Personalization of Interactive Multimedia Services: A Research and Development Perspective
Edited by
José Juan Pazos Arias, Carlos Delgado Kloos and Martín López Nores
Contents
1 Preface ...... 3
Part I Legal and Technological Foundations
2 Legal Framework for Personalization-Based Business Models Teresa Rodríguez de las Heras Ballell ...... 9 3 Technological Platforms, Convergence and Adaptation of Interactive Contents Fabio Paternò ...... 23 4 Filtering Techniques for Selection of Contents and Products Qing Li, Yuanzhu Peter Chen, Zhangxi Lin...... 41 5 Personalization and Contextualization of Media Marcus Specht, Andreas Zimmermann and Tim de Jong ...... 55 6 Contextualized Recommender Systems: Data Model and Recommendation Process Wolfgang Woerndl and Johann Schlichter ...... 65 7 Enhanced Media Descriptions for the Automatic Adaptation of Audiovisual Content Retrieval Iván Cantador, Fernando López, Jesús Bescós, Pablo Castells, and José M. Martínez ...... 77 8 A Semantics-Based Approach Beyond Traditional Personalization Paradigms Yolanda Blanco-Fernández, Alberto Gil-Solla and Manuel Ramos-Cabrer...... 89
Part II Applications
9 PROGNOSIS: An Adaptive System for Virtual Technical Documentation Nikolaos Lagos and Rossitza M. Setchi ...... 105 Contents 1
10 Investigating the Applicability of Content-Based Filtering on Digital Television Advertisements George Lekakos ...... 119 11 The Role of Communities in E-Commerce Rodolfo Carpintier Santana ...... 129 12 Personalized Education Rosa M. Carro, Estefanía Martín ...... 135 13 New Trends for Personalised T-Learning Marta Rey-López, Ana Fernández-Vilas, Rebeca P. Díaz-Redondo ...... 149 14 M-Learning: Personalized and Pervasive Education Anytime, Anywhere and for Everybody Mario Muñoz Organero, Elena Yndurain, Carlos Delgado Kloos ...... 163 15 Self-Adaptive Portals: E-Government Case Study Ljiljana Stojanovic and Nenad Stojanovic ...... 183 16 Use Cases and Challenges of Personalized Healthcare Mikhail Simonov ...... 197 17 Context-Aware Personalization for Services in the Smart Home: A Semantic Approach for OSGi services Rebeca P. Díaz-Redondo, Ana Fernández-Vilas, Manuel Ramos-Cabrer ...... 207
18 New Perspectives to Lifecycle Management of Context-Aware and Personalised Services Bharat Bhushan, Stephan Steglich, Christian Räck and Bernd Mrohs ...... 223 19 The Role of Service Discovery in Context-Aware, Pervasive Computing Applications Carlos García-Rubio and Celeste Campo ...... 233 20 Personalizing Pedestrian Location Services through Ontologies and Rules Vassileios Tsetsos, Vassilis Papataxiarhis and Stathes Hadjiefthymiades ...... 245
Part III Open Problems and Current Research Trends
21 Conclusions and Future Trends José Juan Pazos Arias, Carlos Delgado Kloos and Martín López Nores ...... 259
1 Preface
During the last few years, we have witnessed the deployment of new digital communication tech- nologies such as satellite, terrestrial or cable forms of Digital TV broadcasting, 3G mobile networks, xDSL, domotic networks, WiMax, ..., which make up a wealthy scenario that will be further en- hanced with the ongoing developments of TV broadcasting networks for mobile devices through standards like DVB-H in Europe, DMB in South Korea or MediaFLO in North America. In this new scenario, the users can benefit from an ever growing bandwidth at lower pricer, which promotes the development of new communication services. We may say that we are living a process similar to what has been happening with informatic equipment, characterized by a progressive increase of memory and computational power at decreasing cost, which favored the access to computers to mas- sive sectors of society. By means of diverse initiatives, public administrations are striving to exploit this situation of unnegligible advantages and opportunities, pursuing the goal of achieving universal and ubiquitous access to the Information Society, thus reducing the effects of the digital divide that draws a clear distinction between those who make frequent use of the information and communica- tion technologies and those who do not know these technologies or find severe difficulties in their use. >From these facts, it would be expectable to see a rapid penetration of technology in many differ- ent domains of applications, including entertainment, news, education and training, administration, electronic commerce, healthcare, etc. Certainly, the offerings have augmented much, with a clear example in the number of TV channels that can be already received in homes. However, the pene- tration of information services has been noticeably much slower than expected, which is mainly due to the problem of information overload: frequently, the users are faces with such an overwhelming amount of information that it is very difficult for them to make decisions or remain informed about a given topic. This poses enormous challenges not only to the users themselves, but also to the market forces that would expect to get revenues from the contents and services they develop, because their presence is quickly diluted in a vast disorganized amount of pieces of information and services of a similar