivity would accelerate user acceptance of pervasive computing. Acknowledgements. This work was supported in part by Taiwan’s National Science Council under Grant NSC 95-2221-E-032 -027.
References 1. Ardon, S., Gunningberg, P., Landfeldt, B., Ismailov, Y., Portmann, M., Seneviratne, A.: March: a distributed content adaptation architecture. International Journal of Communication Systems 16(1), 97–115 (2003) 2. Bickmore, T., Girgensohn, A.: Web page filtering and re-authoring for mobile users. The Computer Journal 42(6), 534–546 (1999) 3. Bickmore, T., Girgensohn, A., Sullivan, J.W.: Digestor: device-independent access to the World Wide Web. In: Proceeding of the 6th World Wide Web Conference WWW6, pp. 665–663 (1997) 4. Butler, M.: DELI: A Delivery context library for CC/PP and UAProf, External technical report, HP Semantic Lab (2002) 5. Buyukkokten, O., Garcia-Molina, H., Paepcke, A.: Accordion summarization for end-game browsing on PDAs and cellular phones. In: Proceeding Of the SIGCHI Conference on Human Factors in Computing Systems, pp. 213–220 (2001) 6. Carroll, J., Dickinson, I., Dollin, C., Reynolds, D., Seaborne, A., Wilkinson, K.: Jena: implementing the semantic web recommendations. In: International 13th World Wide Web Conference, pp. 74–83 (2004) 7. Clark, J., DeRose, S.: XML Path Language (XPath) v. 1.0, W3C (1999), http:// www.w3.org/TR/xpath 8. Glover, T., Davies, J.: Integrating device independence and user profiles on the web. BT Technology Journal 23(3), 239–248 (2005) 9. Hua, Z., Xie, X., Liu, H., Lu, H., Ma, W.: Design and performance studies of an adaptive scheme for serving dynamic web content in a mobile computing environment. IEEE Transactions on Mobile Computing 5(12), 1650–1662 (2006)
242
C. Jou
10. Huang, A.W., Sundaresan, N.: A semantic transcoding system to adapt web services for users with disabilities. In: Proceeding of the 4th international ACM conference on Assistive technologies, pp. 156–163 (2000) 11. Hwang, Y., Kim, J., Seo, E.: Structure-aware web transcoding for mobile devices. IEEE Internet Computing 7(5), 14–21 (2003) 12. Klyne, G., Reynolds, F., Woodrow, C., Ohto, H., Hjelm, J., Butler, M., Tran, L. (eds.): Composite Capability/Preference Profiles (CC/PP), W3C (2004), http://www.w3. org/TR/CCPP-struct-vocab 13. Le Hégaret, P., Whitmer, R., Wood, L. (eds.): Document Object Model (DOM), W3C (2005), http://www.w3.org/DOM/ 14. Lum, W.Y., Lau, F.C.M.: A context-aware decision engine for content adaptation. IEEE Pervasive Computing 1(3), 41–49 (2002) 15. Manola, F., Miller, E. (eds.): RDF Primer, W3C (2004), http://www.w3.org/TR/ rdf-primer/ 16. Ohto, H., Hjelm, J.: CC/PP exchange protocol based on HTTP Extension Framework, W3C (1999), http://www.w3.org/TR/NOTE-CCPPexchange 17. OMA: Wireless Profiled HTTP, WAP-229-HTTP-20011031-a of Open Mobile Alliance Ltd., (2001) 18. Raggett, D.: Clean up your web pages with HTML TIDY, W3C (2000), http:// tidy.sourceforge.net/docs/Overview.html 19. Still, M.: The Definitive Guide to ImageMagick. Apress (2005) 20. WAP: WAG UAProf, WAP-248-UAPROF-20011020-a of Wireless Application Protocol Forum Ltd., (2001)
Part III
Society, e-Business and e-Government
Summarizing Online Customer Reviews Automatically Based on Topical Structure Jiaming Zhan1, Han Tong Loh1, and Ying Liu2 1
Department of Mechanical Engineering, National University of Singapore, Singapore {jiaming,mpelht}@nus.edu.sg 2 Department of Industrial and Systems Engineering, The Hong Kong Polytechnic University Kowloon, Hong Kong S.A.R., China [email protected]
Abstract. Online customer reviews offer valuable information for merchants and potential shoppers in e-Commerce and e-Business. However, even for a single product, the number of reviews often amounts to hundreds or thousands. Thus, summarization of multiple customer reviews is helpful to extract the important issues that merchants and customers are concerned with. Existing methods of multi-document summarization divide documents into non-overlapping clusters first and then summarize each cluster of documents individually with the assumption that each cluster discusses a single topic. When applied to summarize customer reviews, it is however difficult to determine the number of clusters a priori without the domain knowledge, and moreover, topics often overlap with each other in a collection of customer reviews. This paper proposes a summarization approach based on the topical structure of multiple customer reviews. Instead of clustering and summarization, our approach extracts topics from a collection of reviews and further ranks the topics based on their frequency. The summary is then generated according to the ranked topics. The evaluation results showed that our approach outperformed the baseline summarization systems, i.e. Copernic summarizer and clustering-summarization, in terms of users’ responsiveness. Keywords: Customer reviews, multi-document summarization, topical structure, web mining.
and it is time-consuming to read through all of them manually. For example, there are hundreds of reviews posted on the web for some popular products in Amazon.com; and thousands of customer emails may be received by the manufacturer regarding one particular product. Some work has been reported dealing with the vast amount of customer reviews [1-3]. All these work focused on opinion mining which was to discover the reviewers’ orientations, whether positive or negative, regarding various features of a product, e.g. the weight of a laptop and the picture quality of a digital camera. However, we noticed that although some comments regarding product features could not be labeled as positive or negative, they were still valuable. For example, the following two sentences are extracted from the customer reviews of mobile phone Nokia 6610 from Hu’s corpus [1]. #1: The phone’s sound quality is great. #2: The most important thing for me is sound quality. Both sentences discuss the product feature sound quality. Unlike the first sentence, the second one does not offer any orientation, either positive or negative, regarding the specific feature of phone Nokia 6610, yet it does provide valuable information for designers and manufacturers about what mobile phone consumers are concerned with. Such neutral comments are currently not considered in opinion mining. Moreover, opinion mining focuses mainly on product features which can not cover all significant issues in customer reviews. Fig. 1 shows some sentences extracted from the customer reviews of Nokia 6610, a flip phone. They have revealed the different perspectives from customers about flip phone. Some customers also elaborate the reasons for their choices. This information is trusted to be valuable for designers and manufacturers. However, using the current method of opinion mining, such important issues would not be pointed out because flip phone is not an explicit product feature of Nokia 6610. • • • • •
As much as I like Nokia phones the flip phones are much better because a) you won’t scratch your screens/keys b) you don’t need to lock your phone all the time to prevent accidentally hitting the keys. Personally I like the Samsung phones better because I found myself liking the flip phones so much more. My past two phones were all flip phones, and I was beginning to tire of them. Nokia was my first non-flip phone, and I'm glad I decided to go with them. This is probably your best bet if you are looking for a phone in this price range, or like me, do not have the patience to deal with annoying flip phones. Fig. 1. Sentences discussing flip phone from customer reviews of Nokia 6610
In this paper, we propose an approach to automatically summarize multiple customer reviews which are related to each other, e.g. reviews discussing the same product or the same brand. In our approach, we intend to discover salient topics among reviews first and then to generate a summary based on these topics. Unlike existing Multi-Document Summarization (MDS) approaches which divide documents into non-overlapping groups and summarize each group of documents individually,
Summarizing Online Customer Reviews Automatically Based on Topical Structure
247
our approach is based on the topical structure of a document collection. The rest of this paper is organized as follows: related work of automatic text summarization is reviewed in Section 2; our summarization approach is presented in Section 3; Section 4 reports the evaluation of summarization results and Section 5 concludes.
2 Related Work of Automatic Text Summarization During the last decade, there has been much interest on automatic text summarization due to the explosive growth of online electronic documents [4-7]. As a result, there is an increasing demand for MDS. Instead of focusing only on a single document, MDS is performed to deal with multiple related documents [8, 9], e.g. news articles regarding an event from various sources. The most popular MDS approach is clustering-summarization [10, 11]. The approach of clustering-summarization first divides a collection of documents into several non-overlapping groups of documents or sentences. Summarization is then performed separately within each group. Two limitations are noted when clustering-summarization approach is applied to the problem of customer reviews: • •
The number of clusters is difficult to be determined without the domain knowledge regarding the collection of reviews. Inappropriately choosing this number will inevitably introduce noisy information and reduce effectiveness. In clustering-summarization, the document set is split into non-overlapping clusters and each cluster is assumed to discuss one topic. However, in a realworld review collection, topics often overlap with each other and are not perfectly distributed in the non-overlapping clusters of documents. Each topic is associated with a number of reviews. Likewise, each review in the collection may discuss several topics, instead of only one, because customers would rather like to comment on various aspects of a product.
In order to tackle the limitations of clustering-summarization method, we propose a summarization approach based on the topical structure of these reviews.
3 Summarization Based on Topical Structure As discussed, the existing MDS approach of clustering-summarization is weak to handle the structure within a real-world document set, i.e. topics are not perfectly distributed in non-overlapping clusters of documents. This situation is more acute in the domain of online customer reviews, since review writers usually write in an arbitrary style and tend to cover different topics in a review. As shown in Fig. 2, it lists some topics in the review collection of Nokia 6610 and review IDs with respect to these topics, review 18 has comments regarding all the topics and some other reviews are also associated with multiple topics. The approach of clusteringsummarization is not suitable in this situation since clustering this collection into nonoverlapping groups will "cut off" the relationship among reviews.
248
J. Zhan, H.T. Loh, and Y. Liu • • • • •
•
Sound quality 8,13,18,20,27,33,34,40 Battery life 2,5,10,13,17,18,26,28,29,30,37 Flip phone 4,18,26,33 Nokia phone 1,2,16,17,18,31,37 Samsung phone 18,40 …
Fig. 2. Some topics from the review collection of Nokia 6610
We propose a summarization approach based on the topical structure demonstrated in Fig. 2. The framework of our approach is shown in Fig. 3. Detailed steps are given as follows.
Intranet and W WW
Extracting FSs and equivalence classes
A collection of customer reviews
Pre-processing: stop words removal, stem ming
Topic identification Ranking FSs and equivalence classes
FSs as topics
Highlighting relevant sentences for each topic
Redundancy reduction in candidate sentences
Equivalence classes as topics
Candidate sentence extraction
Post-processing
Multi-docum ent summary based on topical structure
Fig. 3. Summarization process based on topical structure
3.1 Pre-processing The summarization process starts with a collection of customer reviews as the input. These reviews are collected from WWW or retrieved from Intranet, e.g. all customer emails regarding a product. Pre-processing steps are first applied to the reviews, including stop words removal [12] and term stemming [13]. Stop words are those words which rarely contribute useful information in terms of document relevance, such as a, the and of. Word stemming is the process of reducing inflected or derived words to their stem, base or root form. For example, a stemming algorithm for English should stem the words fishing, fished, fish and fisher to the root word, fish. The purpose of these steps is to reduce the noisy information in the following processes.
Summarizing Online Customer Reviews Automatically Based on Topical Structure
249
3.2 Topic Identification The key step of our framework is to identify topics in the review collection and generate the topical structure based on these topics. Some work of topic identification has been reported in previous literature. The typical method is text segmentation, which is to segment the text by similarity of adjacent passages and detect the boundary of topics [14, 15]. This method works well for single text. For multiple texts, however, it is hard to find such straightforward boundaries. Our process of topic identification is based on Frequent word Sequences (FSs) [16] and equivalence classes [17]. A FS is a sequence of words that appears in at least σ documents in a document collection (σ is the threshold for supporting documents). Algorithm 1 demonstrates the process to extract all the FSs in a document collection. The process starts with collecting all the frequent word pairs, i.e. FSs with length two. These FSs are then expanded with one more word and therefore form a set of word sequences with length three. All the FSs with length three are then expanded. This process is iteratively performed until there is no FS left for expansion. The threshold for supporting documents is chosen according to the size of the review collection. For a small collection, say 20 reviews, a low threshold is chosen to let more important concepts to surface. For a large collection, a high threshold may be considered to reduce noisy information. Algorithm 1. Discovery of all FSs in a review collection
D: a set of pre-processed reviews σ: frequency threshold Fs: a set of FSs //Output: //Initial phase: collecting all frequent pairs 1 For all the reviews d ∈D 2 Collect all the ordered pairs and occurrence information in d 3 Seq2 = all the ordered word pairs that are frequent in D //Discovery phase: building longer FSs k := 2 4 5 Fs := Seq2 6 While Seqk ≠ Φ 7 For all phrases s ∈ Seqk 8 Let l be the length of the sequence s 9 Find all the sequences s’ such that s is a subsequence of s’ and the length of s’ is l+1 10 For all s’ 11 If s’ is frequent S := S ∪ {s’} 12 13 Fs := Fs ∪ S Seqk+1 := Seqk+1 ∪ S 14 k := k+1 15 16 Return Fs //Input:
FSs can be further pruned and grouped into equivalence classes according to their cooccurrences with each other. The equivalence classes are generated in the following way. Let A and B be two FSs. The equivalence class of A, EqA, contains the set of FSs that cooccur with A in almost the same set of reviews, as given by a confidence
250
J. Zhan, H.T. Loh, and Y. Liu
parameter. DetA is the set of FSs that are determined by A, and is required in deciding which FSs belong in EqA. For A and B, if: frequency ( A, B cooccur ) ≥ confidence frequency ( A)
(1)
we add B to the set DetA; A itself is also included in DetA. Other FSs are tested in the same manner, and will be added to DetA if they satisfy the above criterion. EqA is thus made up of all FSs X such that DetX=DetA. A FS or an equivalence class is considered as the representative of one topic in a review collection. In the following experiments, we intend to compare the performance between FSs and equivalence classes as topics. Topics are ranked based on their scores. The score of a FS is calculated in the form of Equation 2. The score of an equivalence class equals to the average scores of its FSs. score = f ⋅ log 2
N +1 ⋅ log 2 (l + 2) n
(2)
where f is the frequency of the FS in the whole review collection, N is the total number of reviews, n is the number of reviews in which the FS occurs, l is the length of the FS. 3.3 Candidate Sentence Extraction For each topic in a collection, all relevant sentences are extracted and added into a pool as candidate segments of final summary until the expected summary length is reached. Each sentence will be accompanied by a label including its source review ID. The method of Maximal Marginal Relevance (MMR) is implemented to reduce the redundancy in the sentence selection process [18]. MMR intends to balance the tradeoff between the centrality of a sentence with respect to the topic (the first part in Equation 3) and its novelty compared to the sentences already selected in the summary (the second part in Equation 3), i.e. to maximize the marginal relevance in the following form:
MR (si ) = λSim (si , D ) − (1 − λ ) max Sim (si , s j ) s j ∈S
(3)
where si is a candidate sentence, D is the set of relevant sentences to a particular topic, S is the set of sentences already included in the summary, λ is the redundancy parameter ranging from 0 to 1. With regard to Sim, we adopt a cosine similarity measure between sentence vectors. Each element of a sentence vector represents the weight of a word-stem in a document after removing stop words. 3.4 Post-processing and Final Presentation The final step is to regenerate sentences from the candidate sentences and present the summary output to users. Fig. 4 shows an example of the summary presented to readers. Topics are ranked according to their saliency in the review collection. Reviews relevant to each topic have been identified and hyperlinked, with their IDs included in the parenthesis
Summarizing Online Customer Reviews Automatically Based on Topical Structure
251
following the topical phrase, to make it easy for users to browse the details of each review article. If users are interested in a particular topic, they can click the unfolding button preceding to the topical phrase to expand this topic and the detailed information will then be presented. In Fig. 4, the topic flip phone is unfolded and all the relevant sentences to this topic are displayed along with reviews’ IDs.
Fig. 4. Summarization output for the review collection of Nokia 6610
4 Evaluation of Summarization Results We compared our summarization approach with the baseline summarization systems of Copernic summarizer [19] and clustering-summarization. Copernic summarizer is a commercial summarization software using undisclosed statistical and linguistic algorithms. The method of clustering-summarization is a popular method for MDS, especially in the context of information retrieval system [10]. In clusteringsummarization, a document collection is separated into non-overlapping clusters and summarization is then performed in each cluster. 4.1 Experimental Data Sets and Parameter Setting The data sets used in our experiments included five sets from Hu’s corpus [1] and three sets from Amazon.com. For each review collection, summaries were generated using Copernic summarizer, clustering-summarization method and our approach based on topical structure. These document sets were normal-sized with 40 to 100 documents per set. Therefore, we extracted FSs with at least three supporting documents in our approach. The confidence level for equivalence classes was set to
252
J. Zhan, H.T. Loh, and Y. Liu
0.9 and redundancy parameter λ in candidate sentence selection was set to 0.5. Since the document sets in our experiments were normal-sized, the clustering number in clustering-summarization method was set to five. The clustering algorithm was implemented in Cluto [20]. The compression ratio of summarization was set to 10%, i.e. the length ratio of summary to original text was 10%. The summary generated by Copernic was a set of ranked sentences. The summary generated by clustering-summarization was divided into clusters, as shown in Fig. 5 (only three clusters are shown here). Cluster 1 (4 reviews) Sound - excellent polyphonic ringing tones are very nice (check cons) it also doubles as a radio, which is a nice feature when you are bored. Cons: ring tones only come with crazy songs and annoying rings, there is only one ring that sounds close to a regular ring. Games kind of stink and you cant download them you have to get the link cable to get additional games. … Cluster 2 (3 reviews) Nice and small and excellent when it comes to downloading games, graphics and ringtones from www.crazycellphone.com I thought this was the ultimate phone when it comes to basic features, but I was dissapointed when I saw that it was only a gsm comaptible phone. … Cluster 3 (17 reviews) I've had an assortment of cell phones over the years (motorola, sony ericsson, nokia etc.) and in my opinion, nokia has the best menus and promps hands down. No other color phone has the combination of features that the 6610 offers. From the speakerphone that can be used up to 15 feet away with clarity, to the downloadable poly-graphic megatones that adds a personal touch to this nifty phone. ... Fig. 5. Summary generated by the method of clustering-summarization for the review collection of Nokia 6610 (only three clusters are shown here)
4.2 Intrinsic Evaluation and Extrinsic Evaluation The methods of summarization evaluation can be classified into intrinsic method and extrinsic method. Intrinsic method compares candidate summaries with reference summaries [21]. Reference summaries are usually generated manually and are therefore biased by human authors. Extrinsic method requires no reference summary and is task-oriented or user-oriented [22]. In our case, since it is hard to define an ideal reference summary to fulfill the diverse information requirements of different users, extrinsic evaluation is more suitable. We evaluated summarization performance according to users’ responsiveness. Human assessors were required to give a score for each summary based on its structure
Summarizing Online Customer Reviews Automatically Based on Topical Structure
253
and coverage of important topics in the review collection. The score was an integer between 1 and 5, with 1 being least responsive and 5 being most responsive. In order to reduce bias in the evaluation, three human assessors from different background joined the scoring process. For one collection, all the peer summaries were evaluated by the same human assessor so that the hypothesis testing (paired t-test) could be performed to compare the peer summaries. 4.3 Evaluation Results Table 1 shows the average responsiveness scores of Copernic summarizer, clusteringsummarization method and our approach based on all the review collections. Table 2 presents the results of paired t-test between our approach (using FSs as topics) and other methods. Table 1. Average responsiveness scores
Table 2. Hypothesis testing (paired t-test) Null hypothesis (H0): There is no difference between the two methods. Alternative hypothesis (H1): The first method outperforms the second one. P-value Frequent word Sequences (FSs) vs. Copernic summarizer 2.26×10-5 Frequent word Sequences (FSs) vs. Clustering-summarization 2.43×10-4 Frequent word Sequences (FSs) vs. Equivalence classes 7.68×10-4
It can be found that the approach based on topical structure performed the best amongst all the peer methods (Table 1 & 2), because this approach better represents the internal structure of a review collection than clustering-summarization. We also analyzed the clustering quality in the clustering-summarization method. Table 3 shows the intra-cluster similarity and inter-cluster similarity for the review collection Nokia 6610. As can be seen, there was not much difference between intra-cluster similarity and inter-cluster similarity, especially for cluster 4 and 5 which were the two major clusters in the collection. This implies that the review collections are difficult to be clustered into non-overlapping clusters. As shown in Table 1 & 2, we found that using FSs as topics was significantly better than equivalence classes with the p-value of 0.0008 in paired t-test. Review writers usually write in an arbitrary style and cover different topics in a review rather than focus on only one topic. Therefore, using equivalence classes might introduce much noisy information, since equivalence classes are grouping topics based on their cooccurrences.
254
J. Zhan, H.T. Loh, and Y. Liu
Table 3. Intra-cluster similarity and inter-cluster similarity of the review collection Nokia 6610 (41 reviews, 5 clusters) Cluster ID 1 2 3 4 5
Copernic summarizer performed worse than other summarization methods. The possible reason is that Copernic summarizer does not take into account the situation of MDS and treats all sentences from a review collection as the same in the pool of candidate segments for summarization. In the aforementioned experiments, for each document set, our approach generated the summary based on the top 10 significant topics. The number of topics in the summary would probably affect the summarization performance, which is similar to the concept of compression ratio in summarization [23]. Too short summaries discard a lot of useful information, while too long summaries cost more reading time. The summarization system should find an optimal summary length so that important information is kept and the reading time is reduced to minimal. In our case, it is necessary to find an optimal number of topics in the summary. Therefore, we varied the number of topics to three levels: 5, 10 and 20, and investigated their effects on summarization performance. The experimental result is given in Table 4. Table 4. Hypothesis testing (paired t-test) Null hypothesis (H0): There is no difference between the two methods. Alternative hypothesis (H1): The first method outperforms the second one. P-value Our approach 10 topics vs. 5 topics 1.06×10-4 Our approach 20 topics vs. 10 topics 0.175
As can be found, increasing the number of topics from 5 to 10 could significantly improve the summarization performance, while there was no significant difference between 10 and 20 topics. The experimental results may suggest that for the moderate-sized document sets in our experiments, the top 10 topics are enough to cover the most important information and enough to satisfy the information need for most readers.
5 Conclusions Summarization of mutiple online customer reviews is a process to transfer reviews from unstructured free texts into a structured or semi-structured summary that reveals the commonalities and links among reviews. The automation of this process, in the
Summarizing Online Customer Reviews Automatically Based on Topical Structure
255
context of e-Commerce and e-Business, should be able to assist potential consumers in seeking information and to facilitate knowledge management in enterprises as well. We proposed an approach to automatically summarize multiple customer reviews based on their topical structure. As we have observed that topics often overlap with each other in a collection of reviews, we extracted topics across different reviews, instead of dividing reviews into several non-overlapping clusters and summarizing. Evaluation results demonstrated that our approach achieved better summarization performance and users’ satisfaction when compared to the baseline systems, i.e. Copernic summarizer and clustering-summarization method. Moreover, this approach is able to address different concerns from potential consumers, distributors and manufacturers. Potential consumers usually concentrate on the positive or negative comments given by other consumers. Designers and manufacturers, on the other hand, may be more concerned about the overall important issues and why customers favor or criticize their products. The emergence of Blogs and e-Opinion portals has offered customers a novel platform to exchange their experiences, comments and recommendations. Reviews for a particular product may be obtained from various sources in different writing styles. The integration and processing of such information from different sources will be the focus in our future work.
References 1. Hu, M., Liu, B.: Mining and Summarizing Customer Reviews. In: 10th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Seattle, WA, pp. 168–177 (2004) 2. Popescu, A.-M., Etzioni, O.: Extracting Product Features and Opinions from Reviews. In: Joint Conference on Human Language Technology / Empirical Methods in Natural Language Processing (HLT/EMNLP 2005), Vancouver, Canada, pp. 339–346 (2005) 3. Turney, P.D.: Thumbs up or Thumbs down? Semantic Orientation Applied to Unsupervised Classification of Reviews. In: 40th Annual Meeting on Association for Computational Linguistics, Philadelphia, PA, pp. 417–424 (2001) 4. Barzilay, R., Elhadad, M.: Using Lexical Chains for Text Summarization. In: ACL 1997/EACL 1997 Workshop on Intelligent Scalable Text Summarization, Madrid, Spain, pp. 10–17 (1997) 5. Gong, Y., Liu, X.: Generic Text Summarization using Relevance Measure and Latent Semantic Analysis. In: 24th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, New Orleans, LA, pp. 19–25 (2001) 6. Hovy, E., Lin, C.-Y.: Automated Text Summarization in SUMMARIST. In: ACL 1997/EACL 1997 Workshop on Intelligent Scalable Text Summarization, Madrid, Spain, pp. 18–24 (1997) 7. Yeh, J.-Y., Ke, H.-R., Yang, W.-P., Meng, I.-H.: Text Summarization using a Trainable Summarizer and Latent Semantic Analysis. Information Processing & Management 41(1), 75–95 (2005) 8. Mani, I., Bloedorn, E.: Summarizing Similarities and Differences among Related Documents. Information Retrieval 1(1-2), 35–67 (1999)
256
J. Zhan, H.T. Loh, and Y. Liu
9. McKeown, K., Radev, D.R.: Generating Summaries of Multiple News Articles. In: 18th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, Seattle, WA, pp. 74–82 (1995) 10. Maña-López, M.J.: Multidocument Summarization: An Added Value to Clustering in Interactive Retrieval. ACM Transaction on Information Systems 22(2), 215–241 (2004) 11. Radev, D.R., Jing, H., Styś, M., Tam, D.: Centroid-Based Summarization of Multiple Documents. Information Processing & Management 40(6), 919–938 (2004) 12. Van Rijsbergen, C.: Information Retrieval. Butter Worths (1979) 13. Porter, M.F.: An Algorithm for Suffix Stripping. Program 14(3), 130–137 (1980) 14. Choi, F.Y.Y.: Advances in Domain Independent Linear Text Segmentation. In: 1st North American Chapter of the Association for Computational Linguistics, Seattle, WA, pp. 26– 33 (2000) 15. Hearst, M.A.: TextTiling: Segmenting Text into Multi-paragraph Subtopic Passages. Computational Linguistics 23(1), 33–64 (1997) 16. Liu, Y.: A Concept-Based Text Classification System for Manufacturing Information Retrieval. Ph.D. Thesis, National University of Singapore (2005) 17. Ahonen, H.: Finding All Maximal Frequent Sequences in Text. In: ICML 1999 Workshop on Machine Learning in Text Data Analysis, Bled, Slovenia (1999) 18. Carbonell, J., Goldstein, J.: The Use of MMR, Diversity-Based Reranking for Reordering Documents and Producing Summaries. In: 21st Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, Melbourne, Australia, pp. 335–336 (1998) 19. Copernic summarizer, http://www.copernic.com 20. Karypis, G.: Cluto: A Software Package for Clustering High Dimensional Datasets, Release 1.5. Department of Computer Science, University of Minnesota (2002) 21. Jing, H., Barzilay, R., McKeown, K., Elhadad, M.: Summarization Evaluation Methods: Experiments and Analysis. In: AAAI 1998 Workshop on Intelligent Text Summarization, Stanford, CA, pp. 60–68 (1998) 22. Tombros, A., Sanderson, M.: Advantages of Query Biased Summaries in Information Retrieval. In: 21st Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, Melbourne, Australia, pp. 2–10 (1998) 23. Mani, I., Klein, G., House, D., Hirschman, L., Firmin, T., Sundheim, B.: SUMMAC: A Text Summarization Evaluation. Natural Language Engineering 8, 43–68 (2002)
The Structure of Web-Based Information Systems Satisfaction: An Application of Confirmatory Factor Analysis Christy M.K. Cheung1 and Matthew K.O. Lee2 1
Department of Finance & Decision Sciences, Hong Kong Baptist University Kowloon Tong, Kowloon, Hong Kong [email protected] 2 Department of Information Systems, City University of Hong Kong Kowloon Tong, Kowloon, Hong Kong [email protected]
Abstract. User satisfaction has become one of the most important measures of the success or effectiveness of information systems; however, very little is known about its structure and dimensionality in the web environment. In the current study, five latent variable models were empirically tested and compared to describe the relationships among 21 observable variables concerned with web-based information systems satisfaction. Using data from a sample of 515 university students, a second-order model was retained based on statistical and theoretical criteria. At the first-order level were six highly interrelated primary factors; understandability, reliability, usefulness, access, usability, and navigation. These six factors were explained by two interrelated second-order factors of web information satisfaction and web system satisfaction. Overall, the model provides a good-fit to the data and is theoretically valid, reflecting logical consistency. Implications of the current investigation for practice and research are provided. Keywords: User Satisfaction, Web-based Information System, Information Quality, System Quality.
system responsive time is critical to WIS satisfaction. As a result, the user satisfaction measurement instrument for traditional IS may fail to fully capture user satisfaction in the Internet-based environment (e.g., [1]). A typical research cycle for developing a standardized measurement instrument to measure a new construct involves two stages of investigation [24]. In the first stage, exploratory studies are performed with the aim to develop hypothesized measurement model(s) via the analysis of empirical data from a referent population. A common technique used in this stage is exploratory factor analysis. In the following stage of investigation, confirmatory studies (e.g., confirmatory factor analysis through a LISREL framework) are performed against new data from the same referent population in order to test the hypothesized measurement model(s) resulting from the previous stage of investigation. An examination of the existing studies on web-based information systems satisfaction revealed that none of these studies have gone beyond the stage of exploratory studies identifying only first order factors. In fact, there are many plausible competing measurement models but very little understanding exists on the relative veracity of these models. Confirmatory studies are thus needed to complete the research cycle. Researchers in the social sciences have long recognized the need for validating theoretical hypotheses at the construct level. Their higher-order factor model are used not only to validate theoretical hypotheses at the construct level, but also to alleviate the multicolinearity that may be presented among the observed variables and the firstorder unobservable constructs [30, 33]. Research on End-User Computing (EUC) satisfaction has already distinguished the different levels of abstraction at the construct level. Doll et al. [11] performed a confirmatory factor analysis of the EUC satisfaction instrument and tested the alternative factor structures of EUC satisfaction. Their results not only provide strong support for their EUC satisfaction instrument, but also demonstrated the second-order factor structure of EUC satisfaction; however, despite the apparent richness of higher-order factor structure, very little IS research has been done on second (or higher) order factor solutions. The purpose of this study is thus to investigate of the web-based information systems satisfaction construct, especially its structure and dimensionality by developing and testing several competing theoretical measurement models. The results are anticipated to increase our understanding of WIS satisfaction and help to enrich and validate WIS satisfaction measurement instrument, which may serve as a practical evaluation tool for evaluating web-based information systems.
2 Theoretical Background End-User Computing satisfaction is one of the most important measures of information systems success [8, 9, 29, 34]. EUC satisfaction is generally defined as an overall affective evaluation an end-user has regarding his or her experience related to a specific information system. Several models for measuring user satisfaction have been developed. For example, Bailey and Pearson [2] developed a tool for measuring and analyzing computer user satisfaction formed by 39 items. Expanded upon the foundation built by Bailey and Pearson [2], Doll and Torkzadeh [10] developed a second-order factor model of EUC satisfaction that consists of five first order factors
The Structure of Web-Based Information Systems Satisfaction
259
(i.e., content, accuracy, format, ease of use, and timeliness) measured by 12 items using a five point Likert scale. This instrument was one of the best known and frequently employed measurements of end-user computing satisfaction (e.g., [25]). In 1994, Doll and his colleagues [11] revisited their model and tested five alternative configurations for EUC satisfaction through a confirmatory factor analysis of the EUC satisfaction instrument. Their results provide strong support for their EUC satisfaction instrument. Otto et al. [28] pioneered in conceptualizing and validating the construct of web user satisfaction, which has five elements, including content, format, graphics, ease of use and responsiveness. Cho and Park [7] also attempted to develop their own sets of instrument for measuring user satisfaction in the web environment. Muylle et al. [27] defined and developed the construct of web site user satisfaction, including information (relevance, accuracy, comprehensibility, and comprehensiveness), connection (ease-ofuse, entry guidance, structure, hyperlink connotation, and speed), and layout. These studies focus on web user satisfaction, seeing the Internet merely a communication channel rather than an integrated platform for hosting web applications. The richness of information and the nature of unstructured and highly individually customizable interactions typically exhibited by WIS redefine the standard of user satisfaction in the web environment. Therefore, web satisfaction should be analyzed at information level and system level [26], captured by the dimensions web information quality satisfaction (Web-IQ satisfaction) and web system quality satisfaction (Web-SQ satisfaction) respectively.
3 Factor Structure for WIS Satisfaction McKinney et al. [26] conceptualized web-based information system satisfaction as a multidimensional concept that was made up of Web-IQ satisfaction and Web-SQ satisfaction, each of which, in turn, was comprised of three dimensions respectively (see Table 1). Table 1. Dimensions of web information satisfaction Dimensions Understandability
Manifest Variables Clear in meaning Concerned with such issues as clearness Easy to understanding and goodness of the information Easy to read Trustworthy Concerned with the degree of accuracy, dependability, and consistency of the Accurate information Credible Users’ assessment of the likelihood that Informative the information will enhance their decision Valuable Refers to the speed of access and Responsive availability of the web site at all times Quick loads Simple layout Concerned with the extent to which the web site is visually appealing, consistent, Easy to use fun and easy to use Well organized Easy to go back and forth Evaluates the links to needed information A few clicks
260
C.M.K. Cheung and M.K.O. Lee
This hierarchy implies that users evaluate WIS performance at multiple levels with multiple dimensions, and ultimately combine these evaluations to arrive at an overall WIS satisfaction perception. Negative performance on each of these dimensions of a WIS is found to have a more significant impact on satisfaction than positive performance [6]. WIS satisfaction is therefore the overall affective evaluation a user has regarding his or her experience related to the web-based information system. Figure 1 shows the third-order factor model of WIS satisfaction.
Fig. 1. McKinney et al.’s [26] web satisfaction model
4 Competing Models for WIS Satisfaction In this study, we followed the approach proposed by Doll et al. [11] to test the five alternative factor structures of WIS satisfaction with 21 observable reflective items. Models 1 to 3 represent the non-hierarchical structure with only first-order factor, and Model 4 and 5 represent the hierarchical structure with more than one level of abstraction. Model 1 is a first-order factor model (see Figure 2a). One factor (WIS satisfaction) is hypothesized to account for all the common variance among the 21 reflective items. This is consistent with the idea used in the end user computing satisfaction literature, adding the item scores to obtain a total satisfaction score. Model 2 hypothesizes that there are six orthogonal or uncorrelated first-order factors (i.e., understandability, reliability, usefulness, access, usability, and navigation) in the model (see Figure 2b). McKinney et al. [26] performed an exploratory factor analysis resulting in six factors. Thus, Model 2 is considered a plausible alternative model of underlying data structure. Model 3 is a first-order factor model with the six factors correlated with each others to represent different dimensions of the concept of WIS satisfaction (see Figure 2c). Assuming the six factors are correlated allows us to capture the common variance in the model.
The Structure of Web-Based Information Systems Satisfaction
(a)
(b) Fig. 2. (a) Model 1. (b) Model 2. (c) Model 3. (d) Model 4. (e) Model 5.
261
262
C.M.K. Cheung and M.K.O. Lee
(c)
(d) Fig. 2. (continued)
The Structure of Web-Based Information Systems Satisfaction
263
(e) Fig. 2. (continued)
Model 4 hypothesizes six first-order factors and two second-order factors, i.e., web information satisfaction and web system satisfaction (see Figure 2d). Based on McKinney et al.’s model [26], understandability, reliability, and usefulness are dimensions of information quality. We believe these three first-order factors are highly correlated, and their covariations can be captured by a second-order factor (Web information satisfaction). Similarly, access, usability, and navigation are closely related to system quality, and a second-order factor (Web system satisfaction) is proposed to capture their covariations. Model 5 assumes that the two second-order factors in Model 4 are correlated (see Figure 2e). Similar to Model 3, we assume the correlations between the two secondorder factors, so that the common variation in the model can be captured.
5 Research Method The sections below describe the details of data collection procedure, measurement, data analytical approach, and model competing criteria. 5.1 Data Collection The web-based information system in question is known as “Blackboard Learning System (http://www.blackboard.com)”, an Internet-based learning portal for students in campus-based education institutions. Perceived usefulness and perceived enjoyment of
264
C.M.K. Cheung and M.K.O. Lee
such a portal shapes the students’ attitude towards it and ultimately their intention to use [21]. Therefore, it is important that the portal allows students to access to course materials, course announcements, and other relevant documents of each course they are enrolled in. It should also contain communication facilities (e.g., discussion forums, group pages and virtual classrooms) for students to exchange ideas and opinions. The web-based portal was introduced to the first-year undergraduate students at the beginning of the semester. After six-week’s usage, students were invited to voluntarily complete an online questionnaire that covered all the measures of the constructs in this study. A total of 515 usable questionnaires were collected. The respondent rate was 64.4%. Among the respondents, 54.8% were female and 45.2% were male. 5.2 Measurement Table 2 lists the measures used in this study. Basically, we borrowed the measures from McKinney et al. [26] but modified the wordings so as to fit them to this particular context of web-based information systems user satisfaction. The measurements employed a seven-point Likert scale, from “1=never” to “7=always”. Table 2. Lists of the measures used in this study Dimensions Understandability
The information on Blackboard is clear in meaning The information on Blackboard is easy to comprehend The information on Blackboard is easy to read In general, information on Blackboard is understandable for you to use The information on Blackboard is trustworthy The information on Blackboard is accurate The information on Blackboard is credible In general, information on Blackboard is reliable for you to use The information on Blackboard is informative to your usage The information on Blackboard is valuable to your usage In general, information on Blackboard is useful for you to use Blackboard is responsive to your request Blackboard is quickly loading all the text and graphic In general, Blackboard is providing good access for you to use Blackboard is having a simple layout for its contents Blackboard is easy to use Blackboard is of a clear design In general, Blackboard is user-friendly Blackboard is being easy to go back and forth between pages Blackboard is providing a few clicks to locate information In general, Blackboard is easy to navigate
5.3 Data Analytical Approach The proposed factor structures were examined through the LISREL VIII framework. LISREL is one of the most widely used Structural Equation Modeling (SEM) techniques in IS. According to Chin [5], if SEM is accurately applied, it can surpass the first-generation techniques such as principle components analysis, factor analysis, discriminant analysis, or multiple regression. Specifically, SEM provides a greater
The Structure of Web-Based Information Systems Satisfaction
265
flexibility in estimating relationships among multiple predictors and criterion variables. It allows modeling with unobservable latent variables, and it estimates the model uncontaminated with measurement errors. As suggested by Doll et al. [11], competing models should be specified based on logic, theory, and prior studies. The LISREL framework offers us a systematic approach to statistically compare the theoretical models using the goodness-of-fit indexes. The best model is then selected as representing the factor structure and dimensionality of WIS satisfaction in the sample data. Further, the psychometric properties (i.e., reliability and validity) of the selected model are examined. 5.4 Criteria for Comparing Model-Data Fit The determination of model fit in structural equation modeling is not as straightforward as it is in other statistical approaches in multivariate procedures. Chisquare test is the only statistical test that identifies a correct model given the sample data. In contrast to traditional significance testing, the researcher is interested in obtaining a non-significant chi-square. Such a finding indicates that the predicted model is congruent with the observed data. Another alternative is the ratio of the chisquare to the degrees of freedom. Researchers have recommended using Normed ChiSquare as low as 2 or as high as 5 to indicate a reasonable fit [14]. However, Chisquare test is highly sensitive to the sample size and the departures from multivariate normality of the observed variables [4]. Given its sensitivity to many factors, researchers are encouraged to complement the chi-square measure with other fit indexes [15]. In IS research, absolute fit indexes and incremental fit indexes are the two most widely used measures to determine how well the data fits the proposed model. For instance, Doll et al. [11] used absolute fit indexes, including the Goodness-of-Fit Index (GFI) and the Root Mean Square Residual (RMSR), to evaluate individual models. They also used incremental fit indexes, including the Normed Fit Index (NFI) and the Adjusted Goodness-of-Fit Index (AGFI), to reflect the improvement in fit of one model over an alternative. Some researchers [15, 19] provided the criteria and interpretation of these measures.
6 Results The results are reported in this section. The models are analyzed using confirmatory maximum likelihood estimation. We start with the check for multivariate normality, and then we report the model estimation and compare the goodness of fit. The models are analyzed using confirmatory maximum likelihood estimation. Finally, we examine and report the psychometric properties of the model that provides the good-fit to the data and is theoretically valid. 6.1 Checking for Multivariate Normality Multivariate normality is an important assumption of confirmatory factor analysis. To check if our observations are independently and identically distributed, we examined the skewness and kurtosis for each scale. Skewness refers to the lack of symmetry of
266
C.M.K. Cheung and M.K.O. Lee
a data distribution, while kurtosis refers to whether the data distribution is peaked or flat relative to a normal distribution. Skewness for scale items ranged from 0.014 to 0.336 and kurtosis ranged between 0.015 and 1.041 were well within the robustness thresholds for normality. 6.2 Model Estimation Specification of the models included fixing one of the paths from each of the six primary factors at 1.0, and the factor variance for the higher-order factor at 1.0. These are important for model identification. Model 1 fixes the factor variance for the single first-order factor (WIS satisfaction) at 1.0 and allows the 21 observable variables to be free. For Models 2 and 3, the first path for each of the six first-order factors (i.e., understandability, reliability, usefulness, access, usability, and navigation) is fixed to 1.0. For Model 2, the covariances among the six first-order factors are fixed to zero. For Models 4 and 5, the first path for each of the six first-order factors (i.e., understandability, reliability, usefulness, access, usability, and navigation) is fixed to 1.0. The first path for each of the two second-order factors (i.e., Web-IQ satisfaction and Web-SQ satisfaction) is fixed to 1.0. For Model 4, the covariances between WebIQ satisfaction and Web-SQ satisfaction is fixed to zero. For all five models, the number of available data point is p(p+1)/2 = 21 x 22 / 2 = 231. For Model 1, there are 42 free parameters that include 21 error variances for the measured variables and 21 factor loadings. This leaves (231-42) = 189 degrees of freedom for Model 1. There are 42 free parameters for Model 2 which include 21 error variables, a total of (21-6) = 15 factor loadings, and 6 first-order factor variances. This results in 189 degrees of freedom for Model 2. The free parameters for Model 3 include 21 error variables, 15 factor loadings, 15 covariances among the first-order factors, and 6 first-order factor variances. Thus, Model 3 has 174 degrees of freedom. For Model 4, there are 46 free parameters that include 21 error variances for measured variables, 15 first-order factor loadings, 4 second-order factor loadings, 6 primary factor disturbances. This leaves 185 degrees of freedom. Finally, there are 47 free parameters for Model 5, including 21 error variances for measured variables, 15 first-order factor loadings, 4 second-order factor loadings, 6 primary factor disturbances, and 2 second-order factor variances, and 1 covariance between secondorder factors. This provides 182 degrees of freedom. 6.3 Goodness-of-Fit Table 3 summarizes the goodness-of-fit indexes for the five competing models. As expected, the large sample size causes the chi-square statistics of all models statistically significant with p-value < 0.0001. Models 1 to 3 are the first-order factor models. Both Models 1 and 2 provide poor fit to the data, where their fit indexes do not fulfill the recommended acceptance levels. Model 3 provides a good fit to the data with desirable goodness-of-fit indexes, and demonstrates a significant improvement over Model 2. The NFI index increases significantly from 0.66 (Model 2) to 0.94 (Model 3), and the AGFI index improves from 0.43 (Model 2) to 0.87 (Model 3). Model 4 and Model 5 represent the second-order factor models. The two models provide reasonable model-data fit, and their fit indexes are close to the recommended
The Structure of Web-Based Information Systems Satisfaction
267
levels. Comparing Model 4 and Model 5, Model 5 performs slightly better than Model 4, with a lower value of normed chi-square (Model 4: 7.67, Model 5: 3.46) and a higher value of GFI (Model 4: 0.86, Model 5: 0.89). Like Model 3, Model 5 also provides substantial improvement over Model 4. The NFI index increases significantly from 0.86 (Model 4) to 0.94 (Model 5) and the AGFI index improves from 0.82 (Model 4) to 0.87 (Model 5). Table 3. Goodness of fit indexes for competing models (n=515)
Key: GFI – Goodness of Fit Index RMSR – Root Mean Square Residual AGFI – Adjusted Goodness of Fit Index NFI – Normed Fit Index
In comparing the goodness-of-fit among all competing models, we notice that the first-order model (Model 3) performs the best. As suggested by Marsh and Hocevar [23], the purpose of higher-order model is to explain the covariation among the lowerorder factors in a more parsimonious way. In fact, even the higher-order model can explain the factor covariation effectively, its goodness-of-fit can never be better than the corresponding first-order model. Harlow and Newcomb [16] further suggested four guidelines for model selection, including (a) logical or formal consistency, (b) empirical adequacy, (c) the ability to capture most of the essential relations among the variables, and (d) simplicity. Based on these criteria, only Model 3 and Model 5 could be retained. Among the two models, Model 5 is more theoretically valid, reflecting the logical or formal consistency. Similar to the case in Harlow and Newcomb [16], Model 5 presents the relationships in the data in an organized and conceptually descriptive manner. In sum, Model 5 is the most appropriate model to capture the structure of WIS satisfaction. Figure 3 depicts the hierarchical structure of Model 5 with their respective factor loadings and residual variances. Each of the factor loadings is large and highly significant with correspondingly low residence variances, offering further support for Model 5. Indeed, we believe that a third-order factor model with six first-order factors (i.e., understandability, reliability, usefulness, access, usability, navigation), two secondorder factors (Web-IQ satisfaction and Web-SQ satisfaction), and one third-order factor (WIS satisfaction) may provide a richer explanation of the underlying structure of WIS satisfaction. However, this model cannot be uniquely determined and hence cannot be estimated. According to Rindskopf and Rose [31], there must be at least three second-order factors (for the third-order factor model) if the model is to be identified.