nature. A similar situation started to be noticeable during the mid 1990s with the exponential growth of the Internet, that made the users fell disoriented among the myriad of web contents available through their PCs. This context gave birth to a set of tools commonly referred to as search engines (e.g. Google and Yahoo), that would retrieve relevant web pages in response to user-entered queries. That model proved effective and successful, with millions of people using them to find pieces of in- formation and services in the World Wide Web. However, the advent of new access devices different from the PC (DTV set-top boxes, mobile and smart phones, PDAs, media players, car navigators, etc) introduces a new range of consumption habits and usage contexts that render the search engine 4 paradigm insufficient. In this broader picture, it is no longer realistic to think that the users will bother to purposefully visit a given site, enter a query with the details of what they want, and then select the contents they might be interested in from among those in a list. The reasons may have to do with the users adopting an inactive or predominantly passive role (e.g. while driving or watching TV in the living room); the devices providing no accessible means to enter text (as it happens with remote controls); only a downstream communication channel being available (as in broadcasting environments); or users feeling uneasy with the interfaces provided. To tackle these issues, a large body of research is being devoted nowadays to the design and provision of personalized information services, with a new paradigm of recommender systems proactively selecting the particular pieces of information and functionality that match the preferences, interests and needs of each individual user at any given moment. In this book, we have gathered contributions of experts from academia and industry worldwide, in order to provide a research and development perspective on personalized information services, to help the reader understand their goals, the major steps of their evolution, the functionalities they can deliver to the users and the main challenges they still face. The contributed chapter have been arranged into three parts: • Part I is devoted to presenting the legal and technological foundations involved with the devel- opment of personalized information services. – To begin with, it is noticeable that personalization is typically opposed to privacy, since its effectiveness depends on accumulating information about the users, while there is a body of legislation to restrict the management of personal information by commercial or admin- istrative entities. In Chapter 2, Teresa Rodríguez de las Heras Ballell provides an in-depth (though accessible) discussion of the most significant legal issues arising from personal- ization strategies, together with a description of personalization-based business models and legal mechanisms to implement personalization. – As mentioned above, personalization arose in the context of the Internet as accessed through personal computers, but its scope is now growing much to become a multi-device experience. In Chapter 3, Fabio Paternò analyzes the aspects related to the provision of personalized ser- vices over various technological platforms, addressing important questions of convergence, adaptation of interactive contents and migration of user interfaces, with a comprehen- sive list of the many dimensions of the related research. – Personalization is achieved by matching characterizations of the contents and services avail- able with the information stored in profiles that capture the preferences, interests, likings and needs of the users. This process is commonly referred to as filtering. In Chapter 4, Qing Li provides insight into the existing approaches to do so, which lie at the core of most recom- mender systems regardless of particular application domains. – One of the features that is concentrating most efforts in personalization research is context awareness, which aims at acquiring information about the physical and social situation in which computational devices are embedded to maximize the value of the information de- livered to the user. In Chapter 5, Marcus Specht, Andreas Zimmermann and Tim de Jong explain the phases of information acquisition, representation and management in context- aware systems, with illustrative examples in ubiquitous computing for learning and informa- tion delivery. – Traditionally, the research and development efforts involved with filtering techniques paid little attention to the advances in context awareness, and vice versa. In Chapter 6, Wolfgang Woerndl and Johann Schlichter provide a comprehensive overview about the state-of-the-art 1 Preface 5