Fig. 3. Factor loadings and residence variances in Model 5
6.4 Psychometric Properties After examining the overall model fit, we turn to examine the parameters estimates for Model 5. Table 4 presents the statistical significance of the estimated loadings, their corresponding t values, and R-square values for the 21 observed variables. All items present significant factor loadings, each with a t-value higher than 2.00, on their underlying latent factor. Fornell and Larcker [12] stated that a loading of 0.70 to latent variable is considered to be a high loading since the item explains almost 50 percent of the variance in a particular construct. In our study, all items have high loadings (0.71 or above) to its respective construct. Composite reliability (CR) and average variance extracted (AVE) are also computed to assess the construct validity. A composite reliability of 0.70 or above and an average variance extracted of more than 0.50 are deemed acceptable [15]. As
The Structure of Web-Based Information Systems Satisfaction
269
shown in Table 4, all the measures fulfill the recommended levels, with the composite reliability ranges from 0.87 to 0.93 and the average variance extracted ranges from 0.68 to 0.79. Overall, the measures of the selected model have desirable psychometric properties. Table 4. Parameter estimates for Model 5 (six first-order factors and two second-order factors) Latent Variable Understandability CR = 0.91 AVE = 0.72 Reliability CR = 0.91 AVE = 0.71 Usefulness CR = 0.92 AVE = 0.79 Accountability CR = 0.87 AVE = 0.68 Usability CR = 0.93 AVE = 0.77 Navigation CR = 0.89 AVE = 0.72
Key: CR – Composite Reliability AVE – Average Variance Extracted
7 Discussion and Conclusions The results suggested that WIS satisfaction can be assessed by a large number of highly related factors. The second-order factor model (Model 5) with six first-order factors (i.e., understandability, reliability, usefulness, access, usability, and navigation) and two correlated second-order factors (i.e., web information satisfaction and web system satisfaction) provides a good-fit to the data and is more theoretically valid, reflecting the logical or formal consistency. 7.1 Managerial Implications Understanding WIS satisfaction is particularly important because a high level of WIS satisfaction is associated with several key outcomes, including enhanced IS continuance usage [3], the realization of IS success [8, 9, 32], and improved user performance [13]. In the current study, our higher-order factor model can greatly assist web designers in understanding how users assess web-based information systems satisfaction. Essentially, the model can help explain three basic issues: (1) what defines WIS satisfaction,
270
C.M.K. Cheung and M.K.O. Lee
(2) how WIS satisfaction is formed, and (3) which attributes are relatively important to the formation of WIS satisfaction. These three issues require managerial attention in efforts to improve user satisfaction with the web-based information systems. Thus, we believe our hierarchical structure model can substantially enhance web designers’ conceptualization and understanding of WIS satisfaction. The first issue is that what defines WIS satisfaction. The research results show that all items have high loadings to their respective construct and that they all contribute to end users’ overall WIS satisfaction. These items therefore can serve as a checklist for practitioners, such as web system designer and webmasters, to evaluate end users’ WIS satisfaction of the WIS when designing or maintaining the WIS. Thus, we believe our hierarchical structure model can substantially enhance web designers’ conceptualization and understanding of WIS satisfaction. In addition, the multilevel conceptualization of WIS satisfaction allows for analysis at different levels of abstraction. Web designers can use the complete scale to determine an overall WIS satisfaction, or they can focus on specific areas that are in need of attention (e.g., web information quality or web system quality). The second issue is that how WIS satisfaction is formed. The study results provide sufficient support to the proposition that the six first-order factors forms the two second-order factors and the two second-order factors are correlated to each other. This correlation implies that there is a common variation in the model can be captured. For example, a quality system interface contributes to both well organization of the system contents (i.e., usability of the system, which determines SQ satisfaction) and high easiness to read (i.e., understandability of the system, which determines IQ satisfaction). Practitioners should be aware of the hierarchical structure of the components of WIS satisfaction, but they should consider WIS satisfaction holistically in designing and implementing WIS. The last issue is that which attributes are relatively important to the formation of WIS satisfaction. The results of this study remind practitioners that both IQ satisfaction has the same weight as SQ satisfaction in contributing to end users’ overall web information system satisfaction so that both aspects are worth their attention. As a result, practitioners should spend their effort in both aspects when performing system maintenance. On one hand, irregular technical system quality enhancement of the WIS, such as adding new system features and changing attractive system interface, amazes end users and raises their Web SQ satisfaction. On the other hand, frequent information quality maintenance of the WIS, such as adding new information and amending system contents, impresses end users and raises their Web IQ satisfaction. 7.2 Research Implications This study also has significant implications for academics. In response to the call for developing standardized instruments and completing a research cycle [11], the current study performed a confirmatory factor analysis on the. satisfaction instrument from [26] to test the alternative factor structures of WIS satisfaction and to assess the psychometric properties of the factors and their measuring items. Our results provide a strong support for McKinney et al.’s instrument. A measurement scale with clear
The Structure of Web-Based Information Systems Satisfaction
271
dimensions and reasonable structure like this does not only enrich the literature, but also provides a steady foundation for the research stream. Researchers in social sciences argued that the use of hierarchical factor structure can enhance the conceptualization and the estimation of human judgment models [16, 23]. Similarly, we believe our higher-order factor model can capture users’ overall evaluation of WIS satisfaction through the underlying commonality among dimensions in the second-order factor. Finally, this study demonstrates the advantages of using confirmatory factor analysis (CFA) for comparing alternative factor structures. CFA facilitates researchers to define alternative models for the testing of competing models and to generate parameter estimates of the models. Also, researchers can easily perform model comparisons using subjective indicators. However, indeterminacy of hierarchical models is common when sufficient restrictions are not imposed. This work has been restricted to estimating only second-order hierarchical models. Future research must attempt to find means to estimate higher-order structures. 7.3 Future Research The focus of this study is to examine web-based information systems satisfaction. We adopted classical test theory [21] to investigate the key dimensions of WIS satisfaction and their level of abstraction. Classical test theory assumes that covariation among measures is caused by the underlying latent factor. In other words, changes in the underlying construct will cause changes in the indicators and these measures are referred to as reflective. The reflective measurement approach is used in the current study, as we attempt to complete a research cycle and perform a confirmatory factor analysis on the satisfaction instrument from [26] where their scale development procedures focusing on latent constructs with reflective indicators and first-order factors. In addition, the first-order factors of WIS satisfaction are expected to be closely related to each other, the reflective measurement approach can alleviate the problem of multicolinearity. In recent years, researchers started to raise their concerns on the use of formative- and reflective-indicator measurement models. For instance, Jarvis et al. [18] developed a set of conceptual criteria to determine whether a construct should be modeled as having formative or reflective indicators. MacKenzie et al. [21] suggested some practical guidelines for developing and evaluating constructs with formative indicators. Future studies should pay attention to the choice between formative and reflective models. The WIS adopted in this study is an Internet-based learning portal. Its main users are university students. An obvious extension of this research is to conduct replication studies for other web-based information systems, and to explore the adaptation of this scale in other online environment. Future study also needs to include various user groups and investigate differences in their behavior. For example, future research could consider variations in system navigation, information presentation as well as web support required for different types of online computer application Finally, this study demonstrates the advantages of using confirmatory factor analysis for comparing alternative factor structures. It facilitates researchers to define alternative models for the testing of competing models and to generate parameter estimates of the models. Also, researchers can easily perform model comparisons
272
C.M.K. Cheung and M.K.O. Lee
using subjective indicators; however, indeterminancy of hierarchical models is common when sufficient restrictions are not imposed. This work has been restricted to estimating only second-order hierarchical models. Future research must attempt to find means to estimate higher-order structures. Acknowledgements. The work described in this paper was substantially supported by a grant from the Research Grants Council of the Hong Kong Special Administrative Region, China [Project No. CityU 1361/04H].
References 1. Abdinnour-Helm, S.F., Chaparro, B.S., Farmer, S.M.: Using the End-User Computing Satisfaction (EUCS) Instrument to Measure Satisfaction with a Web Site. Decision Sciences 36, 341–364 (2005) 2. Bailey, J.E., Pearson, S.W.: Development of a Tool for Measuring and Analyzing Computer User Satisfaction. Management Science 29, 530–545 (1983) 3. Bhattacherjee, A.: An Empirical Analysis of the Antecedents of Electronic Commerce Service Continuance. Decision Support Systems 32, 201–214 (2001) 4. Bollen, K.A.: Structural Equations with Latent Variables. John Wiley and Sons, New York (1989) 5. Chin, W.W.: Issues and Opinion on Structural Equation Modeling. MIS Quarterly 22, 7– 16 (1998) 6. Cheung, C.M.K., Lee, M.K.O.: The Asymmetric Effect of Web Site Attribute Performance on Web Satisfaction: An Empirical Study. e-Service Journal 3, 65–86 (2004/05) 7. Cho, N., Park, S.: Development of Electronic Commerce User-Consumer Satisfaction Index (ECUSI) for Internet Shopping. Industrial Management & Data Systems 101, 400– 405 (2001) 8. DeLone, W.H., McLean, E.R.: Information Systems Success: The Quest for the Dependent Variable. Information Systems Research 3, 60–95 (1992) 9. DeLone, W.H., McLean, E.R.: The DeLone and McLean Model of Information Systems Success: A Ten-Year Update. Journal of Management Information Systems 19, 9–30 (2003) 10. Doll, W.J., Torkzadeh, G.: The Measurement of End-User Computing Satisfaction. MIS Quarterly 12, 259–274 (1988) 11. Doll, W.J., Xia, W., Torkzadeh, G.: A Confirmatory Factor Analysis of the End-User Computing Satisfaction Instrument. MIS Quarterly 18, 453–461 (1994) 12. Fornell, C., Larcker, D.F.: Evaluating Structural Equation Models with Unobservable Variables and Measurement Error. Journal of Marketing Research 18, 39–50 (1981) 13. Gelderman, M.: The Relation between User Satisfaction, Usage of Information Systems and Performance. Information and Management 34, 11–18 (1998) 14. Hair, J.F., Anderson, R.E., Tatham, R.L.: Multivariate Data Analysis: With Readings, 3rd edn. MacMillan Publishing Co., New York (1992) 15. Hair, J.F., Anderson, R.E., Tatham, R.L., Black, W.C.: Multivariate Data Analysis, 5th edn. Prentice-Hall, Englewood Cliffs (1998) 16. Harlow, L.L., Newcomb, M.D.: Towards a General Hierarchical Model of Meaning and Satisfaction in Life. Multivariate Behavioral Research 25, 387–405 (1990) 17. Isakowitz, T., Bieber, M., Vitali, F.: Web Information Systems. Communications of the ACM 41, 78–80 (1998)
The Structure of Web-Based Information Systems Satisfaction
273
18. Jarvis, C.B., MacKenzie, S.B., Podsakoff, P.M.: A Critical Review of Construct Indicators and Measurement Model Misspecification in Marketing and Consumer Research. Journal of Consumer Research 30, 199–218 (2003) 19. Joreskog, K.G., Sorbom, D.: LISREL 8: User’s Reference Guide. Scientific Software International, Chicago, IL (1996) 20. Kaschek, R., Schewe, K.-D., Wallace, C., Matthews, C.: Story boarding for Web-based Information Systems. In: Taniar, D., Rahayu, J.W. (eds.) Web Information Systems, pp. 1– 33. IDEA Group Publishing Inc., Hershey (2004) 21. Lee, M.K.O., Cheung, C.M.K., Chen, Z.: Acceptance of Internet-Based Learning Medium: The Role of Extrinsic and Intrinsic Motivation. Information & Management 42, 1095– 1104 (2005) 22. MacKenzie, S.B., Podsakoff, P.M., Jarvis, C.B.: The Problem of Measurement Model Misspecification in Behavioral and Organizational Research and Some Recommendaed Solutions. Journal of Applied Psychology 90, 710–730 (2005) 23. Marsh, H.W., Hocevar, D.: Application of Confirmatory Factor Analysis to the Study of Self-Concept: First- and Higher-Order Factor Models and their Invariance across Groups. Psychological Bulletin 97, 562–582 (1985) 24. McBrath, J.: Towards a Theory of Method for Research in Organization. In: Mowday, R.T., Steers, R.M. (eds.) Research in Organizations: Issues and Controversies, pp. 4–21. Goodyear Publishing, Santa Monica (1979) 25. McHaney, R., Hightower, R., Pearson, J.: A Validation of the End-User Computing Satisfaction Instrument in Taiwan. Information and Management 39, 503–511 (2002) 26. McKinney, V., Yoon, K., Zahedi, F.M.: The Measurement of Web-Customer Satisfaction: An Expectation and Disconfirmation Approach. Information Systems Research 13, 296– 315 (2002) 27. Muylle, S., Moenaert, R., Despontin, M.: The Conceptualization and Empirical Validation of Web Site User Satisfaction. Information & Management 41, 543–560 (2004) 28. Otto, J.R., Najdawi, M.K., Caron, K.M.: Web-User Satisfaction: An Exploratory Study. Journal of End User Computing 12, 3–10 (2000) 29. Rai, A., Lang, S.S., Welker, R.B.: Assessing the Validity of IS Success Models: An Empirical Test and Theoretical Analysis. Information Systems Research 13, 50–69 (2002) 30. Reddy, S.K., LaBarbera, P.A.: Hierarchical Models of Attitude. Multivariate Behavioral Research 20, 451–471 (1985) 31. Rindskopf, D., Rose, T.: Some Theory and Applications of Confirmatory Second-Order Factor Analysis. Multivariate Behavioral Research 23, 51–67 (1988) 32. Seddon, P.B., Staples, S., Patnayakuni, R., Bowtell, M.: Dimensions of Information Systems Success. Communications of the Association for Information Systems 2 (1999) 33. Weeks, D.G.: A Second-Order Longitudinal Model of Ability Structure. Multivariate Behavioral Research 15, 353–365 (1980) 34. Zviran, M., Erlich, Z.: Measuring IS User Satisfaction: Review and Implications. Communications of the Association for Information Systems 12, 81–104 (2003)
Part IV
e-Learning
Introducing Communities to e-Learning Gavin McArdle, Teresa Monahan, and Michela Bertolotto School of Computer Science and Informatics University College Dublin, Belfield, Dublin 4, Ireland {gavin.mcardle,teresa.monahan,michela.bertolotto}@ucd.ie
Abstract. The popularity of e-learning continues to grow as the convenience which it offers appeals to a wide spectrum of people. A significant number of elearning applications have been developed offering tools of benefit to both tutors and students. Although online communities are popular in other areas, the establishment of these communities within e-learning has been slow. Communication and social interaction among students plays an important part of learning within the classroom and so should be incorporated into online learning environments. This paper presents guidelines for the development of e-learning systems which recognise the importance of communities and social interaction within an online learning scenario. In particular, we describe our own e-learning systems, CLEVR and its mobile counterpart, mCLEV-R, along with details of the user trials used to evaluate them. Results indicate that these systems create an effective online learning community for students. Keywords: Web-Based Education, Virtual Learning Communities, Social Awareness, Virtual Reality, M-Learning.
1 Introduction The use of web-based applications and services in the areas of e-commerce, online banking and e-government has grown over the last decade. These technologies are also used for e-learning, which has emerged as one of the most popular online services [1]. Its use by universities and other teaching institutions has established it as a universally accepted means of learning, while the convenience it offers has made it appealing to students. The availability of broadband continues to grow and the increased connection speed allows web-based services to be more sophisticated. This is particularly evident within online learning where e-learning websites are no longer mere repositories of course material which students access. Today, complete systems are available to manage learning content and to support course instructors and their students. These online systems are known as Learning Management Systems (LMSs) or Course Management Systems (CMSs) and examples, shown in Table 1, include BlackBoard and Moodle. They offer mechanisms for tutors to create and present learning content, monitor the level of student participation, and appraise student performance. They also provide communication facilities for students, such as discussion boards and forums, where interaction can take place. J. Filipe and J. Cordeiro (Eds.): WEBIST 2007, LNBIP 8, pp. 277–291, 2008. c Springer-Verlag Berlin Heidelberg 2008
278
G. McArdle, T. Monahan, and M. Bertolotto Table 1. Widely used Learning Management Systems Platform Name BlackBoard WebCT Top Class Moodle Classfronter
Web Address www.blackboard.com www.webct.com www.wbtsystems.com www.moodle.org www.fronter.info
Commercial Status Commercial Commercial Commercial Open Source Open Source
The learning experience offered by these systems does, to some extent, mimic the learning scenario in a physical college or university. Tutors provide course notes, set assignments for the students and then monitor and assess their progress. However, unlike the real world where students interact continually with each other, social interaction within these systems is often limited to discussion boards and forums, which can hinder the sense of community experienced by users. This lack of timely interaction between users is one of the major drawbacks of standard LMSs available at present, as it can lead to feelings of isolation and loneliness for students [2]. Many studies highlight the significance of social awareness and collaboration in any learning situation [3] [4] [5]. Therefore, in order for an e-learning environment to be effective, it should promote a strong sense of community among students. The inclusion of such elements will aid them in their education and also assist in their social development. To facilitate such community building in the e-learning domain, new technologies which can provide more interactive techniques need to be explored. In particular, instant communication, file sharing and multimedia, which have proved popular for building communities in other online domains, can also be applied to the e-learning sector. Such technologies not only allow students to interact freely but also create a motivating and stimulating environment in which to learn. Some e-learning applications have examined these issues. In particular, the use of 3D environments for learning is a growing area of research. For example, they have been explored for teaching specific topics, such as engineering tasks and laboratory protocol [6] [7] [8] [9]. 3D environments have also been developed as interfaces to access online courses [10] [11]. The tools provided within these interfaces vary from system to system. Many present learning content through multimedia and, while some support real-time communication among students, they do not provide dedicated community building tools. To address this issue, we have developed a system called CLEV-R which provides a 3D e-learning environment with specific features to promote and encourage social interaction and awareness among users. Facilities for delivering learning content in real-time are also provided. This makes the learning experience more appealing to users and also proves effective for supporting several educational activities. The remainder of this paper is organised as follows; in Section 2 the emergence of online communities and the technologies used to create them are explored. Section 3 discusses communities in the online learning scenario and provides some guidelines for the development of an effective e-learning system. In Section 4, CLEV-R, the 3D multiuser interface for e-learning, we have developed, is discussed. This interface promotes the formation of an online community and encourages social interaction among users, while also supporting a variety of learning activities. Section 5 introduces mCLEV-R, a
Introducing Communities to e-Learning
279
mobile learning (m-learning) system which we have created for use on portable devices as an additional resource for accessing learning material anytime anywhere. Details of user trials involving both of these systems are discussed in Section 6, while finally Section 7 summarises the research presented here and presents some future directions for this work.
2 Online Communities As more and more individuals have access to the Internet, there has been a proliferation in the use of online communities. These communities allow groups of people in different geographical locations to come together and communicate their ideas, thoughts and feelings with other, often similar minded people [12]. Such communities have been in use since the advent of the Internet and, in their simplest form, can consist of an email distribution list to share information about specific topics. Web-based discussion boards and forums are another form of online community where users can share expertise and give advice on certain subjects. More recently, e-commerce websites are incorporating these techniques into their business to encourage the formation of communities among their customers. People can leave feedback, discuss products and receive after-sales service through this medium. More advanced technologies have also emerged which can further enhance online communities by enabling increased interaction through the use of synchronous and instant communication methods. These techniques are an effective means to allow people to share information and easily maintain contact with others. One such real-time communication method which has grown in popularity is instant messaging. It can be used as a stand alone communication tool, but it can also be integrated into websites to create a community between users. Chat rooms are a prime example of these. A wide variety now exist, many specialised to a particular theme, where interested parties can communicate using text-based messages. Some of these communities are now using instant audio communication methods and webcam broadcasts. Onscreen Virtual Reality (VR) can also be used to form online communities where a shared 3D world is rendered, in which each member is represented by an onscreen persona or character. Table 2 includes details of ActiveWorlds and Second Life which are two of the most popular 3D social communities. Users become citizens of an online 3D world and can design and build their own areas of this 3D community. Text-chat is the primary communication method within these 3D environments. In addition to those discussed above, a new type of online community is now emerging. Social networking websites, such as those highlighted in Table 2, offer users the ability to build a social network of friends online. Bebo and MySpace are two such websites [13]; within these communities, members of the public can produce a web page containing information about themselves along with photos and videos. They can also link to their friends’ pages to build up the community network. Other users can browse these web pages, post messages and share ideas while building new friendships online. Within these social networks instant communication services, such as those shown in Table 3, can also be used. A number of dedicated websites have developed recently, which promote file sharing in a similar way to social networking websites. For example, Flickr, (see Table 2), is a
280
G. McArdle, T. Monahan, and M. Bertolotto Table 2. Online Systems for Social Interaction Platform Name Active Worlds Second Life Bebo MySpace Facebook YouTube Flickr
Web Address www.activeworlds.com www.secondlife.com www.bebo.com www.myspace.com www.facebook.com www.youtube.com www.flickr.com
Description 3D Social Interaction Platform 3D Social Interaction Platform Social Networking Website Social Networking Website Social Networking Website Video Sharing Website Photo Sharing Website
Table 3. Instant Communication Services Service Name MSN Messenger Yahoo! Messenger Skype
Web Address http://messenger.msn.com http://messenger.yahoo.com http://www.skype.com
photo sharing community platform, where members of the public can provide photos for others to view online [14]. Users can annotate these with descriptions which assist others when searching for a particular image. Group discussions are also facilitated through message boards. In a similar vein, YouTube is an online community for sharing videos. Members can upload video clips to the YouTube website, where others can then view them and comment on their content. As is the case with the other online communities discussed, the popularity of these file sharing websites has been growing immensely.
3 Community in e-Learning As the wide diffusion of the above technologies highlights, there is a great interest in online communities. The use of the latest technologies allows groups of people to extend their social network beyond the normal geographical boundaries and this appeals to users and is something they have come to expect in their online activities [15]. Communities form online, in a similar way to that in the real world. For example, people who attend the same places and have similar interests naturally form a social network. The same is true in a learning scenario where social interaction and networks are very important. Research has shown that people can learn as much from each other in informal meetings as they do in a structured learning session [3] [4]. Despite this, opportunities for social interaction and networking are often absent in conventional e-learning systems where the primary focus is to provide access to learning materials. Many research studies highlight the importance of communication within e-learning systems [16] [17] [18]. However, traditional e-learning applications offer minimal support for person-toperson communication via asynchronous communication methods such as discussion boards, bulletin boards and online forums. Real-time communication is often absent and this hinders the formation of social communities, which can lead to a feeling of
Introducing Communities to e-Learning
281
isolation for students [2]. In other cases this lack of contact with tutors and fellow students can cause students to withdraw from a course prematurely [19]. With these issues in mind, e-learning platforms need to be more engaging and motivating by allowing users to build social networks and collaborate with each other in real-time. Thus, the following specifications should be considered in the design of e-learning systems. 1. Students should be aware of others in their class and interact socially with them. 2. Students should be able to communicate with others, tutor and students alike. 3. Students should be able to collaborate on learning tasks and group assignments. We have developed an e-learning application called CLEV-R, and an m-learning system called mCLEV-R which focus on social interaction and the provision of collaboration tools, while still effectively supporting different learning activities. The following two sections provide an overview of these systems, in particular describing the methods employed to realise an effective online learning community.
4 CLEV-R CLEV-R is a web-based collaborative learning environment which uses virtual reality to present the learning experience to users. An onscreen 3D university is used as an interface where synchronous lectures and group meetings are facilitated. The 3D environment is augmented with tools which support communication among students. Learning activities within the environment are supported through multimedia presentations, reviewing and downloading of course notes, and an extensive communication network. A more in depth description of CLEV-R, including details of a sample session are described in [20]. The 3D environment together with the communication tools play an important role in developing an online community and, when combined, they are key in realising the three specifications outlined above. Figure 1 displays the CLEV-R interface as it appears within a web browser. 4.1 Awareness The use of 3D in CLEV-R creates an interactive and visual learning experience and makes users aware that they are present within a shared multi-user learning environment. It also plays an important role in making students aware of others taking part in the same activities. Each user is represented within the environment by a 3D character known as an avatar (see Figure 1). When a student registers for a course they select this character, which remains their onscreen representation for the duration of that course. Users can move their avatar within the virtual 3D university and, as they do so, their current location is displayed in the 3D environments of all other users which creates a sense of presence for them within a shared space. Each avatar is unique with different hairstyles and clothing, and has its user’s name displayed above it so that they can be easily recognised by their peers. The interface also features a map of the 3D environment, which shows the location of all other connected users. The Graphical User Interface (GUI) which hosts the communication controls for the system, shown in Figure 2, also promotes awareness of others by displaying a textual list of all other users. These features make students visibly aware of others within the system.
282
G. McArdle, T. Monahan, and M. Bertolotto
Fig. 1. A Synchronous Lecture Taking Place in the CLEV-R 3D Environment
4.2 Communication CLEV-R supports multiple modes of communication including text-chat, audio-chat and webcam broadcasts, all of which are live and occur in real-time. These methods can be used for learning activities, such as lectures where they can be used by tutors to present lecture commentary or by students to pose questions and receive immediate feedback. Figure 2 highlights the text and audio communication controls on the GUI. The text-chat controls facilitate the sending of public messages to all other users or private messages to individual users within the 3D environment. The audio communication is broadcast on specific channels which correspond to different locations within the 3D environment. For example, students in a virtual meeting or social room can broadcast their voice and listen to others who are also present in this location by selecting the appropriate broadcast channel. Live webcam feeds which can also be displayed in certain rooms, work in a similar way. Avatar gestures are another form of communication
Introducing Communities to e-Learning
283
Fig. 2. The Graphical User Interface Housing the Communication Facilities
Fig. 3. Students Share a YouTube Video in the CLEV-R 3D Environment
within the 3D environment. Users can instruct their onscreen character to raise or lower their hand and to nod or shake their head. These actions allow users to communicate certain desires visually which others can easily understand. For example, a student can raise their hand during an online lecture to indicate that they wish to ask a question.
284
G. McArdle, T. Monahan, and M. Bertolotto
4.3 Collaboration Within CLEV-R, collaboration is facilitated through the sharing of files and actively communicating with one another. In online lectures the tutor can upload learning material to the presentation board in the lecture room for all students to see. However, in group meeting rooms students also have this ability and so can share their files with each other to aid group discussions and project work. Students can also use these techniques to socialise with one another. Dedicated rooms and areas of the 3D environment support informal interaction among students. Media boards can be used to share music and video files, where students can upload their personal files or even their favourite YouTube videos (see Figure 3). Together with the instant communication methods described above, students can thus work, learn and socialise with each other within the CLEV-R environment.