in contextualized recommender systems, which introduce context as a new dimension of the user profiles to drive the selection of contents and services. – To enable automatic selection of contents and services, it is necessary to define standards for the characterization of resources, considering their format, length, topics, functionality, etc. Also, normalization is needed in what concerns the data structures and fields to store and manage information about users and context. In Chapter 7, Jesús Bescós et al. enumerate solutions available for these tasks, building upon the expressive capabilities of the MPEG standards. – To finish with the technological foundations of personalization, in Chapter 8, Yolanda Blanco-Fernández et al. look at the internals of the filtering techniques, explaining the ongo- ing shift from syntactic reasoning processes (predominant until very recently) to semantic reasoning ones. The semantic approach yields more powerful means to discover relation- ships between user preferences, contexts, pieces of content and services, which in turn results in greater personalization quality. • Part II provides an overview of the functionalities available nowadays in diverse areas of appli- cation. – The first studies on personalized information services arose in the field of adaptive con- tent retrieval, aimed at providing the users with pieces of content tailored to their evolving interests and needs. Recent advances in this line can be found in Chapter 9 by Nikolaos Lagos and Rossitza M. Setchi, instantiated in a system that deliver personalized technical documentation taking into account user and context characteristics. – Personalized e-commerce services have been around on the Internet for some years. In Chapter 10, Rodolfo Carpintier Santana analyzes those development in the light of the ongo- ing shift to a new way of conceiving the World Wide Web around communities of users (the so-called Web 2.0). As electronic commerce also evolves to embrace new technological plat- forms, Chapter 11 by George Lekakos reports the results of experiments with personalized advertising in Digital TV. – Personalized learning systems developed hand in hand with adaptive content retrieval from the late 1990s, with the common goal of delivering the best suited educational activities to the users. Notwithstanding, the same goal is pursued in very different ways depending on the devices employed to interact with the learning material. In this regard, Chapter 12 by Rosa M. Carro explores the perspective of learning through Internet-enabled personal computers (e- learning); Chapter 13 by Marta Rey-López et al. focuses on Digital TV settings (t-learning); and Chapter 14 by Mario Muñoz Organero et al. finally deals with the learning opportunities enabled by mobile devices (m-learning). – The need for personalization in e-government arises from the fact that users, especially those with little web experience, often get lost in the information space of a portal and need specific hints that are usually provided in an administrative building. In Chapter 15, Ljiljana and Nenad Stojanovic present a novel approach to create personalized e-government portals harnessing the power of semantic reasoning technologies. – Personalized healthcare services aim at supporting the users’ well-being when they are healthy and their medical treatments when they are sick, taking into account the individ- ual’s unique biological, social and cultural characteristics. To help realize the wide range of technologies and applications involved, Mikhail Simonov describes in Chapter 16 a number of use cases of personalized healthcare: prevention, alert service, care and nutrition, food, informed consent, and home living and elderly. – Personalized is also envisaged to play a role in the networked home of the future, which may become the central point for all of a user’s interactions with information services and devices. 6
In Chapter 17, Rebeca P. Díaz-Redondo et al. look at the most relevant issues involved with personalization in domotic environments, with a semantic approach to orchestrating the many appliances that may be installed in a house. – One of the fields that is concentrating more attention nowadays is that of ubiquitous com- puting (also referred to as “pervasive computing” or “ambient intelligence”), which consists of embedding context-aware information processing into everyday objects and activities, also out of the home. To gain insight into the development and deployment of ubiquitous and personalized services, Chapter 18 by Bharat Bhushan et al. explains relevant issues of life- cycle management, whereas Chapter 19 by Carlos García-Rubio and Celeste Campo looks at the important tasks of service discovery. Finally, in Chapter 20, Vassileios Tsetsos et al. present an application of these technologies plus a core of semantic modeling and reasoning in pedestrian wayfinding and location-based services. • To finish, Part III of the book provides a discussion of the most important open problems and challenges in personalization research overall, identified jointly by the editors and some of the chapter authors. Overall, the technical level of the book makes it suitable for third-year or posterior undergraduate courses on information systems, and also for graduate ones. Notwithstanding, we believe it should be appealing for any one involved with the design and development of information services, in any of the many applications and technological platforms considered.
Hope you will have a nice read!
The editors
José Juan Pazos Arias Carlos Delgado Kloos Martín López Nores Part I
Legal and Technological Foundations
2 Legal Framework for Personalization-Based Business Models
Teresa Rodríguez de las Heras Ballell
Universidad Carlos III de Madrid [email protected]
Summary. Unlike mainstream belief, Internet does not embody a global marketplace providing uniform, stan- dardized and massively produced goods and services intended to attend a purported universal and undistin- guished demand. Contrarily, Internet offers an unthinkable range of personalizing possibilities. Even more, success in Internet appears today not to repose on the “dimensions” of the demand yet, but to depend on the “precision” of the offer. In contrast with its rival mass media largely assuming passive users, Internet does not only encourage user interaction, but it is indeed nourished by contents, data, information inferred from behav- iors, searches, responses performed by its own users acting as prosumers (neologism coming from a combina- tion of consumer and producer). Companies have become aware of the extraordinary power of such an input data flow to model access profiles, to tailor information items with regard to access devices or applications, to recommend goods and services according to users’ interests and preferences, to place ads closely connected to search results, emails texts or prior consultations, to set up specialized business platforms (e-marketplaces) suited to members’ needs. Such strategies, fuelled by sensitive personal data and behavioral information, are enough to evoke certain legal concerns. Business models based on personalization strategies, recommender sys- tems or context-aware services encounter delicate privacy issues. Assessment of such privacy concerns entails the prior understanding of legal infrastructure enabling personalization within electronic relationships frame- work. The aim of this paper is to explore the most noteworthy legal issues arising from personalization strate- gies. The first step is to gain insight of how legal instruments help business to manage personalization in so-called “age of access”, where products and services are defined indeed not as close items but as open “set of uses” (functionalities) in accordance to agreed license terms. License agreements embody an appealing inherent paradox as far as they symbolize the struggle for combining the standard form contract logic of massive trade with the most demanding requirements for personalizing. Contractual terms of use are a key legal tool to rule users browsing web site, to govern participants accessing e-marketplaces, to tailor access profiles to contents, to conduct recommendation processes. Several legal concerns are aroused by use licenses as personalization managing devices; among them, certain challenging issues regarding privacy. A thin line separates reasonable personalizing strategies from unlawful invasion of users’ privacy. Accordingly, the paper next recreates several scenarios illustrating the unsettled relationship between personalization and privacy protection and traces key principles and rules to manage this tough conflict.
2.1 Personalization in the Digital Economy
The advent of digital world, settled on the infrastructure provided by Internet, has turned information into the life blood of global economy that struggles to trigger its metamorphosis to knowledge. Information processing embodies the keys of the digital economy’s success, but at the same time conceals all the risks attributed to new technologies. 10 Teresa Rodríguez de las Heras Ballell
Unlike mainstream belief, Internet does not embody a global even space (marketspace) providing uniform, standardized and massively produced goods and services intended to attend a purported universal and undis- tinguished demand. Contrarily, Internet offers an unthinkable range of personalizing possibilities. Even more, success in Internet appears today not to repose on the "dimensions" of the demand yet, but to depend on the “precision” of the offer. Collecting, managing and transferring information are the critical tasks supporting the provision of value-added services and the devising of successful personalization-based business strategies. Companies have become aware of the extraordinary power of such data flow to model access profiles, to tailor information items with regard to access devices or applications, to recommend goods and services accord- ing to users’ interests and preferences, to place ads closely connected to search results, emails texts or prior consultations, to set up specialized business platforms (e-marketplaces) suited to members’ needs. Nonethe- less, information manipulation arouses, at the same time, delicate privacy issues. Accordingly, such strategies, launched by personalization-based business models, recommender systems or context-aware service providers, as for fuelled by sensitive personal data and behavioral information, are enough to evoke certain legal concerns. Most representative actors of digital economy, such as Google, have embarked on a personalizing strategies race. Google, the world’s biggest search engine, collects data from users’ searches to provide the best possible search for the user, improve its value-added services, sharpen ads’ target, tailor contents according to users’ preferences and develop new services and products in users’ benefit. In the wild new world of Internet, where individuals fight not to be drowned in the flood of information lacking of credibility references, personalizing strategies, recommender systems and context-aware services are unquestionably the most value-added business offers. Nevertheless, recent declarations of intent to offer new sophisticated personalized services have been received with mistrust and certain alarmism from diverse opposing constituencies. Google mission statement condensed in the sentence “the goal is to enable users to be able to ask the question such as “What shall I do tomorrow?” and “What job shall I take?” has aroused the most catastrophist concerns of an Orwellian vision. Concerns that have been aggravated by Google bid for DoubleClick, a company that helps to place and track on- line advertisements through sophisticated targeting capabilities. To reinforced privacy worries, complaints from