5 Mobile CLEV-R The appeal of mobile devices has increased over the last few years. As they become more powerful, they have emerged as an ideal medium for providing various services to users ’anytime, anywhere’. The provision of learning services on these devices has recently been examined by the research community. The term m-learning is now used to describe learning services provided on mobile devices such as mobile phones, smartphones and Personal Digital Assistants (PDAs). The benefits of mobile devices within learning have led us to explore this area. We have developed a system for PDAs that acts as an accompaniment to CLEV-R. This mobile application called mobile CLEV-R (mCLEV-R), provides users with access to course content and communication facilities while on the move. Acting as a supplement to the desktop system, it can be used by students who cannot be at a fixed location for their online learning. Due to some limitations with PDAs, such as the smaller screen size and decreased memory and processing power, all of the functionality provided in the desktop system cannot be extended to the mobile platform. However, mCLEV-R still permits mobile users to actively partake in learning, collaboration and social activities and the 3 specifications outlined in this paper are also addressed for this version of the system. A 3D environment is used to present the learning material to students. However for the mobile platform it has been simplified and the multi-user features found in the desktop version are not supported. It instead uses a single user office style environment seen in Figure 4(A). Students interact with this environment to access course materials and the communication facilities. While avatars are not used within mCLEV-R, user awareness is created through the communications GUI. Text and audio communication is supported via this GUI shown in Figure 4(B). Using these mediums, mobile users can communicate freely with all currently connected users; both those connected through the desktop CLEV-R interface and those using the mCLEV-R system. As in the desktop version, the text-chat facility can be used to send public and private messages to other users. The communication GUI provides a list of all other connected users which can be used to select the receiver of a message. The audio communication facility allows users to broadcast their voice live into different areas of the 3D university environment provided in the desktop system.
Introducing Communities to e-Learning
(A)
285
(B)
Fig. 4. The 3D Interface and Communication GUI of mCLEV-R
As mobile users have no access to the shared 3D environment, they need another means to determine the location of others. When a mobile user selects a room to broadcast to using the communication GUI, they are presented with a list of all users currently in that location. Therefore they are aware of who is listening to their broadcast. Users accessing CLEV-R via the desktop, are aware of the presence of mobile users as their list of other connected users reflects their presence. These communication techniques can also be used in the context of collaboration and social interaction where mobile users can actively partake in group meetings and discussions without actually being restrained to a fixed desktop PC. Students can also partake in online lectures by downloading the course notes and listening to the live commentary by the lecturer. Full details on how users can interact with the system for learning and communicating can be found in [21]. mCLEV-R provides an effective means for users to participate in synchronous learning and collaborating while on the move, and also creates a social awareness of other students in their course.
6 Evaluation Initial evaluation studies of CLEV-R and mCLEV-R have been completed. The focus of our research is not concerned with the learning content, but rather on the methods used to present it to students and the studies were conducted in this context. In these studies nine test subjects interacted with the 3D environment of CLEV-R, while five participants used the mCLEV-R application. The test subjects were instructed to carry out a number of distinct tasks to ensure they were exposed to the various features of the learning interfaces. These included attending an online lecture, partaking in a group meeting
286
G. McArdle, T. Monahan, and M. Bertolotto
Fig. 5. User Feedback Regarding the Social Aspects of CLEV-R
and using the tools provided to collaborate on a group task and solve a given puzzle. Social interaction was also encouraged and participants used the features provided in the social areas to communicate and share photos with each other. Following the user trials the subjects were given a questionnaire to obtain their opinions on various aspects of the systems. In particular, we wished to discover any usability problems which were experienced by users during their evaluation sessions. Participants’ views regarding the social aspects of both interfaces and the educational benefits that they provide were also obtained. The results from these questionnaires provided invaluable feedback which enabled the interfaces to be improved to give greater support for students. 6.1 Usability Results Overall, positive results were returned from test subjects in relation to the usability of both applications and the various tools they provide. 100% of users found navigation within the 3D environment of CLEV-R intuitive and the interface easy to use. All test users of this desktop interface actively took part in an online lecture and successfully viewed and downloaded course notes. They also uploaded their own files during a group meeting and used the communication controls effectively. While all subjects were able to use the system without any major issues arising, a small number of minor issues were revealed. These included problems using the audio communication feature and navigating through doorways within the 3D environment. These issues have since been addressed.
Introducing Communities to e-Learning
287
Similarly all test subjects of mCLEV-R found the 3D environment appealing. They successfully accessed the communication facilities, announcements page and downloaded course material using the 3D interface. Each participant gave a positive response to the intuitiveness of accessing these facilities. The majority of users found the communication facilities effective for partaking in the various tasks, however some usability issues relating to the communication GUI of mCLEV-R were raised. In particular, participants found scrolling between the text and audio communication sections of the interface frustrating and many suggested that the controls should be placed in closer proximity to each other. While using the audio communication tool, mobile users disliked that they had no visual indication of who the current speaker was. The mCLEV-R interface has since been modified to address these concerns. 6.2 Social Aspects One key area of interest for this study was the reaction of users to the social tools provided in the CLEV-R and mCLEV-R systems, and in particular the sense of community experienced. Feedback in relation to this aspect of the desktop CLEV-R interface is displayed in Figure 5. Overall participants’ responses were extremely positive. While using the system 78% of the test subjects felt part of a group, they were all continually aware of each other and no one felt isolated during the evaluation. Furthermore, 89% felt the environment was an effective means of social interaction. When questioned about the effectiveness of the tools for sharing files, such as photos, 22% did not agree that the tools provided were an effective medium for this. It was later discovered that these test users had not uploaded their photos correctly and so others did not see them, which could account for their negativity towards this feature. Clearer instructions on how to use these tools have now been added to the system, which should prevent similar experiences in the future. Communication within CLEV-R was deemed to be effective by 78% of the test subjects. Those that returned a negative response to this question were persons who experienced echo and feedback problems during the trial, a problem that has now been rectified. One test subject specifically highlighted the real-time aspect as being particularly valuable. The results for the community and social aspects of the system show that CLEV-R, along with the social tools it provides, is an effective medium for interaction among users. Feedback regarding the social aspects of the m-learning application was also gathered. Within the mCLEV-R application, the communication interface aims to promote awareness of others and form an online learning community allowing students to interact and learn together. 100% of the participants in our evaluation study stated that they were aware of the presence of all other users and 80% felt that the presence of others enhanced their learning experience. However, due to the usability problems outlined above, some users gave negative feedback in relation to the effectiveness of the communication facilities for interacting socially with others. A redesign of the communication interface to address the usability problems discovered should improve these results. In addition, some of the participants were inexperienced PDA users and as a result found the stylus and virtual keyboard input method of these devices difficult to use. We believe that this will not be an issue when users become more familiar with this input method.
288
G. McArdle, T. Monahan, and M. Bertolotto
Fig. 6. User Feedback Regarding the use of mCLEV-R for M-Learning
6.3 Additional Issues As CLEV-R and mCLEV-R are learning environments, we were interested in ascertaining participants’ views on the value that the desktop and mobile systems can bring to learning. Again results relating to the desktop CLEV-R application were positive with all participants agreeing that it has potential for improving online learning. The results show that 100% of the test subjects were engaged in the interactive environment and their interest in learning was maintained throughout the trial. Users of traditional e-learning platforms often cite boredom as a major drawback with such systems and so this positive result for CLEV-R is very encouraging. 100% of test subjects saw the lecture room and the facilities it provides as an effective means for online learning. The collaboration tools that the students used to work together functioned as expected and the subjects employed them effectively to solve a given puzzle. Although geared towards the educational end of the spectrum, the users saw these tools for group work as a social feature, which made learning more fun. The ability to converse with other members of the group about the task was seen as very beneficial. Feedback in relation to the benefits of mCLEV-R as a tool for accessing learning material was equally positive. All test subjects could see the advantages of providing learning tools for mobile users and their general feedback regarding the use of mCLEVR for m-learning is displayed in Figure 6. 100% of the participants felt that mCLEVR provides a good medium for learning. They agreed more strongly that it is a good learning tool for users when they cannot be in a fixed location, perhaps suggesting that
Introducing Communities to e-Learning
289
a desktop learning application is more favourable to mCLEV-R when possible. We are satisfied with this response as mCLEV-R was not developed as a standalone application and users were happy with the functionality it provides for users on the move. 80% of participants had a high enjoyment level when using the system and when asked whether they would use mCLEV-R if it was provided with an e-learning course, all test-subjects replied positively. These evaluation studies did uncover some usability problems with the applications, which helped us to improve the functionality provided by the systems. Overall feedback was encouraging. Test subjects who had previous experience of e-learning systems were positive about CLEV-R and highlighted the collaborative and community aspects of the system as particular benefits. Further more extensive user evaluations will be conducted in the near future where the trials will focus on measuring the usability of the systems, gauging opinions about the social community aspects as well as determining how beneficial the 3D interfaces are within the e-learning paradigm.
7 Conclusions Social networks, supported by computer mediated communication on the Internet, continue to grow in popularity. As they do so, an increased number of services are emerging in this area. Social networks and a social support infrastructure is important in the traditional face-to-face learning scenario. However its introduction into e-learning has been slow. Conventional e-learning applications tend to be text-based where communication is limited to asynchronous message boards and forums. This can lead to a lack of interaction both with learning content and other users, and does not readily support the formation of an online community. In this paper we have provided a set of suggestions to address this issue. We have used these guidelines to design and create a 3D online e-learning application called CLEV-R. CLEV-R provides an interactive VR environment in which students are immersed. The use of an avatar to represent each user creates an awareness of others. Several communication tools provide a mechanisms for users to interact with each other synchronously. These instant communication techniques help to foster a community among students while collaborative tools encourage students to work together. The tools provided for teaching permit discussion and interaction to take place during online lectures, while features in the social areas of the environment allow informal interaction and sharing of files among students. In addition to providing the CLEV-R interface for use on a desktop computer, it has also been adapted for use on mobile devices(m-CLEV-R). In the mobile application, similar functionality allows students to partake in synchronous online lectures and access course material. Real-time communication is supported via a communication interface and allows students to interact socially with each other. Initial evaluations have shown that both CLEV-R and mCLEV-R performed favourably and that the test subjects involved in the study appreciated the community and social features provided. These initial evaluations are a precursor for a larger study, incorporating more test subjects and an increased emphasis on monitoring the use of the social and learning tools provided. The specifications outlined here, and adhered to during the implementation of our systems, will be of benefit within the e-learning paradigm. The
290
G. McArdle, T. Monahan, and M. Bertolotto
community and collaboration aspects will attract more students to e-learning, engage them and maintain their interest in their studies. Acknowledgements. The CLEV-R and mCLEV-R projects have been partly funded by the Irish Research Council for Science, Engineering and Technology (IRCSET) under the National Development Plan. We also wish to thank the volunteers who took part in our user trials and evaluation.
References 1. Hiltz, S.R., Turoff, M.: Education Goes Digital: The Evolution of Online Learning and the Revolution in Higher Education. Communications of the ACM 48(10) (2005) 2. Kamel Boulos, M.N., Taylor, A.D., Breton, A.: A Synchronous Communication Experiment within an Online Distance Learning Program: A Case Study. Telemedicine Journal and eHealth 11(5), 283–293 (2005) 3. Redfern, S., Naughton, N.: Collaborative Virtual Eenvironments to Support Communication and Community in Internet-Based Distance Education. In: Proceedings of the Joint International Conference on Informing Science and IT Education, Cork, Ireland, pp. 1317–1327 (2002) 4. Laister, J., Kober, S.: Social Aspects of Collaborative Learning in Virtual Learning Environments. In: Proceedings of the Networked Learning Conference, Sheffield, UK, March 26 28 (2002) 5. Kitchen, D., McDougall, D.: Collaborative Learning on the Internet. Telemedicine Educational Technology Systems 27, 245–258 (1998) 6. Chee, Y.S., Hooi, C.M.: C-VISions: Socialized Learning through Collaborative, Virtual, Interactive Simulations. In: Proceedings of the Conference on Computer Support for Collaborative Learning (CSCL), Boulder, Colorado, USA, January 7 - 11, 2002, pp. 687–696 (2002) 7. Nijholt, A.: Agent-Supported Cooperative Learning Environments. In: Proceedings of the International Workshop on Advanced Learning Technologies, Palmerston North, New Zealand, December 4 - 6 (2000) 8. Rickel, J., Johnson, W.L.: Virtual Humans for Team Training in Virtual Reality. In: Proceedings of the Ninth International Conference on AI in Education, Le Mans, France, July 19 23, pp. 578–585 (1999) 9. Bouras, C., Giannaka, E., Kapoulas, V., Nani, M., Tsiatsos, T.: Virtual Learning Communities-Theory and Practice: the Case of VirRAD. In: Proceedings of E-learn 2003, Phoenix, Arizona, USA, November 7 - 11, pp. 495–498 (2003) 10. Bouras, C., Giannaka, E., Tsiatsos, T.: Virtual Collaboration Spaces: The EVE Community. In: Proceedings of the Symposium on Applications and the Internet (SAINT), Orlando, Florida, USA, January 27 - 31, pp. 483–455 (2003) 11. Bouras, C., Triantafillou, V., Tsiatsos, T.: Collaborative Learning Environment using Distributed Virtual Environments. In: Proceedings of the World Conference on Educational Multimedia, Hypermedia & Telecommunications (ED-MEDIA), Tampere, Finland, June 25 - 30, pp. 173–178 (2001) 12. Wellman, B.: Community: From Neighbourhood to Network. Communications of the ACM, vol 48(10), 53–55 (2005) 13. Scott, M.: MySpace: No Free Ride in Europe. Business Week, Published (September 11, 2006) 14. Weiss, A.: The Power of Collective Intelligence. Collective Intelligence 9(3), 16–23 (2005)
Introducing Communities to e-Learning
291
15. Aizlewood, A., Doody, M.: Seeking Community on the Internet: Ethnocultural Use of Information Communication Technology. In: Proceedings of the International Symposium on Technology and Society (ISTAS), Raleigh, North Carolina, USA, June 6 - 8, pp. 5–12 (2002) 16. Hamburg, I., Lindecke, C.: Social Aspects of E-Learning and Blending Learning Methods. In: Proceedings of the Fourth Europen Conference E-Comm Line, Bucharest, Romania, September 25 - 26, pp. 11–15 (2003) 17. Salmon, G.: E-tivities: The Key to On-line Learning. RoutledgeFalmer, Oxford, UK (2002) 18. Thurmond, V., Wambach, K.: Understanding Interactions in Distance Education: A Review of the Literature. Instructional Technology and Distance Learning 1(1), 9–33 (2004) 19. Serwatka, J.A.: Improving Retention in Distance Learning Classes. International Journal of Instructional Technology and Distance Learning 2(1), 59–64 (2005) 20. McArdle, G., Monahan, T., Bertolotto, M.: 3D Collaborative Virtual Environments for ELearning and M-Learning. In: Proceedings of the International Conference on Web-Based Education (WBE), Peurto Vallarta, Mexico, January 23 - 25, pp. 1–6 (2006) 21. Monahan, T., McArdle, G., Bertolotto, M.: Interactive and Collaborative M-Learning. In: Proceedings of the International Conference on Interactive Computer Aided Learning (ICL), Villach, Austria, September 27 - 29, pp. 1–6 (2006)
Evaluating an Online Module on Copyright Law and Intellectual Property Carmel McNaught1, Paul Lam1, Shirley Leung2, and Kin-Fai Cheng1 1
Centre for Learning Enhancement and Research 2 University Library System The Chinese University of Hong Kong, Hong Kong {carmel.mcnaught,paul.lam,shirleyleung,gap.cheng}@cuhk.edu.hk
Abstract. Various forms of knowledge can be distinguished. Low-level learning focuses on recognition and remembering facts. Higher level learning of conceptual knowledge requires the development of some form of mental structural map. Further, application of knowledge requires learners to put theories and concepts into use in authentic and novel situations. This study concerns learning at a number of levels. The context is a fully online module on copyright laws and intellectual property, designed as an introductory course for all postgraduates at a university in Hong Kong. The paper also explores whether the knowledge learnt through the web-based medium was retained after three to six months. Findings ascertained the effectiveness of the new medium, not only in delivering facts but also for assisting the learning of higher level knowledge. As expected, the performance of students declined in the delayed post-tests but not to any alarming degree. Retention of factual knowledge, however, was much lower than retention of other forms of knowledge. This perhaps suggests that the role of e-learning, just as in face-to-face classes, should focus on concepts and the applied knowledge, rather than on memorization of facts alone. Keywords: Specific facts, schematic knowledge, schema theory, knowledge retention, e-learning, copyright laws, intellectual property.
Evaluating an Online Module on Copyright Law and Intellectual Property
293
There are has been a great deal of debate over the ‘knowledge’ level which is somewhat problematic because the word knowledge, in common usage, has a broad range of meanings. The revised Bloom’s taxonomy [2], [3] tackles this challenge and contains two dimensions instead of one – a knowledge dimension and a cognitive process dimension. The knowledge dimension now clearly classifies and distinguishes between forms of knowledge: factual knowledge, conceptual knowledge, procedural knowledge and metacognitive knowledge (Table 1). Anderson and Krathwohl (2001) [2] described factual knowledge as “knowledge of discrete, isolated content elements”; conceptual knowledge as involving “more complex, organized knowledge forms”; procedural knowledge as “knowledge of how to do something”; and metacognitive knowledge as involving “knowledge about cognition in general as well as awareness of one’s own cognition” (p. 27). As educators we are interested in students acquiring conceptual, procedural and metacognitive knowledge, as well as factual knowledge. It is somewhat paradoxical that formal education has often overemphasized factual knowledge in beginning classes, calling such knowledge ‘foundation knowledge’, and then expected students to make the transition to other forms of knowledge with little overt support. For example, Students who achieve higher grades on essay-based examinations show conceptual organization of knowledge while simple listings of facts and concepts are correlated with low grades [4]. The development of mental structural maps of knowledge [5] and “accompanying schematization of knowledge is what educators surely hope to occur in their students” ([6], p. 633).
2 Learning and Knowledge Retention in e-Settings The use of the web as a strategy to deliver learning activities has been of growing importance as technology advances. Research studies have been carried out to evaluate the effectiveness of e-learning in achieving learning outcomes. While many studies claimed that students learn well in the new media, most of these studies did not differentiate or compare the forms of knowledge being investigated. This paper compares and contrasts students’ learning on four levels of knowledge in an online course. The first objective is to investigate whether e-learning can support the acquisition of higher order knowledge. For e-learning to be an effective learning tool, it has to be able to facilitate acquisition of knowledge at the higher levels. The second objective of the study is to explore how well the knowledge acquired at these various levels is retained. The study of knowledge retention in non-web settings in general tends to show that the retention rate for specific facts falls behind that for a broader base of more general facts and concepts [7]. For example, Conway, Cohen and Stanhope (1991) [8] studied very long-term knowledge retention by monitoring the performance of 373 students over ten years on tasks related to a cognitive psychology course. They found that “the decline in retention of concepts is less rapid than the decline in the retention of names” (p. 401). This finding supports Neisser’s (1984) [9] schema theory that describes how conceptual knowledge is developed when students construct linkages between specific facts in their minds. Such linkages or webs or maps are called knowledge schema.
294
C. McNaught et al.
They are more resistant to forgetting than isolated pieces of detailed knowledge. There might be exceptional cases, though, if the specific facts are involved in very personal contexts. Herbert and Burt (2004) [10] suggested that context-rich learning environments (such as problem-based tasks or tasks with connections to learners’ own lives) allow the building of a rich episodic memory of specific facts and this improves the motives of learners to pay attention to learning. Learners are “more likely to then know the material and schematize their knowledge of the domain” (p. 87). Relatively little is known, however, about learning and knowledge retention patterns in e-settings. Yildirim, Ozden and Aksu (2001) [11] compared the learning of 15 students in a hypermedia learning environment with that of 12 students in a traditional situation. They found that students learnt and retained knowledge better in the computer-based environment, not only in the lower-level domains that were about memorization of declarative knowledge, but also in the higher domains of conceptual and procedural knowledge. Bell, Fonarow, Hays and Mangione (2000) [12], however, in their study with 162 medical students, found that “the multimedia textbook system did not significantly improve the amount learned or learning efficiency compared with printed materials … knowledge scores decreased significantly after 11 to 22 months” (p. 942). The problem with many of these studies is that the design of the online module does not provide any advantage over the printed version from the students’ perspective [13]. We were conscious of the need to design for a learning advantage when deciding to use a fully online module. This paper aims to provide further information about knowledge retention in an online course through analysing student performance levels on a fully online introductory course for postgraduate students on copyright law and intellectual property. The course was structured to include learning activities on four levels: (1) specific facts, (2) more general facts and rules, (3) concepts, and (4) applied knowledge. These are related to the revised Bloom’s taxonomy in Table 1. For the fourth category, we will use the term ‘applied knowledge’ but, as shown in Table 1, the tasks in this category require some analytic skills. These categorizations are only indicative. Table 1. Knowledge levels in the online module and the revised Bloom’s taxonomy Knowledge Factual Conceptual Procedural Metacognitive Rem=Remember Anal=Analyze
Rem (1)
Cognitive process App Anal
Und
Eval
Cre
(2) (3)
Und=Understand Eval=Evaluate
(4) App=Apply Cre=Create
3 The Context of the Study The topic of avoiding infringement of copyright law is central to research ethics and includes issues of honesty, credit-sharing and plagiarism. As the computer is used increasingly to disseminate information, teaching professionals also must have knowledge of the applications of the law to this developing technology [14]. There is a
Evaluating an Online Module on Copyright Law and Intellectual Property
295
growing need to introduce copyright policies into university libraries [15]. It is vital to equip students with knowledge about copyright and intellectual property, and to warn them against plagiarism. The situation is particularly true in Hong Kong as the issue of intellectual property and copyright law in research and study-related environments is currently receiving a great deal of attention in the academic community. The Government has recently revised the ordinances and laws governing copyright in Hong Kong. These laws are being interpreted and re-interpreted by many different people and interest groups. The need to educate students properly on these issues is thus particularly important. The University Library of The Chinese University of Hong Kong (CUHK) teaches all postgraduate students a module titled ‘Observing intellectual property and copyright law during research’. This course is a compulsory module for research postgraduate students; students need to complete this online module before their graduation. As the situation is very fluid in Hong Kong, the course has been designed to tackle the issue in as many practical ways as possible. Whatever the laws in Hong Kong are, it must be clearly stated that many of the issues surrounding intellectual property in academic circles are universal, and not just applicable in Hong Kong. There are two major components on the online module: the learning resources and the test. The module was originally conducted solely through face-to-face workshops organized by the Library. The problem with this method was that teaching was restricted to designated times and places. In recent years, the Library has been investigating the potential benefits of putting the course online, similar to the University of Illinois at Chicago [16]. An online version of this particular topic was deemed to be an appropriate strategy for the following reasons: The course is an introductory course: Most of the study materials are easy to understand. This type of content is good for self-learning through students’ individual reading and consideration of the online materials. Students are from various disciplines: Gathering students physically for a lesson has always been difficult, as they have conflicting timetables. With e-learning methods, learning can take place on-demand, and students can be given greater control over their learning than before [17]. Online learning might be effective: Our reading provided sufficient examples of studies where higher level learning seemed to be supported by an online environment. For example, Iverson, Colky and Cyboran (2005) [18] compared introductory courses held in the online format and traditional format. Their findings suggested that online learners can gain significantly higher levels of enjoyment and significantly stronger intent to transfer their learning to other contexts. With effect from 2004–05, the format of this compulsory module was changed from lecture-based to online-based. The online version of the course was run the second time in the academic year 2005–06. The online module is offered four times a year from September to April each year and it is offered under the ‘Research’ section of CUHK’s Improving Postgraduate Learning programme (http://www.cuhk.edu.hk/clear/library/booklet/29.htm). At the time of writing, the module had been run eight times.
296
C. McNaught et al.
4 Online Module 4.1 Procedure All postgraduate students are entitled to enrol in the course. In fact, they are required to take (and pass) the course before their graduation. There are four cohorts each year, and each cohort last for about two months. Eligible students may enrol themselves into any one of the cohorts; they then have to complete the course and the course-end test within the two-month duration.
Fig. 1. Flow of learning activities of the online course
The flow of the course is illustrated in Figure 1. In order to complete the course, students are required to complete the following four tasks in sequence within the course period: 1) 2) 3) 4)
read ALL the course materials; take the online exam; fill in the online survey (not viewable until the exam is submitted); and view their own exam results (not viewable until the survey is submitted).
Students can attempt the summative test only after they have completed reading all the course materials. 4.2 Course Content The learning resources consist of 28 pages of course content that focuses on five areas of issues: copyright around the world and copyright in Hong Kong cover specific facts such as history and the enactment bodies of copyright laws in Hong Kong and around the world; permitted act for research and private study in Hong Kong introduces the more general facts and rules governing the accepted academic practices; avoiding plagiarism is a conceptual section as it defines plagiarism and explains various related concepts; lastly, intellectual property & copyright focuses on the applied knowledge by showcasing various real-life situations and commenting on appropriate practices. Each page contains easy-to-read materials; some are linked to PowerPoint slides and/or further readings. In all, the course gives students a clearer understanding of the core issues of intellectual property, copyright law and plagiarism
Evaluating an Online Module on Copyright Law and Intellectual Property
297
in academic research. The course provides advice about compliance in ‘dealing with’ intellectual property ‘fairly’. 4.3 Course-End Test The summative test consists of 20 questions randomly selected from a pool of 29 questions. The pass mark of the test was 10. Students fail the test if they score 9 or below. If they do so, they need to retake the course. The test questions followed the course structure and asked students’ knowledge on the five themes described above. The questions were set at different levels of knowledge. Specific facts are “facts that referred to details of specific theories and findings highlighted in the course” ([8], p. 398). They are related to a restricted setting. Example test questions include “Where was the Convention signed in the 19th century which protects literary and artistic works?” and “The Hong Kong Ordinance on Copyright was substantially revised in which year?” General facts and rules are the “more global aspects of theory” ([8], p. 398). Rules are general facts in this sense as they are set procedures that are true in a wider context. Questions that fall into this category include: “Printing out any records or articles from the electronic resources subscribed by the Library will infringe the copyright; True/ False?” and “Copying by a person for research is fair dealing if the copying will result in copies of the same material being provided to more than one person at the same time and for the same purpose; True/ False”. Concepts are explanations and definitions of theories and ideas, and clarifications of the linkages between these theories and ideas. They are “highly familiar, generalized knowledge which students tend to simply know” ([10], p. 78). Test questions in the course concerning concepts include “Which of the following actions is regarded as plagiarism?” and “Leaving out some words in a quoted passage without any indication is plagiarism; True/ False?” Lastly, there are questions that required application of knowledge, and students were asked to make decisions based on theories and concepts learnt in highly-specific situations. For example, there are questions “You want to set up a factory in Shenzhen to make a black and gold-coloured pen, and you want to call the pen a ‘MAN BLANK’. Which of the following would you need to check?” and “You want to use a photograph of a painting by Leonardo Da Vinci (1452–1519) in your dissertation. Who owns the copyright?” Table 2 illustrates the relationships between the content themes of the questions and their respective knowledge levels. Table 2. Categorization of the exam questions Knowledge levels Specific facts General facts & rules Concepts Applied knowledge
Content themes Copyright around the world Copyright in HK Permitted act for research & private study in HK Avoiding plagiarism Intellectual property & copyright
4.4 Evaluation Strategies CUHK is a strongly face-to-face university in its teaching style and e-learning is not used extensively [19]. It is therefore especially important to evaluate innovations, especially in courses that are conducted totally online. We devised an evaluation plan which is composed of multiple evaluation instruments. The evaluation questions that interested the course organizers include: accessibility – whether students can readily access the course; learning – whether students can learn the concepts of the course effectively through online means; and retention of learning – whether the learning is retained. Concerning accessibility, the research team kept detailed records on the access and activity logs of the students’ visits to the various pages on the site and their attempts at the tests. We will illustrate this aspect by quoting the logs kept in the eight cohorts across two academic years (2004–05 and 2005–06). Regarding students’ learning, the data came from students’ test scores and their opinions elicited through surveys conducted in the same eight cohorts in the academic years 2004–05, and 2005–06. The surveys collected students’ feedback on how much they valued the course, and how much they thought they learnt from the course. Lastly, regarding retention of knowledge, two attempts to invite students to take retests were carried out. During the 2004–05 academic year, the first trial of this study was carried out. The retest was launched in June 2005 for both students in Groups 1 and 2 in the 2004–05 cohort. Group 1 students originally took the online course test in October 2004 and the original test period of the Group 2 students was December 2004. Therefore, there was a time gap of six to eight months between the first time the students did the test and the retest. The content of retest was the same as the original examination, and consisted of 20 multiple choice questions randomly selected from a pool of 29 questions. The retest received a relatively low completion rate in the first trial: 16.5% (52 did the retest out of the 315 students who were in either Group 1 or 2 and had taken the original test). Thus, in order to boost the response rate, a lucky draw prize ($HK500 – ~Euro51 – book coupon) was offered in the second trial in the 2005–2006 academic year. The second study was launched in March–April 2006. This time we invited students in Groups 1 and 2 of the 2005–06 cohort to take the retest. The original test period of the 2006 Group 1 was 3–28 October 2005 and that of the Group 2 students was 21 November–16 December 2005. Thus, the time gap between the exam and the retest ranged from three to six months. The retest invitation was sent to those Group 1 and Group 2 students who had taken the course test. No retest invitation was sent to those who did not take part in the examination. The number of students who received the invitation of retest was 387. Reminders were sent twice. At the end, there were a total of 148 retest participants. The completion rate for the second trial is 38.2% (148/387).
5 Findings The online course was readily accessed by students. For example, in the academic year 2005–06, the 571 students who took the course and finished the online test had visited the site (recorded by the counter on the first page of the site) a total of 5,786
Evaluating an Online Module on Copyright Law and Intellectual Property
299
times, meaning that each student on average accessed the site 10.1 times. The counters on the 28 course content pages, on the other hand, recorded a total of 119,034 visits. Thus, on average, each student accessed these pages 208.5 times to prepare for the course-end test. Most students who registered the course actually finished it. A total of 1,278 students registered for the course in all the eight cohorts, and among them 1,134 successfully completed the course-end test. The completion rate was 88.7%. Overall, students answered 17.8 questions correctly out of the 20 attempted questions, a percentage score of 88.9%. A total of 1,120 students answered the opinion survey attached with the course-end test (out of the 1,134 students who completed the course; response rate being 98.8%). The students were assured that their feedback on the survey would not in any manner affect their scores on the test. The survey in general affirmed that the course was overwhelmingly welcomed by students. For example, the average score on the question “The modules achieved the stated objectives” was 4.0 in a 5-point Likert scale in which 1 stands for strongly disagree and 5 means strongly agree. This is very high for a compulsory module. The following sections explore the performance of a subset of the students (the 200 students who completed both the test and retest in our two study trials) in their learning and retention of the knowledge acquired in the course. 5.1 Learning of Knowledge The learning outcomes of the students can be gauged by the performance of the students in their original course-end tests. The 200 students performed very well in the original test, achieving a percentage score of 93.4% among the questions they attempted. Their scores of each of the knowledge levels were slightly different, though still very high in general. They scored, on average, 91.3% correct in questions that were about specific facts, 93.7% in the questions about general facts and rules, 97.4% in questions on concepts, and 93.0% in questions about applied knowledge. The distribution of the marks is illustrated in Figure 2. It is also noted that the performance of the 52 students in the first 2004–05 study trial in general showed the same pattern as that of the 148 students in the second 2005–06 trial. One-way ANOVA found that the between-group differences were statistically significant at the 0.01 level. Post-hoc Scheffe tests were then carried out which established that the main difference was from the exceedingly high marks on the concepts category. The differences between students’ performances on questions related to concepts and those in questions related to other knowledge domains were all statistically significant at the 0.05 level. 5.2 Retention of Knowledge Retention of knowledge was investigated by comparing the 200 students’ performances in their original tests and re-tests. Paired-sample t-tests were used to test for any differences between the mean scores of the examination and the retest. Although the first trial of the study in 2004–05 had a much lower response rate than the second test–retest study in 2005–06, the two set of results were actually very similar.
300
C. McNaught et al.
100%
95% Overall 90%
04-05 05-06
85%
80% Specific facts General facts/ rules
Concepts
Application of concepts
Fig. 2. Performance in different knowledge levels
In 2004–05, students scored on average 94.4% in their original test while they scored 78.0% in their postponed retest. In 2005–06, the scores were 93.0% and 77.9% respectively. Overall, the 200 students scored 93.4% and 77.9% in their original tests and retests. The result from the paired-sample t-test revealed that the differences between these original test scores and retest scores are statistically significant at the 0.01 level. It is worthwhile to note that although the students’ performance in the retest declined significantly; nevertheless, their performances were quite reasonable, with an average percentage score of 77.9%. A closer look at the data on the various knowledge levels revealed the patterns portrayed in Table 3 and Figure 3. Table 3. Retention by knowledge domains Knowledge domains Specific facts
Exam
Retest
Diff.
91.3%
52.7%
38.6%
General facts and rules
93.7%
81.2%
12.5%
Concepts
97.4%
87.1%
10.3%
Applied knowledge
93.0%
82.5%
10.5%
The data show the sharpest decline in performance on question items that relate to specific facts when compared with the other knowledge domains. This was a 38.6% drop (91.3% to 52.7%) while the declines in performances in the other three questions levels were only 12.5%, 10.3% and 10.5%, respectively. This represents more than three times the percentage change when compared with the other changes.
6 Discussion The online module appears to be an effective learning tool. The scores on the original tests were all very high, showing that e-learning is good not only in delivery of facts, but also in explaining concepts (in fact, students’ scores on the questions related to concepts were the best), and teaching applied knowledge.
Evaluating an Online Module on Copyright Law and Intellectual Property
301
100% Concepts Percentage correct
90% Application of concepts
80%
Principles / rules
70% 60%
Specific facts
50% Exam
Retes t
Fig. 3. Decline in performance by knowledge domains
Students performed slightly worse in the retests than in the first tests. Compared with the very high scores in the original test, students’ scores in the retests were clearly poorer. This drop in scores, however, is quite expected as time is always regarded as affecting retention of knowledge. In fact, students still managed to achieve relatively good performance in the retests and this shows that e-learning can have extended effects on students’ learning, contrary perhaps, to the observations of Bell, Fonarow, Hays and Mangione (2000) [12], who found material learnt on computer is not retained; but more or less in line with the position of Yildirim, Ozden and Aksu (2001) [11] that e-learning can produce long-term learning. While this small study in no way solves the ambiguity in the research literature, it does contribute to our understanding. While students generally found all categories easy (above 85% of the answers in all categories were correct), they found one category increasingly more difficult as time passed. This is the category of specific facts. Students differed in their retention of different forms of knowledge. Knowledge of specific facts tended to drop to a far greater extent than learnt knowledge in the other domains. The decrease of scores in this category significantly outnumbered those in the other categories, dropping more than 35% while the other declines were in the 10% level. Unrelated facts are difficult to remember in traditional classroom teaching [8] and we now have evidence that, although e-learning can be used to disseminate facts, facts learnt in ‘e-classrooms’ are not retained over time. There is thus a resemblance between knowledge retention in the two learning environments. Education is concerned with the development of the higher cognitive reasoning skills rather than memorization of facts and unrelated concepts. The findings of this study seem to support the role of web-assisted teaching as not being limited to delivery of isolated facts and information. The web can be effective in facilitating learning at higher levels. Knowledge of isolated specific facts is not retained while acquired knowledge concerning more general rules and concepts, and their applications, appears to be more worthwhile as the focus of online materials. The findings of the study provided timely feedback to the development team about which questions in the module to consider for revision and how we might refocus some of the information in the module. There have thus been tangible benefits from the study.
302
C. McNaught et al.
The present study has clear limitations. First, the delayed retests took place after a relatively short period of time (three to six months) and so the retention pattern of these various forms of knowledge in a more extended period of time is largely unknown. Nevertheless, many previous studies have shown that the period immediately after the learning activity is actually the most critical as this is when the decline in knowledge retained is most serious [20], [21], [8]. Second, we are aware of the fact that many factors, such as individual differences, prior knowledge of learners, content organization and structure, etc., affect learning and memory retention [22], [7]. The present study on a single online module utilizing one specific way of content design is far from being able to make any general claims about retention of knowledge forms in e-medium learning environments.
7 Conclusions The study confirms that e-learning can be an effective tool not only in the dissemination of facts, but can also effectively explain concepts and assist students in applying knowledge. Specific factual knowledge is hard to retain. The findings of this study suggest that the role of e-learning, just as the role of traditional teaching, should focus on concepts and applied knowledge rather than on memorization of facts alone. The data from this study come from one course alone and so must be treated as indicative. We are continuing to collect evaluation data to enable continued improvement of this module. However, additional further studies in a range of discipline areas are warranted.
References 1. Bloom, B.S.(ed.) Taxonomy of Educational Objectives: The Classification of Educational Goals: Handbook I, Cognitive domain, New York, Longman (1956) 2. Anderson, L.W., Krathwohl, D.R.: A Taxonomy for Learning, Teaching, and Assessing: A Revision of Bloom’s Taxonomy of Educational Objectives. Boston, Allyn & Bacon (2001) 3. Krathwohl, D.R.: A Revision of Blooms’ Taxonomy: An Overview. Theory into Practice 41(4), 212–218 (2002) 4. Conway, M.A., Gardiner, J.M., Perfect, T.J., Anderson, S.J., Cohen, G.M.: Changes in Memory Awareness during Learning: the Acquisition of Knowledge by Psychology Undergraduates. Journal of Experimental Psychology: General 126, 393–413 (1997) 5. Novak, J.D., Gowin, D.B.: Learning How to Learn. Cambridge University Press, Cambridge (1984) 6. Herbert, D.M.B., Burt, J.S.: Memory Awareness and Schematization: Learning in the University Context. Applied Cognitive Psychology 15(6), 613–637 (2001) 7. Semb, G.B., Ellis, J.A., Araujo, J.: Long-term Memory for Knowledge Learned in School. Journal of Educational Psychology 82(2), 305–316 (1993) 8. Conway, M.A., Cohen, G., Stanhope, N.: On the Very Long-term Retention of Knowledge Acquired through Formal Education: Twelve Years of Cognitive Psychology. Journal of Experimental Psychology: General 120(4), 395–409 (1991) 9. Neisser, U.: Interpreting Harry Bahrick’s Discovery: What Confers Immunity against Forgetting? Journal of Experimental Psychology: General 113, 32–35 (1984)
Evaluating an Online Module on Copyright Law and Intellectual Property
303
10. Herbert, D.M.B., Burt, J.S.: What do Students Remember? Episodic Memory and the Development of Schematization. Applied Cognitive Psychology 18(1), 77–88 (2004) 11. Yildirim, Z., Ozden, M.Y., Aksu, M.: Comparison of Hypermedia Learning and Traditional Instruction on Knowledge Acquisition and Retention. The Journal of Educational Research 94(4), 207–214 (2001) 12. Bell, D.S., Fonarow, G.C., Hays, R.D., Mangione, C.M.: Self-study from Web-based and Printed Guideline Materials: A Randomized, Controlled Trial among Resident Physicians. Annals of Internal Medicine 132(12), 938–946 (2000) 13. Reeves, T.C., Hedberg, J.G.: Interactive Learning Systems Evaluation. Educational Technology Publications, Englewood Cliffs, New Jersey (2003) 14. Van-Draska, M.S.: Copyright in the Digital Classroom. Journal of Allied Health 32(3), 185–188 (2003) 15. Gould, T.H.P., Lipinski, T.A., Buchanan, E.A.: Copyright Policies and the Deciphering of Fair Use in the Creation of Reserves at University Libraries. The Journal of Academic Librarianship 31(3), 182–197 (2005) 16. Rockman, H.B.: An Internet Delivered Course: Intellectual Property Law for Engineers and Scientist. Prontiers in Education 34, S1B22–S1B27 (2004) 17. DeRouin, R.E., Fritzsche, B.A., Salas, E.: Optimizing e-Learning: Research-based Guidelines for Learner-controlled Training. Human Resource Management 53(2–3), 147–162 (2004) 18. Iverson, K.M., Colky, D.L., Cyboran, V.: E-learning Takes the Lead: An Empirical Investigation of Learner Differences in Online and Classroom Delivery. Performance Improvement Quarterly 18(4), 5–18 (2005) 19. McNaught, C., Lam, P., Keing, C., Cheng, K.F.: Improving eLearning Support and Infrastructure: An Evidence-based Approach. In: O’Donoghue, J. (ed.) Technology supported learning and teaching: A staff perspective, Hershey, PA, pp. 70–89. Information Science Publishing (2006) 20. Bahrick, H.P.: Semantic Memory Content in Permastore: Fifty Years of Memory for Spanish Learned in School. Journal of Experimental Psychology 113(1), 1–29 (1984) 21. Bahrick, H.P., Hall, L.K.: Lifetime Maintenance of High School Mathematics Content. Journal of Experimental Psychology: General 120(1), 20–33 (1991) 22. Semb, G.B., Ellis, J.A.: Knowledge Taught in School: What is Remembered? Review of Educational Research 64(2), 253–286 (1994)
Using Games-Based Learning to Teach Software Engineering Thomas M. Connolly, Mark Stansfield, and Tom Hainey School of Computing, University of the West of Scotland, High St, Paisley, PA1 2BE, Scotland {thomas.connolly,mark.stansfield,hain-ci0}@uws.ac.uk
Abstract. For some time now, computer games have played an important role in both children and adults’ leisure activities. While there has been much written on the negative aspects of computer games, it has also been recognised that they have potential advantages and benefits. There is no doubt that computer games can be highly engaging and incorporate features that are extremely compelling. It is these highly engaging features of computer games that have attracted the interests of educationalists. The use of games-based learning has been growing for some years now, however, within software engineering there is still a dearth of empirical evidence to support this approach. In this paper, we examine the literature on the use of computer games to teach software engineering concepts and describe a computer game we have been developing to teach these concepts. Keywords: Games-based learning, software engineering, collection and analysis, project management, team leading.
requirements
1 Introduction It is generally accepted that computer games are an extremely motivating and engaging medium and represent a new form of popular culture. There is also a growing recognition of the potential benefits that can be gained in using computer games within teaching and learning, although there are still many critics of this approach. We have been attempting to use a games-based learning approach to support the teaching of requirements collection and analysis within a software engineering course in Higher Education for some years now [1]. Recently we have teamed up with a Scottish games-based learning company to develop an expanded version of the game for both the academic and training communities. In this paper, we examine the literature on the use of games-based learning within software engineering and also examine some of the issues underlying the teaching of the abstract and complex domain of requirements collection and analysis and, more generally, software engineering. We then discuss the high-level requirements for our game and provide an overview of the game play and an outline subsystem design.
Using Games-Based Learning to Teach Software Engineering
305
difficult to recognize as such because of complex interdependencies [2]. According to Armarego [3], there is an educational dilemma in teaching such problems in software engineering because: • • • •
complexity is added rather than reduced with increased understanding of the problem; metacognitive strategies are fundamental to the process; a rich background of knowledge and intuition are needed for effective problemsolving; a breadth of experience is necessary so that similarities and differences with past strategies are used to deal with new situations.
Oh and Van der Hoek identify a number of other issues that complicate the teaching of the software process [4]: •
•
•
• •
Software Development is Non-linear: activities, tasks and phases are repeated and multiple events happen at the same time. Managing two similar projects in the same way may not produce the same outcome due to the presence of several (possibly unexpected) factors (e.g., technical advances, client behaviours or expectations). Software Development Involves Several Intermediate Steps and Continuous Choices between Multiple, viable alternatives: even with careful planning, not all events that can occur can be anticipated at the start of a project. Difficult decisions must be made, tradeoffs considered and conflicts handled. Software Development may Exhibit Dramatic Effects with Non-obvious Causes: while software development has several cause-and-effect relationships (e.g., it is more cost-effective to identify flaws in the earlier phases of development than to identify them in the later phases), there are other situations that may arise in which the cause is not so apparent. For example, Brook’s Law states that adding people to a project that is already late typically makes that project later. Software Engineering Involves Multiple Stakeholders: clients and nondevelopment personnel in an organization all make decisions that impact development. Software Engineering often has Multiple, Conflicting Goals: software development includes tradeoffs between such things as quality versus cost, timeliness versus thoroughness, or reliability versus performance.
Two further issues arise with teaching software development that we are interested in taking into consideration in any learning environment we develop are: •
•
Communication: software engineers must be able to communicate, both verbally and in writing, with staff internal to the project (project manager, team leaders, analysts, designers, developers, testers, quality assurance) as well as with external stakeholders. Pedagogical Praxis: Shaffer [5] proposes a theory of ‘pedagogical praxis’, which links learning and doing within an extended framework of communities of practice ([6], [7]). Pedagogical praxis is based on the concept that different professions (for example, lawyers, doctors, software engineers) have different epistemologies (epistemic frames) – different ways of knowing, of deciding what
306
T.M. Connolly, M. Stansfield, and T. Hainey
is worth knowing and of adding to the collective body of knowledge and understanding. For a particular community, the epistemic frames define “knowing where to begin looking and asking questions, knowing what constitutes appropriate evidence to consider or information to assess, knowing how to go about gathering that evidence, and knowing when to draw a conclusion and/or move on to a different issue” [8, pp. 4]. Implementation of pedagogical praxis requires a faithful recreation of the professional community, one that is “thickly authentic”; that is, one where (a) learning is personally meaningful for the learner, (b) learning relates to the real-world outside the classroom, (c) learning provides an opportunity to think in the modes of a particular profession and (d) learning where the means of assessment reflect the learning process [9]. Connolly and Begg have suggested that the term thickly authentic be extended to incorporate: (e) learning using the tools and practices of the modern-day professional [10]. According to Schön [11, 12] the following are some of the key problems in teaching an abstract subject of this nature: • • • •
It is learnable but not didactically or discursively teachable: it can be learned only in and through practical operations. It is a holistic skill and parts cannot be learned in isolation but by experiencing it in action. It depends upon the ability to recognize desirable and undesirable qualities of the discovered world. However, this recognition is not something that can be described to learners, instead it must be learned by doing. It is a creative process in which a designer comes to see and do things in new ways. Therefore, no prior description of it can take the place of learning by doing.
Students often have considerable difficulty comprehending implementationindependent issues and analyzing problems where there is no single, simple, wellknown, or correct solution [10]. They have difficulty handling ambiguity and vagueness and they can also display an inability to translate tutorial examples to other domains with analogous scenarios, betraying a lack of transferable analytical and problem-solving skills [13]. Kriz [14] highlights the point that the majority of students are not competent enough to put their knowledge into practice and they are unable to cope successfully with the everyday tasks associated with the practice of their chosen field. These problems can lead to confusion, a lack of self-confidence and a lack of motivation to continue. Many of the above characteristics make teaching requirements collection and analysis and, more generally, the software development process, problematic using didactic approaches to teaching and learning and the practical experience provided falls far short of what a student can expect “in the real world”. Instead, these issues suggest that students can only learn about software engineering by doing software engineering and rely less on overt lecturing and traditional teaching. This approach requires a shift in the roles of both students and teachers, with the student becoming an apprentice, exploring and learning about the problem in the presence of peers (who may know more or less about the topic at hand) and the teacher moving from being the “knowledgeable other” towards becoming a facilitator, who manages the context
Using Games-Based Learning to Teach Software Engineering
307
and setting, and assists students in developing an understanding of the material at hand [15]. We advocate an alternative teaching paradigm for software engineering based on constructivism. Cognitive constructivism views learning as an active process in which learners construct new ideas or concepts based upon their current/past knowledge. The learner selects and transforms information, constructs hypotheses and makes decisions, relying on a cognitive structure to do so [16]. Social constructivism, seen as a variant of cognitive constructivism, emphasizes that human intelligence originates in our culture. Individual cognitive gain occurs first in interaction with other people and in the next phase within the individual [17]. These two models are not mutually exclusive but merely focus upon different aspects of the learning process. In fact, Illeris [18] believes that all learning includes three dimensions, namely, the cognitive dimension of knowledge and skills, the emotional dimension of feelings and motivation, and the social dimension of communication and cooperation – “all of which are embedded in a societally situated context”. Figure 1 provides a representation of the environment we will use as the basis for the development of the games-based learning application and the problems to be addressed.