its nearest competitors about the risk of the deal to damage fair competition have joined. Meanwhile, European data protection and competition watchdogs, in contrast with other alarming voices, have been nonetheless more cautious. On the one hand, Article 29 Data Protection Working Party has expressly revealed its concern just as regards information storage period policy ruling Goggle services. On the other hand, European Commission is investigating eventual antitrust considerations of envisaged Google-DoubleClick merger to decide to clear the deal or to open an in-depth probe. No privacy inquiry is nevertheless hinted thereupon. Even if it is undeniable that delicate issues are involved, perception of personalization strategies seems to be too influenced by alarmist visions on the scope of privacy right and the threats that new technologies impose thereupon. Hence, a careful analysis, firstly, of the scope and constituent elements of Data Protection and Privacy regulations and, secondly, of the extent of personalization and the mechanisms supporting such business strategies is due in order to readjust the discussion framework. Against such a backdrop, this chapter aims at exploring the most noteworthy legal concerns arising from personalization-based business strategies in the digital space. The starting hypothesis is that, insofar credibility and attention are the scarcest resources in the digital economy, recommender systems and trusted third par- ties provide extremely appealing and welcome value-added services. Personalization-based business models create value. At the same time, such strategies may imply risks to be efficiently manage. Despite the main- stream perception, any information collecting and processing do not entail the encroaching on privacy spheres insofar as the principle of consent is regarded. A thin line separates reasonable personalizing strategies from unlawful invasion of users’ privacy, but in fact it is a matter of limits. Privacy does not mean anonymity but control on personal data. Section 2.2 of this chapter is devoted to define and distinguish these blurred notions of personalization, privacy and anonymity. On the foundations of such distinction, Section 2.3 aims to trace the guidelines of data protection and privacy regulation. In conformity with such legal background, personalizing and recommending business strategies must be assessed. The first step is to gain insight of how companies implement such strategies considering the diversity of activities nourishing digital economy. Section 2.4 does then explore the range of business models competing in the market in order to find out how personalization aims are devised. From such a market overview, it is inferred that license agreements are the main legal devices applied by business to manage personalization in so-called “age of access”, where products and services are de- 2 Legal Framework for Personalization-Based Business Models 11
fined indeed not as close items but as open “set of uses” (functionalities) in accordance to agreed license terms. Contractual terms of use are key legal tool to rule users browsing web site, to govern participants accessing e-marketplaces, to tailor access profiles to contents, to conduct recommendation processes. Basics of license agreements, most challenging issues arising from the formation process and proposals of guidelines to ensure valid and enforceable contracts are tackled in Section 2.5 of the chapter. An underlying conclusion permeates the whole reasoning. New technologies play as regards the matter in hand two and paradoxically opposing roles. From the perspective of a harming action, network effects turn new technologies into pernicious multiplier of damages or potential victims. Such an undesirable risk is aggravated by the difficulties commonly associated to digital space in order to achieve an efficient and enforceable regula- tion thereof. Internet, a new space lacking of centralized supervisor and seasoned by a feasible anonymity and a natural delocalization, is then perceived as a dangerous receipt for impunity and anarchy. Regulation, either initiatives on international-basis, domestic legislation or self-regulation (codes of conducts), is nevertheless possible and viable. Uncertainty provoked by delocalization in digital space may be overcome by redefining territory-based connecting factors to decide the law to apply to situations taking place in a new space. Verifying such available solutions renders a calming effect that leads us to the other face of new technologies. From the perspective of control possibilities, new technologies enable unimaginable capabilities of monitoring, tracking and ruling conducts. Everything in the digital space is represented by traceable data, behaviors, movements, actions, searches, decisions (transactional data). Although these controlling functions may be exploited in an Orwellian manner1 (Echelon and Carnivore are patent evidences), new technologies truly hold a critical and powerful position within available regulation techniques in the digital space. Technological architecture assists the enforcement process of law and contract in the new space. Technology is indeed an allied actor to tradi- tional regulating mechanisms (market rules, social norms and legislation). Access to an e-marketplace may be controlled by implementing passwords, electronic signatures or biometric identification devices; complying with digital auction rules may be ensured by designing an interface allowing only the option among technically available actions; eventual copyright infringements may be, not only discouraged but also, and above all, pre- vented by rendering technically impossible the implementation of certain violating actions; the enforcement of a software license may be achieved by deactivating or suspending the software in case that the user breaches contractual obligations incumbent on him or her; even privacy itself may be protected with the assistance of a software supporting the automatic compliance of the privacy level preselected by the user (P3P, the Platform for Privacy Preferences). Hence, rather than assuming an avoidable and permanent danger of privacy under new technologies’ reign, privacy enhancing technologies offer a totally different perspective over the problem.
2.2 Personalization, Privacy and Anonymity
Personalization is a value-creating business strategy. Assuming such a premise, the immediate question is why it does then arouse so controversial issues. Business models based on personalized services and recommender systems need to collect and process sensitive data, and they are indeed fuelled by wide-ranging information concerning users’ habits and preferences supplied by searching records, browsing behavior or profiling devised by invisible data-capturing devices (“cookies”). In a very simplistic manner, it may be upheld that collecting too much information for personalizing purposes may imply an unlawful invasion of privacy, whereas an abso- lute anonymity prevents business from providing personalized services. Such an alternative, in the way it was depicted, leads to an inefficient and undesirable result. Rethinking the involved concepts renders the discussion more reasonable. Privacy, rather than anonymity, has to be interpreted as a right to control our own data and any information related to our preferences, behaviors and interests. It is indeed a question of limits; it is a matter of control. Further precision in the definition of the three concepts concerned may be helpful to unravel the tricky interrelation among them. Firstly, personalization does not name a unique and homogeneous business strategy, but encompasses diverse business models ranging from internet access providers, servers, browsers or searchers to content 1 Rodotà [2004] suggests very skillfully that such a society under surveillance evokes the “Panopticon” of Jeremy Bentham rather than the “Big Brother” of George Orwell. 12 Teresa Rodríguez de las Heras Ballell providers, a large number of varied on-line retail shops or business-to-business e-marketplaces. As detailed below under Section 2.4, wide-ranging formulas competing in the market share personalization as business mission. In addition to a variety of customizing techniques, a clear line sharply distinguishes customizing models from recommender systems. Even if both business strategies are based in tailoring the service or the product being aware of user’s needs, surroundings circumstances or location, the latter provide a value-added service, the recommendation. Whereas the former merely implies the provision of the user with well-tailored information, the latter entails the enhancement of the supply of information with the act of recommending. Recommendation value depends on the reliability of recommender. Information may be true or wrong, accurate or inexact; nevertheless qualifying a recommendation requires using a wider range of grey. Provider’s position is different in each case. Both strategies run on data collected by diverse devices in the course of business rela- tionships. Customization demands a large amount of detailed data about the user, recommendation commonly requires crossing data of a cluster to identify similar users’ profiles. Secondly, privacy is a broader concept than personal data protection [Ballesteros Moffa, 2005, Bergadamo et al., 2005, Kuner, 2007, Herrán Ortiz, 2003, Lisi, 2003, Masimini, 2002]. Even if collecting and processing personal data may involve the encroaching on the data subject’s private sphere, other kind of privacy invading actions does not always concern the manipulation of personal data but the breach of security systems, confi- dentiality standards or unsolicited commercial communications rules. Albeit the mainstream use of “privacy policy” in Internet, the right to respect for the private and family life, home and communications and the right to the protection of the personal data are separately enshrined by, among other texts, the Charter of Funda- mental Rights of European Union in Article 7 and 8 respectively. Such an independent treatment endorses the distinction upheld above, but its relevance dilutes when specific legislation on protecting privacy is analyzed. From the perspective of the privacy infringement, both risks are nonetheless tackled jointly by regulation ap- plicable to business in digital space. Thus, within the European market, Directive 2002/58/EC of the European Parliament and of the Council, dated on 12th July, concerning the processing of personal data and the pro- tection of privacy in the electronic communications sector (hereinafter, Directive on privacy and electronic communications)2 approaches the problem from the view of the risks created by electronic communications, either entailing the invasion of privacy or an unlawful manipulation of personal data. Any attempt to abstractly define the concept of privacy is a cumbersome task that will surely lead to an open notion of blurred outlines. Addressing the meaning of privacy from the perspective of the potential dan- gers