Fig. 1. Environment for teaching software engineering (and problems that need to be addressed)
3 The Software Engineering Game The development of the game is being underpinned by Participatory Design principles with users and other stakeholders playing a prominent role in all the stages relating to design, development and evaluation. The benefits of Participative Design are that it can provide better project control, better communication, more satisfied users and
308
T.M. Connolly, M. Stansfield, and T. Hainey
participants, lessens the need for costly corrective action post implementation and can provide more innovative and creative solutions than might have otherwise been possible [19, 20]. To support the design, development and evaluation of the game, it was decided to establish a steering committee comprising senior representatives from industry and commerce, a number of academic representatives, the project managers and the developers of the game. By drawing upon the expertise and views of senior managers from industry and commerce it is hoped that the game will have a greater degree of relevance and significance to a wider audience other than students in higher education. In addition, it is hoped the game will utilise and develop a wider range of skills and knowledge that might be transferable across a wider section of industry and commerce. We now discuss the high-level objectives of the game, the game play and then provide an outline design of the game itself. 3.1 High-Level Objectives The games-based learning environment should provide a rich learning experience through the creation of a range of project scenarios that will: • • • • • • •
Promote an engineering ethos that emphasizes fitness for purpose as the guiding principle in the design, development and assessment of information systems and their components. Enable the learner to take a disciplined approach to requirements collection and analysis, and to the high level specification, design and implementation of information systems and their components. Enable the learner to handle complexity, vagueness and ambiguity during the project. Enable the learner to develop a range of project management skills. Assist the learner to develop analytical and problem-solving skills and transferable skills. Assist the learner to develop the skills required for both autonomous practice and team-working. Assist the learner to develop reflection and metacognitive strategies.
In discussion with the advisory group, the following requirements were identified: • • •
• •
The game will be targeted at both university students in a computing-related subject and also the professional training market. The game must support a number of players carrying out different roles (for example, analyst, developer, project manager) as well as a facilitator. Communication between players should be supported. The facilitator will be able to see what the players are doing, will be able to intervene in the game (for example, to modify the frequency of new projects, to modify the number of people assigned to a project) and will be able to call team meetings to discuss issues that have arisen in the team’s play. Ideally, in a team-based activity when a player is not available the game (AI) should play that role. The game must be scenario-based to allow the players access to a range of project scenarios to provide practical experience.
Using Games-Based Learning to Teach Software Engineering
•
•
309
The game must have a reasonably authentic underlying business model to model clients, projects, staff, suppliers and competitors. The model should take cognisance of a range of project variables such as project budget, time, staff, staff specialisations, staff costs, resource costs. These variables would be scenariospecific. The game should run in an online environment.
Game play should be recorded wherever possible to support debriefing, post-game analysis and evaluation.
3.2 Game Play The basic idea of the game is for the team (comprising one or more players) to manage and deliver a number of software development projects. Each player has a specific role, such as project manager, systems analyst, systems designer or team leader. A bank of scenarios have been created based on case studies the authors have been using for many years in teaching and learning; for example, the DreamHome Estate Agency [21], the StayHome Online DVD Rentals company and the Perfect Pets Veterinary Clinic [22], the Blackwood Library and the Fair Winds Marina [1]. Each scenario has an underlying business model; for example, there will be a budget for the delivery of the project, a set timescale for the delivery of the project and a set of resources (for example, staff with specified technical specilisations) that can be used on the project. Additional resources can be brought in for a project although this will have a cost and timescale (delay) associated with it. The project manager has overall responsibility for the delivery of each project on budget and on time and is given a short brief for each project. Communication is one of the key aspects of the game and the project manager must communicate relevant details of the project to the other players. This will be done using a message metaphor – any player can communicate with any other player(s) by sending a message. Players have a message board that indicates whether there are any unread messages. The player(s) assigned to the system analyst role has to identify the requirements for the project. To do this, the player must move through the game and ‘talk’ to the nonplayer characters (NPCs) in the game, as illustrated in Figure 2. In addition, there are objects in the game that can also convey relevant information when found (for example, a filing cabinet may convey requirements). For the prototype game we are using written transcripts in place of NPC speech. We hope shortly to use lip synching within the game to have the NPCs ‘talk’ to the system analyst. Each NPC’s ‘speech’ will contain some general background details and a number of requirements (the analyst has to distinguish the requirements from the general details). Visiting the same NPC may generate the same speech or a new speech. Each speech will generate a transcript that the analyst can visit at any point in the game. The transcript is presented as a numbered list of requirements. During the play, the analyst can use the transcripts to produce an initial ‘wishlist’ of requirements, which can be refined until such time as the analyst believes all requirements have been identified, at which point the analyst can send the completed requirements to the project manager. The project manager now has two choices: send the requirements to the designer to produce an outline high-level design or consider the requirements to be incorrect and ask the analyst to rework the requirements (asking for rework will have a ‘cost’ associated with it).
310
T.M. Connolly, M. Stansfield, and T. Hainey
During this period, the designer will be provided with some background information relevant to the design phase (for example, high-level components that the company might have developed previously, technical experience of the staff, technical resources the designer has access to and software and hardware that can be bought externally). Upon receiving the requirements, the designer must produce a high-level design that addresses the clients’ requirements and must identify what will be developed ‘in-house’ and what software/hardware will be bought in. In addition, the designer must provide some estimate of cost and timescale to implement the system. Again, the design will go back to the project manager to accept or reject (in which case the design must be reworked by the designer at ‘cost’). The implementation phase is handled by the team leader who is given a brief by the project manager (high level design, available budget, available staff), as illustrated in Figure 3. The team leader is responsible for the delivery of the implementation phase. However, during this period the team leader may have to handle a number of planned events (such as staff holidays) and unexpected events (such as staff becoming ill, leave, and some activities taking longer than planned). Some events the team leader may be able to handle autonomously within the remit provided; however, with others the team leader may need to consult the project manager to seek a solution.
Fig. 2. Screen during requirements collection
The facilitator will have access to the game play and will be able to intervene during the play. One intervention is to call a (physical or virtual) team meeting because of problems identified with the running of the project. There are a number of other interventions, such as changing the requirements during the design or implementation phase, reducing the number of staff available for the project, making staff go off sick, making staff leave the company.
Using Games-Based Learning to Teach Software Engineering
311
Fig. 3. Screen showing team leader assigning project tasks
Fig. 4. Environment for teaching software engineering (and problems that need to be addressed)
3.3 Game Design The game is based on the traditional multi-client/single-server architecture. The subsystem design is shown in Figure 4: • •
The Scenario Builder is an offline utility to allow us to create and update the game scenarios. These scenarios are stored in the server-side database. The Postgame Analysis is a second offline utility to allow us to provide data to the facilitator on how the team has performed. The utility will also provide us with data to evaluate the impact of the game and to eventually produce longitudinal analyses.
312
•
T.M. Connolly, M. Stansfield, and T. Hainey
The Game Server consists of four main subsystems: The Comms subsystem, which allows players to communicate with each other. The Visualisation/User Interface, which handles what the players see on the screen and what they can do. The Business Model, which implements both the general business rules and the business rules specific to each scenario. This will be loosely based on the SESAM model [23]. The Game AI (Artificial Intelligence), which implements ‘missing players’. At the time of writing, this subsystem has not been fully designed.
4 Conclusions In this paper we have examined previous approaches to the application of games-based learning to software engineering and have found a significant dearth ofempirical research to support this approach. Software engineering has been described as a “wicked problem”, characterized by incomplete, contradictory and changing requirements and solutions that are often difficult to recognize as such because of complex interdependencies. Other issues that complicate the teaching of software engineering are that software development is non-linear, it involves several intermediate steps and choices between multiple, viable alternatives, it may exhibit dramatic effects with non-obvious causes and it involves multiple stakeholders. Finally, we have described the design of a new gamesbased learning application aimed at the teaching of requirements collection and analysis, design and project management aimed at both the academic and training markets. We consider evaluation to be key to the entire development process and have adopted a Participatory Design approach from the outset. In the design of the game we have included a Postgame Analysis utility to support the collection of empirical evidence on the use of this game. The input to date from the advisory group has been extremely useful and has helped shape the design of the game. We have developed a prototype of the game that we are currently using to explore various design ideas and hope to have an initial version of the full system available for evaluation in early 2008. As well as developing the game, we have examined approaches to evaluating games-based learning, particularly within software engineering and have found a dearth of in-depth studies [24]. We are currently working an producing a framework that will allow us to evaluate a range of both hard and soft skills in our software engineering game.
References 1. Connolly, T.M., McLellan, E., Stansfield, M.H., Ramsay, J., Sutherland, J.: Applying Computer Games Concepts to Teaching Database Analysis and Design. In: International Conference on Computer Games, AI, Design and Education, Reading, UK (November 2004) 2. DeGrace, P., Hulet Stahl, L.: Wicked Problems, Righteous Solutions: A Catalog of Modern Engineering Paradigms. Prentice Hall, Englewood Cliffs (1998) 3. Armarego, J.: Advanced Software Design: A Case in Problem-Based Learning. In: Proceedings of the 15th Conference on Software Engineering Education and Training, Covington, Kentucky, USA, February 25–27, 2002, pp. 44–54 (2002)
Using Games-Based Learning to Teach Software Engineering
313
4. Oh, E., Van der Hoek, A.: Adapting Game Technology to Support Individual and Organizational Learning. In: Proceedings of the 13th International Conference on Software Engineering and Knowledge Engineering, Buenos Aires, Argentina (2001) 5. Shaffer, D.W.: Pedagogical Praxis: The Professions as Models for Postindustrial Education. The Teachers College Record 106(7), 1401–1421 (2004) 6. Lave, J.: Situating learning in communities of practice. American Psychological Association, Washington (1991) 7. Lave, J., Wenger, E.: Situated learning: Legitimate peripheral participation. Cambridge University Press, Cambridge (1991) 8. Shaffer, D.W.: Epistemic frames and islands of expertise: Learning from infusion experiences. In: Proceedings International Conference of the Learning Sciences, Santa Monica, CA. Retrieved July 28, 2005, from http://www.education.wisc.edu/ edpsych/facstaff 9. Shaffer, D.W., Resnick, M.: Thick authenticity: New media and authentic learning. J. of Interactive Learning Research 10(2), 195–215 (1999) 10. Connolly, T.M., Begg, C.E.: A constructivist-based approach to teaching database analysis and design. J. of Information Systems Education 17(1), 43–54 (2006) 11. Schön, D.A.: The Reflective Practitioner: How Professionals Think in Action. Basic Books, New York (1983) 12. Schön, D.A.: Educating the Reflective Practitioner: Towards a New Design for Teaching in the Professions. Jossey-Bass Inc., San Fransisco (1987) 13. Connolly, T.M., Stansfield, M.H.: From eLearning to games-based eLearning: using interactive technologies in teaching an IS course. International Journal of Information Technology and Management (2007) 14. Kriz, W.C.: Creating effective learning environments and learning organizations through gaming simulation design. Simulation and Gaming 34(4), 495–511 (2003) 15. Koehler, M.J., Mishra, P.: Teachers Learning Technology by Design. J. of Computing in Teacher Education 21(3) (2005) 16. Piaget, J.: Six Psychological Studies. Vintage Books, New York (1968) 17. Forman, E., McPhail, J.: Vygotskian perspectives on children’s collaborative problemsolving activities. In: Forman, E.A., Minick, N., Stone, C.A. (eds.) Contexts for learning. Sociocultural dynamics in children’s development, Oxford University Press, Oxford (1993) 18. Illeris, K.: Towards a contemporary and comprehensive theory of learning. International J. of Lifelong Learning 22(4), 396–406 (2003) 19. Kensing, F., Blomberg, J.: Participatory design: issues and concerns. Computer Supported Cooperative Work 7, 167–185 (1998) 20. Cherry, C., Macredie, R.D.: The importance of context within information system design: an assessment of participative design. Requirements Engineering 4, 103–114 (1999) 21. Connolly, T.M., Begg, C.E.: Database Systems: A practical approach to design, implementation, and management, 4th edn. Addison Wesley Longman, England (2005) 22. Connolly, T.M., Begg, C.E.: Database Solutions: A step-by-step approach to building databases, 2nd edn. Addison Wesley Longman, England (2002) 23. Mandl-Striegnitz, P.: How to Successfully Use Software Project Simulation for Educating Software Project Managers. In: Proceedings of the 31st Frontiers in Education Conference, Nevada, USA (2001) 24. Connolly, T.M., Stansfield, M.H., Hainey, T.: An Application of Games-based Learning within Software Engineering. British J. of Educational Technology 38(3), 416–418 (2007)
Agility in Serious Games Development with Distributed Teams: A Case Study Manuel Oliveira1 and Heiko Duin2 1
Abstract. This paper describes as a case study the first 12 months of a 2 year project developing a serious game targeted at a number of industrial sectors: aeronautical, automotive, civil construction, software and electronics. The paper presents the devised methodology to address the problems that emerged, mostly associated to the inherent barriers of managing a distributed team with equal participatory roles and responsibilities in the creative process of developing a serious game. Some of the lessons learnt are shared with the reader. Keywords: Distributed brainstorming, conceptual design, serious games, distributed teams.
Agility in Serious Games Development with Distributed Teams: A Case Study
315
from the Agile Programming community, but focused on the end-user from conception to final deployment. 1.1 The PRIME Project The case study presented in this paper is based in the European project PRIME (Providing Real Integration in Multi-disciplinary Environments). The main objective of the PRIME project is to give business professionals a learning environment where they can experiment with new ideas and learn how to handle the entire life cycle of products/processes. PRIME proposes to achieve this by enhancing current work environments with a new paradigm based on serious gaming [2]. Serious games are digital games used as a per-suasion or educational technology. They can be similar to educational games, but are primarily in-tended for an audience outside of primary or secondary education. Serious games might be of any genre and many of them represent a kind of edutainment applications. A serious game could be a simulation which has the look and feel of a game, but corresponds to non-game events or processes, such as business or military operations (e.g. America’s Army or Tactical Iraqi [23]). In PRIME, the adoption of serious game approach allows the user to learn by experience within a Virtual Business Environment (VBE) that is safe and foments risk taking without detrimental impact on business. The context is based on strategic management, including multi-stakeholder negotiation and business connectivity, with a strong focus on manufacturing. The VBE is composed of several systems defining its economic fabric: capital markets, population, governments, logistic support, labor markets, business-toconsumer, consumer associations, labor unions, etc. Within this rich environment, the business professional assumes the role of a Business Unit (BU), which may correspond to either a single site or multiple sites. The management breakdown of a BU is based on eight functional units: Production, Product Development, Sales, Human Resources, Strategic Marketing, Distribution, Finance and Information Systems. Associated to each functional unit is a set of operation processes that are based on an input-transformation-output model. The business professional, as a human player, makes strategic decisions based on information available through the BU and the VBE. However, the decisions themselves are not implemented by the PRIME Serious Game, although the interface and game mechanisms may facilitate the decision processes [14]. The same is not true of the artificial stakeholders, who share similar decision processes, but are required to have them implemented as a decision engine supported by a knowledge rule based database. A more detailed description of the PRIME project is given in [18]. 1.2 Stakeholders The project has a developer team of seven partners (Alfamicro, BIBA, CIST-Sofia, EPFL, Intrapoint, MIP and Sintef), who are geographically distributed across Europe accompanied by six end-user partners from different business sectors (CRF, IAI, Intracom, KESZ, LEGO and Siemens).
316
M. Oliveira and H. Duin
Fig. 1. The PRIME consortium
The need for a geographically distributed team exists from the onset of the project, as evidenced by map illustrated in Fig. 1. Achieving a cohesive multidisciplinary team with the ability of working together irrespective of the cultural differences, work practices, and geographic distribution across 3 different time zones, proved to be a challenge in process engineering and not coupled to technology. The diversity of the end-user organizations from different industrial sectors raises additional challenges to the development process. In addition, each developer organization also had particular expectations for the outcome of the project, which would ultimately influence the development process. Table 1. PRIME stakeholders description Acronym
Longname
Country
Role
Expectations
Alfamicro
Alfamicro
Portugal
Developer
Consultancy company will use PRIME outputs in its consultancy services.
BIBA
Bremen Institute of Industrial Engineering and Applied Work Science
Germany
Developer
Research institute that will use PRIME as a supporting teaching tool in the Bremen university management courses.
CIST
Center of Information Society Technologies, University of Sofia
Bulgaria
Developer
Research centre at the university of Sofia that will use PRIME as a teaching tool in support of their management related courses.
End-User
The research institute for the Fiat group and will use PRIME to train managers in the complexity of managing logistic networks
CRF
Fiat Research Centre
Italy
Agility in Serious Games Development with Distributed Teams: A Case Study
317
Table 1. (continued) Acronym
Longname
Country
Role
Expectations
EPFL
Swiss Federal Institute of Technology
Switzerland
Developer
University aims to exploit PRIME as a teaching tool in support of their management related courses.
End-User
This end-user is involved in the research and manufacturing of military equipment, namely aircraft. The aim is to train managers in their proprietary methodology of product innovation.
IAI
Israel Aircraft Industries
Israel
Intracom
Intracom
Greece
End-User
This end-user is involved in the manufacturing and sales of communication equipment. Their aim is to train managers from product development and marketing to work together to develop new products.
Intrapoint
Intrapoint
Norway
Developer
Software developer that intends to exploit PRIME as a product for training of employees.
KESZ
Central European Building and Construction ltd.
End-User
This end-user is the third largest construction company in Hungary. Their aim is to train managers in project management, handling effectively problems that emerge
Hungary
LEGO
Lego Company
Denmark
End-User
This end-user is a world renowned toy maker and their aim is train managers in the entire process of setting up a product portfolio with all the associated manufacturing decisions, taking into account the market implications to their existing products and their competitors.
MIP
Business School of Politecnico di Milano
Italy
Developer
A top European business school, which intends to use PRIME as a supporting teaching tool in their MBA courses.
End-User
This end-user is a branch of Siemens focusing on software development. Their aim is to train managers in product development taking into account the consumer behavior within the market.
Developer
Largest research institute in Norway and their aim is to use PRIME within the NTNU university and as a consultancy tool with their network of industrial clients.
Siemens
Sintef
Siemens Austria
Sintef Research Institute
Austria
Norway
2 Game Development The process of software development is permeated with creativity (Glass, 1995), and in the case of game development a clear example is embodied in the dynamic
318
M. Oliveira and H. Duin
evolution of a game design document, which final version corresponds to the synergy of all the different stakeholders involved in the development process. The production of games has evolved significantly since the “garage” period when two or three friends would develop a game by doing all the implementation, game design and content with no schedule and minimal budget. Nowadays, a game is usually produced by a multi-disciplinary team of about twenty people or more, with the technical developers being a small core of a few people and the remainder of the team being responsible for content creation and game design. The production cycle usually involves 2 to 2.5 years of desperate development effort chasing ever-shifting deadlines and overcoming problems whilst simultaneously trying to keep to the target release date. The cost involved in maintaining such a large team for an extended amount of time requires an investment budget of a few million Euro. As a result, one verifies in the evidence in many postmortems [9] that it is quite common for projects to exceed the initial budget and overextend the initial time allotted for completion. In many cases, critical problems may be traced to the challenges of managing the creative process of all facets of game development. Although similar to one another, the development of a serious game diverges in many ways from that of a traditional game. In the case of a traditional game, the development process is driven by the vision of a group of people that expect and plan the result to be successful, with a sufficiently high level of market success to justify the development costs. On the other hand, with a serious game, the development process is driven by the user requirements of a well-defined set of users that will and intend to use the resulting serious game. The addition of requirements engineering requires more creativity [22], which according to some [17], adds additional complexity that ultimately may compromise the success of the project. The gaming industry is fully aware of the high risk associated to game development, both traditional genre and serious game. Therefore attempts have been made to improve the development methodology by addressing the challenges raised. A first approach has been to adopt well structured frameworks and methodologies from software engineering community and streamline the development process. There are numerous different development methodologies [25] to choose from, but none is able to guarantee the success of a project. This supports the claim “silver bullets” do not exist [6] to solve the nightmares associated to the software process of complex systems. It is widely accepted that software engineering implies more than sophisticated techniques (e.g. [16]), which is the only way to tackle the complexity of large scale systems that are inherently multi-disciplinary [12]. As a result, there is a growing trend is search alternate development methodologies with some games adopting principles from the agile software development community [5]. Such methods promote a light-weight software process based on iterative short cycles where software engineering is deeply ingrained throughout. The lean development philosophy is achieved by reducing the requirement of supporting documentation and relying on best practices to ultimately increase the productivity. However, an analysis [24] of the foundational principles, along with the necessary assumptions made about the software process, reveals that there continues to be no such thing as “silver bullets”. In addition, the existing agile methods still require some maturity in order to address their limitations [1] namely the disregard to generalization to address the issues of reusability and lack of practical guidance with respect to applicability.
Agility in Serious Games Development with Distributed Teams: A Case Study
319
One of the most popular and predominant agile methods is eXtreme Programming (XP) [5], which proposes twelve practices to replace the traditional development methodologies of the software process. Although the XP methodology is implementation oriented with much diminished emphasis on analysis and design, a small trial study [26] demonstrated the effectiveness of the method in the research arena where the software process is exploratory with an associated high degree of uncertainty to the problem domain. The High Moon Studios, responsible for games such as Dark Watch, have been using SCRUM [7] for project management with agile principles and XP for the actual development [15]. Both agile methodologies have been adapted to fit the nature of the teams involved in the game production. However, none of the methodologies support geographically distributed teams and although the risks of game development are mitigated, the underlying problems persist to plague the projects.
3 PRIME Development Methodology The potential pitfalls associated to game development were known from the onset of the project, and it was recognized the need for an iterative development approach that supported both the exploratory nature necessary in research and a high level of task parallelization. In addition, the PRIME project is obligated to develop a serious game that addresses the needs of 6 end-user companies. This has proven to be an additional challenge as the gaming awareness within the corporate knowledge was very low or even non-existent, which led to a high uncertainty concerning the user requirements [3]. 3.1 Overview The uncertainty of the problem domain entailed the adoption of a development approach that was based on an iterative methodology. However, the classic relevant development models, such as exploratory or spiral, were not sufficiently adequate for PRIME, neither did the methods ensure the successful completion of the entire project within 24 months. A schematic representation of the PRIME development methodology is depicted in Fig. 2, illustrating the iterative approach by the spiral process with four major milestones coinciding with the major releases of the PRIME software. The first milestone at month 6 corresponds to a cardboard prototype of PRIME; the second milestone corresponds to the release of the alpha version of PRIME at month 12, coupled with the testing and integration methodologies; the third milestone is the beta release of the PRIME software coupled with the PRIME-Time methodologies and evaluation framework; and finally at month 24, the final major version of the PRIME software is released along with the result analysis of the evaluation process. Although the milestones are pre-determined, the spiral process implies that the development is dynamic, meaning that the nature of PRIME software adapts to the needs of the six end-users and what is identified by the consortium to be effective generalisations to support strategic decisions in the context of global manufacturing. This approach implies that the involvement of the end-users from the start of the project in truly a user-centred approach.
320
M. Oliveira and H. Duin
Fig. 2. PRIME development methodology based on agile principles
Unlike the classic spiral development methodology, PRIME does not have iterative stages of specifications, design, prototype and evaluation. A more agile approach is taken with the following four deeply ingrained principles: • • • •
People and communication over processes and management tools; Working documents and visual artefacts (prototypes, storyboards, images, animations) over comprehensive documentation; End-user collaboration over frozen functionality; Response to change over established plan.
In addition, the development activities are aggregated into five well-defined strands that operate in parallel: •
•
•
Vision. The aim of the Vision strand is to convey a consolidated view of the project outcome within the consortium, which is essential in managing the user expectations and to provide a common understanding amongst the development team of developers, who are geographically distributed and working in parallel. The vision will continue to evolve as the user centred approach integrates the feedback from the end-users. This strand was initiated 4 months in advance from the official start date of the project. PRIME-Time. The PRIME-Time strand encompasses all the methodologies aimed at the creation of new work environments based on the usage of PRIME integrated into current work environments, in both the Industrial or Academic environments. There are nine key concepts that are part of PRIME-Time: Player (knowledge worker), training, evaluation, time, place, award, motivation, monitoring and management. Design. The Design strand corresponds to the activities that shape and mould the game design of the PRIME serious game, defining the game play and the
Agility in Serious Games Development with Distributed Teams: A Case Study
• •
321
game mechanics, along with the underlying simulation model. The strand takes into account the feedback from the end-users and the concerns related to the integration of PRIME-Time into real work environments. Code. This strand corresponds to many of the traditional implementation activities, namely the user requirements, technical specifications and the actual implementation. Evaluation. The Evaluation strand covers both the testing of the output of the development and the validation of the initial hypothesis that managers will gain experience within a virtual environment that allows soft-failure.
All the strands operate in parallel, with information flowing between them, thus influencing the final output in terms of the PRIME software and PRIME-Time methodologies 3.2 Developer/End-User Partnerships To facilitate the development process, developer-end-user partnerships were established according to Table 2. Each one of the developers acquired de-tailed knowledge of their end-user during the solicitation and elaboration of the user requirements by means of close collaboration. This deep synergy, which was established in the initial phase of the project, allows a developer to champion the interests of the corresponding end-user whilst absent from the developer meetings and technical brainstorming sessions. It was quickly identified that the lack of gaming culture, or awareness of what a serious game was, would add to the development challenges. There-fore, the consortium as a whole and by means of the developer/end-user partnerships worked to create the game culture within the consortium, which included the realization of internal workshops presenting examples of serious games and holding discussions on broad user scenarios. Although 6 partners of the consortium are characterized as endusers, some of the developers themselves are also end-users. As indicated in Table 1, Alfamicro will be using PRIME as a consultancy tool, whilst BIBA, EPFL, MIP, Sintef and Sofia will be using PRIME as an education tool. This promoted ownership and interest in the usage of the PRIME results by all the partners within the consortium. Table 2. Developer/End-user partnerships End-User
Developer
CRF
MIP
IAI
BIBA and Alfamicro
Intracom
EPFL
KESZ
Alfamicro
LEGO
Sintef
Siemens
BIBA
322
M. Oliveira and H. Duin
The developer/end-user relationships will remain in place throughout the duration of the PRIME project. Each pair developer/end-user holds virtual and face-to-face meetings to carry out the necessary work. 3.3 Working Groups During the first twelve months of the project, there has been one kick-off meeting (September 2005 on Madeira Island) and three project meetings (November 2005 in Vienna, March 2006 in Budapest and June 2006 in Athens). The project meeting in Budapest coincided with the six month milestone and the “cardboard” demonstrator was presented to the end-users. With just four project meetings, all the development activities have been done together by means of a collaborative platform and a dynamic flexible management methodology to effectively support creativity within a virtualised space. With seven geographically distributed developer partners, it was necessary to effectively coordinate all the developers involved, which exceeded 25 individuals with differing cultural backgrounds. At the first instance, coordination was based on a classic management hierarchical structure with virtual meetings between workpackage leaders and the corresponding task leaders. The workpackage leaders would in turn have a meeting with the Operation Manager. Only the individuals responsible for active tasks and workpackages needed to participate in the meetings. This structural organization was in operation until the second project meeting (Vienna – November 2005) where it was acknowledged that there were delays in the project. The major culprit was the inherent “waterfall” mentality affecting most of the developers, whom recognised the spiral process of the PRIME development methodology, but remained adamant about sequential task execution. It was evident that an alternative approach was necessary. In addition, the digital platform put in place to support electronic brainstorming was only used by a few individuals. The adopted solution was to set up Working Groups (WGs), which would have a dynamic life cycle, being created to address a particular need and lasting until their purpose had been achieved. These WGs consisted of small number of developers, where each individual would be representative of their own local team. Consequently, the number of participants would not exceed five at the most, but the development team consisted on average of 25 individuals from seven different organizations. On occasion, some additional individuals would participate to a particular WG meeting, but solely as observers. The initial two WGs were focused on the game object model and the system object model respectively. Other WGs were created, dissolved and merged, with the current WGs consisting of (in parenthesis are identified the participating partners with the WG leaders in bold): •
Game Design WG (Alfamicro, BIBA, MIP, Intrapoint and Sintef). This working group is the result of the merging of the Game Object Model and Gamer WG. Their responsibility is to maintain and validate the Game Object Model, the game play and the corresponding game design. The active members of the WG are all individuals who are considered gamers, thus with experience in either playing extensively games during their growing up or developing games.
Agility in Serious Games Development with Distributed Teams: A Case Study
•
•
• •
323
System Design WG (Alfamicro, BIBA and Intrapoint). This working group evolved from the System Object Model group. The group is responsible for the maintenance of the technical specifications and overseeing the implementation of the PRIME server, client and middleware. However, the WG is driven by the outputs of the Game Design WG. Simulation Model WG (Alfamicro, BIBA, EPFL and Intrapoint). This working group is relatively new and emerged from the System Design WG to address all issues concerning the hierarchical simulation model that supports the Virtual Business Environment. Toolset and Artificial Stakeholders WG (Alfamicro, BIBA and Sofia). This working group tackles the orthogonal development activities to the PRIME server and client, namely the PRIME Toolset and the Artificial Stakeholders. PRIME-Time WG (Alfamicro, EPFL and MIP). This working group focuses in developing the PRIME-Time model, identifying barriers and developing effective adoption methodologies.
Initially there would be on average 4-6 virtual meetings throughout a month. With the WGs, the number of meetings increased to an average of eight meetings a week, with some groups having more than two meetings a week. The aim of a WG meet-ing would be to brainstorm using supporting working documents and the existence of clear objectives would keep the duration to less than two hours. Once the WG concept was implemented, it fomented the realization of additional bi-lateral meetings, which would be more brainstorming intensive and exceed the duration of two hours threshold. Although each WG had a group leader, the over-all coordination of all WG was done by someone who assumed the role of Producer. This person be-came responsible for driving all the development activities and being actively involved in all the working groups. The Producer would in turn liaise with the Operation Manager and together, they would identify risks and develop contingency plans that were discussed with the Quality Risk Manager. All the supporting documentation and working documents generated by the WG would be available in an online document management system for easy access by all. 3.4 Development Status Overview The PRIME vision has been developed and matured, providing the framework for all development activities. In addition to the development milestones, many more activities produced results, ranging from surveys to the PRIME portal. At month 6 (end of month February 2006), the release of a cardboard prototype of the PRIME client coupled with a storyboard describing the context of the human player interaction with the VBE. This prototype was limited in functionality, but allowed the end-users to interact with the demonstrator and sign-off their approval. In addition to the demonstrator, a Game Object Model with 200 game entities was developed, along with the game play encompassing the various strategic decisions that a manager of a Business Unit will take. The game design itself has gone through 3 major iterations, taking into account the expectations and gaming background of the end-users.
324
M. Oliveira and H. Duin
The PRIME-Time concept has been developed and matured, thus providing a framework for the integration of PRIME in existing work environments and the evaluation activities. At month 12 (end of August 2006), the implementation of the alpha version has progressed. The feedback of the end-user organizations on the prototypes has enriched the features and contributed to an increase in the complexity of the problem domain. This has led to rescheduling and reprioritization of tasks to accommodate the minor delay introduced.
4 Conclusions In PRIME, it was observed that technology alone is not sufficient to enable a process, namely formulation, discussion and refinement of ideas. However, it is not sufficient either to enforce a singular methodology since the process is dependent on a multidisciplinary team of individuals, each with their own cognitive mental models that most likely differ from one another. Another two important factors are the geographic dispersion of the team and the different contexts associated to a task. The project began with a strict hierarchical managerial approach to the work supported by collaborative tools, namely electronic support for brainstorming. During the third month of the project, it became obvious that delay was incurred and it was necessary to adopt an alternative approach or risking non-completion of the project within the designated time with the allocated resources. In the case of the PRIME team, it was necessary to adopt a flexible project management process that evolved according to the needs as dictated by the circumstances. At the heart of the process was the producer with a hands-on approach, who supports the operational manager of the project. Although during the first twelve months, the producer centralized the information flow, this has gradually changed as the work shifts more into implementation and the creative entropy wanes. The producer is responsible for the coordination of dynamic working groups, which would be responsible for a single or set of tasks for a given time period. In addition to the working groups, there were the end-user/developer partnerships to ensure a user centered development approach. This pairing of partners proved particularly successful with the phase of requirement engineering, where it was very important to guide end-users through user requirement analysis and to inform them at an early stage of the advantages and limitation of the simulation underlying the game. The sharing of information was done via the collaborative platform, which adopted principles of agile programming community, promoting clear communication based on working documents and visual artifacts. The success of the methodology has allowed the project to recuperate the initial delay and accommodate the emerging problems within the development process and the PRIME functionality. The developer partners have initiated the adoption of the methodology into other of their projects, in particular during the phases where creativity is predominant and the team is geographically distributed. The PRIME methodology has a wider applicability, as in the case of the designing and creating the manufacturing processes of a new product, namely involving the use of plastic moulds and special tooling. The design and engineering of a new product
Agility in Serious Games Development with Distributed Teams: A Case Study
325
strongly benefits from a closer contact between product designers, tool makers and product manufacturers. However, further research is necessary to evaluate the effectiveness of the PRIME development methodology in a broader case. Acknowledgements. The authors thank the respective project partners, as well as the European Commission for all scientific, organizational and financial support. The PRIME project is partially funded under contract number FP6-016542 within the Sixth Framework Program-me, Priority IST/NMP.
References 1. Abrahamsson, P., Warsta, J., Siponen, M., Ronkainen, J.: New Directions on Agile Methods: A Comparative Analysis. In: IEEE Proc. International Conference on Software Engineering, Portland (2003) 2. Annetta, L.A., et al.: Serious Gaming: Incorporating Video Games in the Classroom. Educause Quarterly. Number 3 (2006) 3. Baalsrud Hauge, J., Duin, H., Oliveira, M., Thoben, K.-D.: User Requirements Analysis for Educational Games in Manufacturing. In: Proceedings of the 12th International Conference on Concurrent Enterprising: Innovative Products and Services through Collaborative Networks (ICE 2006), Milano, Italy (June 26-28, 2006) 4. Beck, K., et al.: Manifesto for Agile Software Development (2006), last accessed April 2006, http://www.agilemanifesto.org 5. Beck, K.: Extreme Programming Explained: Embrace Change. Addison-Wesley, Reading (2000) 6. Brooks, F.: No Silver Bullet: Essence and Accidents of Software Engineering. Computer 20(4) (1987) 7. Controlchaos (2006), last accessed April 2006, http://www.controlchaos.com 8. Diehl, M., Stroebe, W.: Productivity loss in brain-storming groups: Toward the solution of a riddle. Journal of Personality and Social Psychology 53 (1987) 9. Gamasutra (2006) Last accessed in April 2006, http://www.gamasutra.com 10. Gallupe, R., Dennis, A., Cooper, W., Valacich, J., Bas-tianutti, L., Nunamaker, J.: Electronic Brain-storming and Group Size. Academy of Management Journal 35 (1992) 11. Gouran, D., Hirokawa, R.: Functional theory and communication in decision-making and problem-solving groups: An expanded view. In: Hirokawa, R., Poole, M. (eds.) Communication and group decision making, 2nd edn., Thousand Oaks, CA (1996) 12. Grimson, J., Kugler, H.: Software Needs Engineering – a Position Paper. In: Proceeding of ICSE, Limrick (2000) 13. Glass, R.: Software Creativity. Prentice Hall, Englewood Cliffs (1995) 14. Kracke, R., Baalsrud Hauge, J., Duin, H., Thoben, K.-D.: Training of Strategic Decisions in Collaborative Networks through Serious Games. In: Camarinha-Matos, L.M., Afsarmanesh, H., Ollus, M. (eds.) Network-Centric Collaboration and Supporting Frameworks. IFIP TC5 WG 5.5 Seventh IFIP Working Conference on Virtual Enterprises, Helsinki, Finland, September 25-27, 2006. Springer, Boston (2006) 15. Keith, C.: Agile Methodology in Game Development: Year 3. In: Proc. Computer Games Developer Conference, San Jose (2006) 16. McConnell, S.: The Art, Science, and Engineering of Software Development. IEEE Software (January/February 1998)
326
M. Oliveira and H. Duin
17. Nguyen, L., Carroll, J., Swatman, P.: Supporting and monitoring the creativity of IS personnel during the requirements engineering process. In: Proc. of 33rd Hawaii International Conference on System Sciences (HICSS-33), Maui (2000) 18. Oliveira, M., Andersen, B., Oliveira, A., Rolstadas, A.: Using Serious Games to Improve European Competitiveness. In: Cunningham, P., Cunningham, M. (eds.) Exploiting the Knowledge Economy: Issues, Applications and Case Studies, IOS Press, Amsterdam (2006) 19. Oliveira, M., Andersen, B., Oliveira, A., Rolstadas, A.: A Case Study of Applying Agile Principles to Build a Serious Game in Strategic Manufactuirng. In: Cunningham, P., Cunningham, M. (eds.) 2000. Exploiting the Knowledge Economy: Issues, Applications and Case Studies, IOS Press, Amsterdam (2006) 20. Osborn, A.: Applied Imagination: Principles and Procedures of Creative Thinking. Scribner, New York (1957) 21. Paulus, P., Larey, T., Ortega, A.: Performance and perceptions of brainstormers in an organizational setting. Basic and Applied Social Psychology 17 (1995) 22. Robertson, J.: Eureka! Why analysts should invent requirements. IEEE Software 19 (2002) 23. Squire, K.: Game-Based Learninig. An x-Learn Perspective Paper. Masie Center. ELearning Consortium (2005) 24. Turk, D., France, R., Rumpe, B.: Limitations of Agile Software Processes. In: Proc. of the Third International Conference on eXtreme Programming and Agile Processes in Software Engineering, Alghero (2002) 25. Wiering, R.: A Survey of Structured and Object-Oriented Software Specification Methods and Techniques. ACM Computing Surveys 30(4) (1998) 26. Wood, W., Kleb, W.: Extreme Programming in a Research Environment. In: Wells, D., Williams, L. (eds.) XP 2002. LNCS, vol. 2418, pp. 317–343. Springer, Heidelberg (2002)
Improving the Flexibility of Learning Environments: Developing Applications for Wired and Wireless Use David M. Kennedy1 and Doug Vogel2 1
The University of Hong Kong, Faculty of Education, Hong Kong [email protected] 2 The City University of Hong Kong, Department of Information Systems, Hong Kong [email protected]
Abstract. Mobility is an intrinsic property of learning encompassing spatial, temporal and developmental components. Students’ expectations on how and when they learn are creating increasingly heavier demands upon all aspects of their learning. Young people in particular have made mobile devices fundamental to their daily lives. However, educators and developers are faced with the dilemma: do you develop applications for the mobile or the wired environment? In this paper we argue that learning environments will remain combinations of wired and wireless for the foreseeable future. However, not all affordances offered by wired environments are transferable to small mobile devices. In fact, some tasks are better served by applications that are designed to be entirely mobile. The paper will present initial results and evaluations of five of learning tools with the properties mobility, flexibility and either instructor- or student-generated content. Keywords: Mobile, technology, smart phones, design, software, learning, PDA.
Not withstanding these problems a number of countries are choosing wireless connectivity over extensive wired development. However, the educational issue remains the same—providing flexibility of learning opportunities. Students in many developed countries suffer from problems balancing work and study, which limits their on-campus engagement. Thus, student expectations of when and where learning can occur have evolved to include wireless solutions along with wired internet access at home or university. Therefore, mobile devices may well define the next era in computing as such devices create a new paradigm between anytime/anyplace availability (ubiquity) and computing capabilities (functionality). For example, market trends indicate that sales of notebook computers are currently buoying overall sales of personal computers (PCs), particularly in the developed world, as people move to the convenience of mobile computing (The Register, 2006). Mobile computing devices include a variety of hardware with different functions. They include notebook/laptop computers, personal digital assistants (PDAs), mobile telephones, game consoles, mp3 players (e.g., iPod), DVD and/or CD players and digital cameras. Considerable efforts have already been made to utilise the inherent mobility and flexibility for learning of some of these devices (primarily PDAs and mobile phones—small mobile devices). Small mobile devices (SMDs) have been used successfully in business (e.g., [2], medicine, [3], nursing [4], engineering [5], radiology [6], K-12 schools [7] and more recently as a field data-collection and reflection tool [8]. In the last two years, many of the functions once associated with separate devices are now being incorporated into a single device called a ‘smart phone’ [9]. The original limitations associated with SMDs (e.g., memory, battery life, connectivity, and CPU speed), are rapidly being overcome in smart phones. Moore’s Law suggests that these rapid gains in memory, storage and CPU speed will continue. However, for design purposes SMDs present a different user interface—stylusdriven interactions and handwriting recognition rather than the keyboard and mouse associated with desktop/ notebook computers (D/Ns). In the applications described below, considerable effort has been made to provide: • flexibility of potential use; • alternate, but similar interfaces for wired or wireless devices; and • one-stop authoring of content for both domains. Recent developments (e.g., the iPhone from Apple Inc.) have introduced gesturebased interactions (using a finger rather than a stylus) for navigation and data entry have increased the challenges for researchers in this domain. However, it is the latter (above) that the authors believe will ‘make or break’ uptake of these ubiquitous technologies in the broader educational environment.
2 Mobility and Flexibility for Learning: The New Paradigm? Mobility is an intrinsic property of learning, encompassing spatial (university, workplace, home), temporal (days, evenings, weekends) and developmental (the learning needs/ life
Improving the Flexibility of Learning Environments
329
skills of individuals which change depending upon age, interest or employment) components. Student expectations on how and when they learn are creating increasingly heavier demands upon all aspects of their learning and young people, more than any other group, are making the devices extensions of their personal space fundamental to their daily lives [10, 11]. Figure 1 (after [12], p. 7) is a diagrammatic representation of this view. In Figure 1 the horizontal arrow indicates increasing mobility of people (right to left), while the vertical arrow indicates increasing mobility of the device. Per s onal Mobilit y of t ools
Mobile phones Smart phones
PDAs
Games consoles
Tablet PCs
Clas s room response sy stems
Laptops
Static
Por table Kiosks
Elec tronic whiteboards Video conf erencing
Mobilit y of people
Shar ed
Fig. 1. Mobility of people and tools: Comparing the personal with the shared
The learning tools described in this paper may be associated with the top left quadrant, where people and devices are highly mobile. In considering how to design applications for both mLearning and wired access, a number of key factors need to be addressed. Zheng & Ni ([9] p. 473) suggested that the main elements that need to be addressed in moving from the D/N to an SMD are: • context, where the functional design accounts for screen size, processor speed, and educational needs; • content, resources that can be presented, annotated, queried and answered using the input devices(s); • community, allowing students to share information (text, SMS and images) via Bluetooth and WiFi; • communication, using all of the possible input methods, text recognition, keyboard, stylus; • customisation, so that students have the facility to tailor the device to personal needs; and • connection, via a variety of methods, to support students moving from wireless to wired environments. In practice this often means a decrease in functionality or a change in design features when moving from the D/N to SMDs, but this is not the always the case. In applications that recognise mobility as an intrinsic element needed for learning
330
D.M. Kennedy and D. Vogel
(e.g., field studies, professional practice or engagement outside the classroom), the D/N application may not be suitable. The current developments seek to: 1. develop and research the use of a range of tools for both web and mobile platforms; 2. develop or select software that enables academics to easily publish activities/tasks to the web and/or personal digital assistants (PDAs) or smart phones simultaneously; 3. develop a common set of icons for common tasks (e.g., save, upload, download, login, check answer, hints, information) consistent for each mobile application; and 4. provide pedagogical advice and support (with examples) for developing content suitable for eLearning (web-based and mLearning). In this research a number of applications have been redesigned to provide greater flexibility for student learning by providing access and engagement in the mobile form. The five examples are discussed below. Interactive Graphing Object (IGO): Many concepts are best represented by graphs, where more than two variables can be represented. Many key concepts in science, business and medical sciences are best understood using graphs [13-15]. The Interactive Graphing Object (IGO) was originally developed for the web [13] but was adapted for mobile devices to increase the flexibility of learning available for students. The IGO supports onscreen sketching and curve fitting [13]. The IGO authoring environment also supports web authoring and publishing to either fixed or mobile devices. Tatoes: Tatoes is an XML-based application that converts the highly successful quizgeneration software Hot Potatoes (available free to educational institutions from http://hotpot.uvic.ca/) to a form that functions on a Pocket PC. Authoring is done on a PC and the Tatoes software converts the files (multiple-choice, fill-in-the-gap, ordered lists, matching exercises) to run on any device with the Windows .NET mobile platform installed. An instructor creates a series of questions using Hot Potatoes publishes them to the web or mobile form. Crossword Puzzle: The Crossword tool is part of the Tatoes environment. While generally focusing on knowledge-type questions, it has been popular with students in the mobile form, as they can undertake the exercise when away from any network, wired or wireless. Build-a-PC Tool: The student is required to select from a range of computer components to build a desktop computer system appropriate for specific people in an organisation. Students must make decisions based upon the position held, responsibility level and budget available. Extensive feedback is provided about their decisions. In each of these four examples above, instructors author once and publish to the internet and/or mobile devices with one click. However, the last application is not under direct instructor control, rather it is student-controlled. Phototate: This software utilises the camera feature found in smart phones and PDAs. Students may capture images, annotate the image with a pen tool, and add audio content to the final saved file. In Tables 1 and 2 the differences between the web and mobile forms are shown using screen captures from the online and mobile versions of the same application (except for Phototate). The first four tools have been used with two cohorts of first-year information systems students (totalling 1,600 students) at the City University of Hong Kong
Improving the Flexibility of Learning Environments
331
(CityU), while Phototate has been piloted on a field trip to Norway by third-year science students at CityU.
3 Designing for Fixed and Wireless Devices Development of the mobile applications has been undertaken since 2004. In Figure 2, a framework showing the software developed as part of this project is presented in relation to mobility (horizontal axis) and content creation, instructor or student (vertical axis). In the majority of institutions of higher education, applications have been developed from the premise that connection to computing environments will be at fixed locations using a D/N computer (e.g., university, schools, home, computer laboratories and internet cafes). Table 1 then compares the interfaces of the fixed and mobile versions of each of the tools described above.
Fig. 2. Content creation for applications intended for fixed and wireless devices
The second dimension considers who is responsible for the creation of the content. While the internet has created many opportunities for teaching and learning, and many interactions/tasks can be designed by instructors, the underlying assumption is that the student is the recipient of the content. The major exception to this statement is the use of computer-mediated communication (CMC). Thus far, the use of mobile devices for students to generate content has been limited to text (SMS, email and CMC), digital photographs and video, and data collection using a forms-based interface). In this project providing tools and resources to enable students to learn at a time most suitable for the student, and simple to use authoring interfaces (PC-based) are seen as primary factors guiding the developments. The major expected outcomes are flexibility of learning opportunities. Flexibility incorporates a: • connected mode (at the campus); • nomadic mode (at home or connected to a desktop computer on campus or at home), and • disconnected mode (on public transport, away from wired or wireless connections) (after [9])
332
D.M. Kennedy and D. Vogel
Table 1. Comparing applications, desktop/notebook versus mLearning equivalents (five tools) Desktop or laptop computers Tool: IGO/ mIGO
Small mobile devices
A set of icons has been used to simplify the user interface. Input of actual values in the mobile version is no longer available and, in the mobile version, longer questions must be scrolled to read the entire text. However, both versions support complex curve fitting, iterative forms of feedback, zooming, and saving of files (students may save their work in progress).
Tool: Hot potatoes/ Tatoes
In moving from the D/N to the SMD some limits on text length are needed to reduce the amount of reading on the SMD and the use of images is problematic. However, for text-based questions the mobile version has all of the functionality of the full-sized version.
Tool: Crossword
Improving the Flexibility of Learning Environments
333
Table 1. (continued) The two versions are almost the same, with a much more efficient use of screen real estate in the mobile form. The major difference is the need for the questions and hints on the mobile version to overlap the crossword, unlike the web version where the clues are at the top of screen.
Tool: Build-a-PC
The two versions are effectively, the same. However, what has changed is the way in which students engage with the task. In the D/N version, students use a drag-and-drop method using a mouse. In addition, the mouse-over of a component produces a description of the item. On the SMD version, the tap-and-click method is used (the object is selected and then added using a tap on the appropriate button, or the student selects ‘Detail’ to know more about the component).
Tool: Phototate
Phototate is software that allows students to collect data in the field or undertake reflective practice. Phototate allows students to annotate the photos taken with a PDA or smart phone with a variety of pen colours and attach an audio file to the annotated image. The composite file may then be uploaded to a Course Management System and the student’s ePortfolio for future analysis and reflection. The file on the left is from a science field trip to Norway, uploaded to a webpage. The attached audio file consists of a discussion about nutrient levels versus water clarity between a student and the instructor. On the right is a partially annotated photo of students on a PDA. Research in the use of this application is ongoing.
To date, what has been shown is that the change in interface between the more fully featured web versions of individual applications has not interfered with students who use the mobile form of the tools for self-guided learning. Numerous authors see
334
D.M. Kennedy and D. Vogel
Table 2. Perceived limitations of SMD versus actual experience: Designing for web and mobile applications
Perceived inhibitors
Key functional differences: Web/Mobile
Small screen size
Current screen size creates overlapping text and/or graphics, especially for the feedback to questions. This has not been an inhibitor to the use of the mobile versions of the content in the minds of the students.
Nonergonomic input method
From one perspective, the stylus or onscreen text recognition limits the speed and flexibility of input for an annotation. However, a stylus also allows more flexibility for annotations using Phototate software, and text input has never surfaced as a significant issue (student focus groups).
Slow CPU speed
Limited memory
Limited battery span
Applications run more slowly than on a desktop computer, but the speed at which all of the applications run is deemed satisfactory by students. The size and power requirements of more powerful CPUs limit what can be placed in mobile devices. This is a limitation only on the most demanding of applications. All of the examples described above operate satisfactorily on PDAs up to two years old. The advent of cheap flash memory has overcome storage issues. Up to 100 Phototate files can be stored on one 512 Mb flash memory card. Extended use is limited by current technology, and this remains a problem, particularly the nonpersistence of information with some hardware.
Everchanging OS
This is a current and likely future problem not resolvable in the short term with many competing systems (e.g., Symbian and Windows mobile).
Infrastructure compatibility
There is a plethora of wireless standards, some of which require large amounts of power, but also offer increased bandwidth and transfer speeds.
Connectivity bandwidth
Current wireless networks have sacrificed speed for access, but this is changing rapidly.
Future solution
Flexible film display Voice recognition Projection keyboard Cursive hand-writing recognition improvements New breed of architecture for faster CPU
New breeds of lithium batteries or fuel cells Open-source OS for mobile devices (e.g., Linux has been run successfully on a smart phone) Standards are still being developed to bridge the mobile platform. 3G mobile capacity and Bluetooth v.1.2 More efficient wireless protocols than currently available
Improving the Flexibility of Learning Environments
335
only problems of security, screen size, battery life and CPU speed; however, we focus on opportunities and flexibility for student learning. In Table 2 these issues perceived as problematic are discussed in light of our experiences and feedback from students. The list of problems is adapted from [after 16], but the improvements in mobile affordances since 2004 have been relentless, negating many of the perceived problems just a few short years ago. 3.1 Brief Summary of Current Evaluation Data Qualitative and quantitative evaluations have been undertaken with Tatoes, Phototate and build-a-PC and are reported elsewhere [8, 17] (416 students participated in various tasks from a cohort of an introductory business course). In the research reported thus far, the use of the exercises on the PDAs was not mandated so students tended to download exercises more than upload the completed work. However, correlation with scores on the mid-term exam revealed that students who either downloaded or uploaded Tatoes (for example) scored significantly higher on the midterm exam (p <.01). Since the tutorial scores were a good indicator of general student study performance, we adopted it as an ANCOVA covariant. Students who downloaded or uploaded Tatoes instances had significant higher scores on the midterm exam (p < 0.001) than students who did not participate. Examination of the higher performing (and poorer performing) students in their use of Tatoes in quantity and depth (i.e., exploring alternatives) was revealing. For example, a student who uploaded every exercise and also extensively explored alternatives got the 6th highest score on the midterm. The top 5% of midterm scores (41) represent 26% of the heaviest Tatoes users including 8 out of the top 10 mid-term scores. However, a number of students who uploaded most of the Tatoes exercises but did not do any exploration of alternatives are scattered throughout the midterm exam scores including 39% in the bottom half of the class. The suggestion might be that they just went through the exercise without thinking and the results confirm that behaviour. On of the key findings from the tutors indicated that students who engaged with the use of mobile devices for learning in and out of the classroom found the experience of using a PDA provided a significantly better set of learning experiences, was fun, and useful for motivation and confidence building. Approximately one quarter of the total student cohort of 1,600 have engaged with all or some of the tools in spite of the voluntary nature of the exercises. While the initial evaluation results were encouraging, there still remains a great deal of work to be done to explore more fully the reasons why students used or did not use the exercises. In particular, the ongoing research seeks to determine why students (almost all) did not use the high quality feedback provided (all incorrect alternatives had a detailed description of why the selection was incorrect) with the Tatoes exercises. The expectation was that students would take the opportunity to explore incorrect alternatives in order to improve their understanding: this was not observed. The current cohort of students are experiencing specific teaching interventions in an attempt to determine if their current approach to learning can be modified, and encourage more students to use the exercises.
336
D.M. Kennedy and D. Vogel
4 The Future The future may see a paradigm change as mobile devices become more integrated into educational environments (see Table 2). However, for this to occur we need to establish sound pedagogical frameworks based upon experience and research into how students use the tools in practice: what the specific learning needs are, and how more effective feedback, communication and collaboration can be enhanced. Four tools, Tatoes, Crossword, IGO and Build-a-PC, have been designed with the facility for lecturers to provide high quality feedback to students. The use of the Tatoes environment provides limited evidence that writing multiple-choice questions with detailed feedback may encourage some students to not only look for the correct answers to a question, but to spend time examining what makes other distracters wrong. Phototate provides a tool for reflective practice and data generation by students away from traditional classrooms and lecture halls, providing support for more flexible modes of learning. The focus in the past has been on the perceived limitations of mobile devices, rather than the pedagogical affordances and flexibility offered. The rate of uptake (voluntary) by students has been steady as they take advantage of the: • the quality of the feedback provided for incorrect distracters and correct answers; and • flexibility offered by anytime/anyplace learning. What this project has made clear is that offering students the flexibility of mobile learning options in addition to current wired learning environments is viable now, and not something that educators need to wait for. The key issues are not technological, but pedagogical and institutional.
References 1. Vernon, R.D.: Cornell data networking: Wired vs.wireless? In: Office of Information Technologies, vol. 2006, Cornell University, Ithaca (2006) 2. Easton, J.: Going wireless: Transform your business with mobile technology, Harper Collins, New York (2002) 3. Smørdal, O., Gregory, J.: Personal Digital Assistants in medical education and practice. Journal of Computer Assisted Learning 19, 320–329 (2003) 4. Altmann, T.K., Brady, D.: PDAs bring information competence to the point-of-care. International Journal of Nursing Education Scholarship 2 (2005) 5. Perry, M., Michael Jacob, J.: PDA in the classroom project. In: 2005 Illonois-Indiana Sectional Conference, Americian Society for Engineering Education, Northern Illinois University, DeKalb (2005) 6. Boonn, W.W., Flanders, A.E.: Survey of Personal Digital Assistant use in radiology. RadioGraphics, 537–541 (2005) 7. Johnson, D.W., Rudd, D.: Will handheld computers succeed in college? In: Colton, D., Tastle, W.J., Hensel, M., Abdullat, A.A. (eds.) ISECON 2003. §4113, San Diego (2003) 8. Vogel, D., Kennedy, D.M., Kwok, R.: Do Mobile Device Applications Lead to Learning? In: Montgomerie, C., Seale, J. (eds.) EDMEDIA 2007, Association for the Advancement of Computers in Education, Chesapeake VA, pp. 555–564 (2007)
Improving the Flexibility of Learning Environments
337
9. Zheng, P., Ni, L.M.: Smart phone & next generation mobile computing. Elsevier Inc., San Francisco (2006) 10. Vavoula, G., Sharples, M.: KLeOS: A personal, mobile, knowledge and learning organisation system. In: Milrad, M., Hoppe, U., Kinshuk (eds.) WMTE 2002, Institute of Electrical and Electronics Engineers (IEEE), Vaxjo, Sweden, pp. 152–156 (2002) 11. Sharples, M., Taylor, J., Vavoula, G.: Towards a theory of mobile learning. Mobile technology: The future of learning in your hands. 58 (2005) 12. Naismith, L., Lonsdale, P., Vavoula, G., Sharples, M.: Literature review in mobile technologies and learning, FutureLab, Bristol, UK, vol. 48 (2005) 13. Kennedy, D.M., Vogel, D.R., Xu, T.: Increasing opportunities for learning: Mobile graphing. In: Atkinson, R., McBeath, C., Jonas-Dwyer, D., Phillips, R. (eds.) ASCILITE 2004: Beyond the Comfort Zone, Perth, Western Australia, pp. 493–502 (2004) 14. Kozma, R.B.: The use of multiple representations and the social construction of understanding in chemistry. In: Jacobson, M., Kozma, R. (eds.) Innovations in science and mathematics education: Advanced designs for technologies of learning, Erlbaum, Mahwah, NJ, pp. 11–46 (2000) 15. Kremer, R.: Visual languages for knowledge representation. In: Eleventh Workshop on Knowledge Acquisition, Modeling and Management (KAW 1998), Knowledge Science Institute, Banff, Alberta, Canada, vol. 2006 (1998) 16. Csete, J., Wong, Y.-H., Vogel, D.: Mobile devices in and out of the classroom. In: Cantoni, L., McLoughlin, C. (eds.) EDMEDIA 2004. Association for the Advancement of Computing in Education, Norfolk VA, pp. 4729–4736 (2004) 17. Vogel, D., Kennedy, D.M., Kuan, K., Kwok, R., Lai, J.: Do Mobile Device Applications Affect Learning? In: Sprague, R.H. (ed.) HICSS-40, Hawai’i. Institute of Electrical and Electronics Engineers (2007)
Evaluating an e-Learning Experience Based on the Sakai Environment Félix Buendía García1 and Antonio Hervás Jorge2 1
Dept. of Computer Engineering [email protected] 2 Dept. of Applied Mathematics Universidad Politécnica de Valencia, Spain [email protected]
Abstract. In the context of the Bologna process, e-learning experiences are being promoted in many European universities. Recently, the Universidad Politécnica of Valencia (UPV) has implemented the Poliformat platform that is based on the Sakai environment. Last year, a pilot experience was developed to test this environment for undergraduate courses in a Computer Engineering degree. The current work describes an e-learning application of the Poliformat platform in an Operating Systems online course and how this experience has been evaluated using survey techniques that analyze online learning issues and also technical aspects related to the Poliformat platform. Globally, the experience was rather positive although some problems were detected in the evaluation process that revealed that e-learning methods and tools have to be improved. Keywords: Online learning, e-learning platforms, online courses, evaluation.
Evaluating an e-Learning Experience Based on the Sakai Environment
339
2 Description of Sakai Sakai is a consortium of universities, colleges and commercial affiliates working in open partnership with standards organizations and other open-source initiatives to develop “community-source enterprise-scale software applications to enhance collaboration, research and teaching within higher education” [2]. The core members include MIT, University of Michigan, Stanford, Indiana University, and uPortal. The Sakai environment is based on an extensible service-oriented architecture for building and deploying enterprise-scale collaboration, teaching and research tools and services. Sakai uses Java-based technologies that enable teaching tools such as a complete course management system, support for collaborative activities and a services-based portal. The UPV became a Sakai partner in 2005 and adapted its environment to produce the Poliformat platform. Some of the Poliformat contributions are the integration within the corporate systems and applications, the customization of appearance and the internationalization, including the translation to Spanish language.
Fig. 1. Poliformat main screen
Fig. 1 shows an example of Poliformat screenshot that displays the main window in the Operating Systems course. The Poliformat window is structured into several areas:
The head area displays the available course sites (e.g. Operating Systems). The left area shows the main subject options such as Syllabus or Schedule. The central area displays the subject information. The right area displays other site sections such as announcements, forums or chat messages.
340
F.B. García and A.H. Jorge
3 The e-Learning Experience 3.1 Teaching Context “Operating Systems” is present in most of the computing curricula of both undergraduate and Master’s programs. At the Computer Engineering Department of the UPV more than twenty instructors teach Operating Systems courses to about a thousand students. Table 1 shows the organization of the Operating Systems courses. Each course involves both a theoretical and a practical part. The percentage of the practical part varies depending on the learning goals of each course. Introductory courses such as SO1 require students to acquire a solid theoretical foundation, thus the teaching methodology favors the theoretical part. Intermediate and advanced courses (i.e., SO2, ADS, or ESO) try to balance this percentage by covering new conceptual topics and performing some practical activities. This methodology allows students to learn how operating systems are designed and how to use their services. Table 1. Operating System courses Course name Operating Systems I (SO1) Operating Systems II (SO2) System Administration (ADS) Operating System Study (ESO)
Type Introductory
#Students 567
Intermediate
650
Advanced
187
Advanced
84
Unfortunately, this teaching methodology becomes inefficient in mass courses because when the student/teacher ratio increases, the traditional relationship between the students and the instructor becomes affected so that it is more difficult to provide right feedback and to monitor students’ progress. In order to alleviate the problem of mass courses and to study the requirements and benefits of distance learning, the Computer School at the UPV started in 2002 an innovative experience providing students with distance learning for some Operating Systems courses, including SO1 and SO2. Web repositories and e-mail tools were used to provide remote assistance and automatic assessment. These courses were aimed at students who were not able to attend traditional courses for any reason whatsoever. During the 2005-2006 academic year, the Poliformat platform was used to support a SO2 course that was taught to 48 students (about 650 in classroom-based courses). The students joined voluntary the course and they knew the general conditions (e.g. the assessment system was the same for all students either in classroom-based or in online courses). 3.2 Online Learning Method The current e-learning experience is based on an online learning method to organize the course items and control its delivery [3]. This method must be independent from
Evaluating an e-Learning Experience Based on the Sakai Environment
341
Fig. 2. OS Course flow chart
the e-learning platform that supports the specific course. In this case, the proposed method combines several techniques from self-paced autonomous learning for theoretical issues to programmed instruction used in practical activities. Fig. 2 shows a flow chart that displays the main steps in the course organization. In the current approach, a preliminary evaluation step is required to test the students’ goals and to check if they are able to meet the course conditions. If the stated conditions are accepted then the student is enabled to join the online course (a kind of “learning agreement”) but if students do not “sign up” the agreement they are recommended to join the alternative class-room based course. The next step consists of scheduling the course items using agenda services and assigning a calendar of recommended tasks and activities by weeks. From this step, online learning issues are planned according to two different “parallel paths”:
The acquisition of theoretical knowledge (concepts, facts…) is based on a selfpaced learning in which the student decides what concepts are interested in and the way and time to get them. A collection of resources and material about the target subject are available in the Resource repository.
342
F.B. García and A.H. Jorge
The practical skills are obtained by performing activities that involve the working with real systems or simulations [4] in the case of technical subjects (e.g. Operating Systems). These activities also require some theoretical fundamentals but they have to be acquired under the student responsibility. Nevertheless, the achievement of practical activities is based on a paradigm of programmed instruction. Students receive assignment information about the proposed activities and they have to deliver results in a tightly scheduled way.
After theoretical concepts have been acquired and practical activities performed, a formative assessment is performed in order to evaluate the student’s knowledge and skills. This evaluation is only informative and it has not a grading purpose. Once learning activities have been finished, the next step is a post-course evaluation that queries students their point of view about the course delivery. Finally, all students either in online or classroom-based courses perform the final examination for getting their grades. 3.3 Implementation Details The previous online learning method has been applied to the course of Operating Systems using the Poliformat platform. After the student accesses the platform, he or she views the course site as Fig. 1 displays. The course site provides information about the course items although a specific learning path is not forced to the student. From this point, the student can select the different items that compose the course. The Poliformat platform provides a Schedule service that allows instructors or site organizers to post items in a calendar format. The calendar has day, week, month, year, and flat list views. The information about events, notifications or programmed activities is organized by means of the Announcement service that is used to inform site participants about current items of interest. Other kinds of information related to the course subject are provided by the Resources service. Different types of resources include documents (e.g., word processing documents, spreadsheets, slide presentations, plain text), links to websites, and simple text documents. Fig. 3 shows a screenshot that displays a repository of resources related to the Operating Systems course.
Fig. 3. Example of Poliformat repository
Evaluating an e-Learning Experience Based on the Sakai Environment
343
There are other services mainly addressed to provide communication facilities such as Discussion, Email, Chat or News, or to manage learning information such as Modules, Quizzes, Assignments or Gradebooks. In the context of the current experience, the next services are of particular interest:
The Assignment service that implements the practical activities provided in the course. This service allows instructors to create, distribute, collect, and grade online assignments. Assignments are private; student submissions are not visible to other users of the site. Fig. 4 shows a screenshot that displays some settings about a practical activity in the course example. The Tests & Quizzes section that allows the instructor to create online assessments for delivery via a web interface to the course students. They are used in the formative assessment of the course (see Fig. 2).
Fig. 4. Example of assignment features
4 The Experience Evaluation This section describes how the current e-learning experience has been evaluated in the context of the Poliformat platform. This evaluation is based on subjective methods such as surveys and checklists. In this experience, two different evaluation processes were performed as Fig. 2 shows. A pre-course evaluation reviewed the students’ profiles (e.g. how the student was reported about the online course or if he/she joined the course voluntarily) and asked them if they are able to carry out the proposed course activities. The Survey tool was not available in the Poliformat platform at the beginning of the course and a questionnaire was proposed to the students. The evaluation questionnaire was based on “true/false” questions and most of the student answers agreed with the course conditions. A more detailed evaluation was performed at the end of the course and the Survey tool developed at the UPV was used to elaborate the proposed questionnaire. The
344
F.B. García and A.H. Jorge
evaluation questionnaire was based on “Likert-scale” questions (from strongly disagree-0 valued to strongly agree-10 value). Table 2 shows a set or questions related to general issues of the course evaluation (independent from the e-learning platform) that were submitted to the students. Table 2. Post-course general issues Issue Prerequisites Expectations Scheduling Goals Resources Expected activities Timing Academic level Assessment Methodology Assistance Satisfaction
Question Do you think that the online course would need some prerequisite to join it? Does the course address your expectations? Has the course scheduling been met? Are the course goals adequately stated? Are the available resources adequate for the course? Do you think that programmed activities are the expected one? Is the time assigned to activities enough? Is the academic level of activities adequate? Is the assessment system adequate to the course activities? Do you think that the course methodology is adequate? Is the teacher assistance adequate? Do you think that the course experience has been satisfactory?
The obtained results are displayed on Fig. 5 that provides a view of the weak aspects detected by students. The evaluation shows that prerequisites were not mainly considered by students as a crucial aspect and some complaints were stated about the online e-learning methodology and the academic level of practical activities. Globally, the degree of satisfaction with the course was acceptable.
Evaluating an e-Learning Experience Based on the Sakai Environment
345
Table 3 shows some technical aspects about the Poliformat platform that were asked to students. The proposed questionnaire was also based on “Likert-scale” questions that checked the effect of the e-learning platform in the course example. Table 3. Post-course technical issues Issue Handiness Interface Help Operation Reliability Communication effectiveness Additional tools Activity support Assessment support
Question Is it easy to interact with the platform? Is the user interface of the platform friendly? Are help options in the platform useful? Does the platform operate as expected? Are serious platform errors happened? Are communication tools effective? Do you think that the platform would need other tools? Is the platform support enough for practical activities? Is the platform support enough for assessment?
The results about the evaluation of technical aspects are displayed on Fig. 6. Students were particularly critical with operation and reliability issues. In the first year of the Poliformat application, there were several system crashes and login failures that were noticed by students. Another complaint was focused on the handiness to interact with the platform. A common student point of view was the demand of additional tools for communication purposes.
Post-course evaluation (II) Assessment support
Additional tools
Reliability
Help
Handiness 0
1
2
3
4
5
Fig. 6. Evaluation of technical issues
6
7
8
9
346
F.B. García and A.H. Jorge
5 Conclusions The current work has described an e-learning experience under the Poliformat platform that is based on the Sakai environment. The experience has been focused on a technical subject (Operating Systems) that is part of a Computer Engineering degree. An online learning method has been used to drive the e-learning experience and some implementation details about the role of the Poliformat platform in the experience have been reported. Finally, the experience has been evaluated using a two phase approach (before and after the learning activities) and considering, on the one hand, the general issues about the online learning method and, on the other hand, the technical issues introduced by the Poliformat platform. The obtained results reveal that methodology aspects have to be improved in the course delivery, mainly, the lack of support for teaching theoretical issues, and special attention must be given to system maintenance, avoiding the platform failures and improving the user interaction. Further works are addressed to take advantage of the evaluation results in the management and delivery of future courses. Another aspect to improve is the systematization of evaluation procedures that will help the generation of survey questionnaires. In this context, the use of learning standards such as IMS Learning Design will contribute to facilitate the evaluation of online courses in e-learning platforms. Acknowledgements. This work has been partially supported by the Ministerio de Educación y Ciencia, Spain, under Grant TIN2005-08788-C04-02.
References 1. Mengod, R.: Poliformat, the Sakai-based on-line campus for UPV - history of a success. In: 5th Sakai Conference, Vancouver, BC, Canada, 30 May - 2 June (2006) (Community Source Week) 2. White, A.: Introducing the Sakai Community. In: 4th Sakai Conference, Austin, TX, USA (2005) 3. Anderson, T.D., Elloumi, F. (eds.): Theory and Practice of Online Learning Athabasca. Athabasca University (2003) 4. Buendía, F., Cano, J.-C.: WebgeneOS: A Generative and Web-Based Learning Architecture to Teach Operating Systems in Undergraduate Courses. IEEE Transactions on Education 49(4), 464–473 (2006)
Author Index
Baker, Thar 3 Baroncini, Vittorio 184 Bertolotto, Michela 277 Brucks, Claudine 35 Ceravolo, Paolo 19 Cheng, Kin-Fai 292 Cheung, Christy M.K. 257 Ciuffoletti, Augusto 72 Collet, Christine 109 Congiusta, Antonio 72 Connolly, Thomas M. 304 Cornacchia, Michele 184 Costagliola, Gennaro 213 Damiani, Ernesto 19 Deters, Ralph 124 Duin, Heiko 314 Dyachuk, Dmytro 124 Ferrucci, Filomena 213 Fuccella, Vittorio 213 Fugazza, Cristiano 19 Garc´ıa, F´elix Buend´ıa 338 Garc´ıa-Ba˜ nuelos, Luciano 109 Godart, Claude 60 Guerra, Esther 169 Hainey, Tom 304 Hewett, Rattikorn 45 Hilker, Michael 35 Jankowski, Gracjan 72 Jankowski, Michal 72 Jorge, Antonio Herv´ as 338 Jou, Chichang 230 Kadobayashi, Rieko 157 Kaushik, Saket 201 Kennedy, David M. 327 Krajiˇcek, Ondrej 72 Lam, Paul 292 Langend¨ orfer, Peter
85
Lara, Juan de 169 Lee, Matthew K.O. 257 Leung, Shirley 292 Liu, Ying 245 Livi, Stefano 184 Loh, Han Tong 245 Luoma, Olli 98 Maaser, Michael 85 Malizia, Alessio 169 McArdle, Gavin 277 McNaught, Carmel 292 Meyer, Norbert 72 Miseldine, P. 3 Monahan, Teresa 277 Oliveira, Manuel Ortmann, Steffen
314 85
Portilla, Alberto
109
Randles, M. 3 Reed, Karl 19 Schewe, Klaus-Dieter 139 Schommer, Christoph 35 Stansfield, Mark 304 Taleb-Bendiab, A. 3 Thalheim, Bernhard 139 Vargas-Solar, Genoveva Vogel, Doug 327
109
Wagner, Cynthia 35 Weires, Ralph 35 Wijesekera, Duminda 201 Yildiz, Ustun
60
Zechinelli-Martini, Jos´e-Luis Zhan, Jiaming 245 Zurolo, Luigi 213