that it must face in the digital environment guides to a more useful and pragmatic result. A realistic and perceptive look at the rapid evolvement of electronic communications has enabled to identify certain scenarios of risk. Firstly, a breach of the security of the network may imply serious risks for users’ privacy, therefore ser- vice providers must take appropriate measures to safeguard the security of their services and inform users and subscribers of any special risks thereby and of measures they can take to protect the security of their communi- cations. Secondly, unauthorized access to communications may involve a grave attack to the confidentiality of both contents and any data related to such communications, accordingly, measures should be taken to ensure that confidentiality remains guaranteed in the course of business practice and during the data storage period. Thirdly, certain devices such as so-called spyware, web bugs, hidden identifiers and other similar ones can enter the users’ terminal equipment without their knowledge, gain access to information stored thereupon or trace users’ activities with severe impact on the private sphere of users; consequently, the use of such devices should be allowed only for legitimate purposes and upon the knowledge of the users concerned. Fourthly, traffic data processed and stored by providers contain information on the private life of natural persons and concern the right to respect for their correspondence or concern the legitimate interests of legal persons; hence, they should be erased or made anonymous when they are no longer needed for the purpose of the transmission of a communication of other legitimate purposes (billing, marketing). Fifthly, location data giving the geographic position of the terminal equipment of the mobile user in order to enable the transmission of communications or provide value-added services such as individualized traffic information and guidance to drivers run on data concerning the private sphere of users, so the processing of such data for providing certain services should require the user’s consent. To put an end to this long list of potential risk situations on privacy, it might well be worth alluding to unsolicited communications for direct marketing purposes (automatic calling machines, telefaxes, e-mails or SMS messages) that may represent an additional threat to users’ privacy.
2 O.J. L 201, dated on 31st July 2002. 2 Legal Framework for Personalization-Based Business Models 13
Despite of the first impression that such a lengthy relation of menacing situations may arouse, privacy and personal data protection do not always prevail over any other right, interest or public policy principle in the digital space and, accordingly, certain limits may be imposed thereupon provided that a legitimate purpose is to achieved. In that regard, it is not a question of prohibition, but of balance. Privacy is then a right to control one’s sensitive data. In conformity with such perspective, regulation on protecting privacy and personal data does not intend to ban every collecting or processing action above-mentioned, but establishing the rules allowing users to control their own data instead (see infra Section 2.3). Consistency of such reasoning leads us to affirm that privacy does not mean a right of full anonymity. In other words, privacy and anonymity are to reconcile with other public policy objectives. On the one hand, the possibility of remaining anonymous is essential if the fundamental rights to privacy and freedom of expression are to be maintained in the digital space. On the other hand, the ability to participate in the digital world without revealing one’s identity challenges the viability and the enforceability of other key areas of public policy such as the fight against illegal and harmful content, financial fraud, or copyright infringements3.
2.3 Privacy and Personal Data Protection Keys
Prior to go into the analysis of the personalization-based business models and the potential risks on privacy arising therefrom, it might be well worth tracing the basics of regulation on protecting privacy and personal data with special reference to electronic communications. Such an overview is expected to provide the further discussion with a benchmark to test compliance level of business. The first obstacle that regulation on privacy protection in Internet must face is the territoriality problem. Decentralized structure and a-national character are the most distinctive features of a new space where tradi- tional principles determining applicable law on territory-basis appear to become impracticable. Drafting and agreeing an international standard on personal data and privacy protection would be truly the most convincing and efficient solution. Nevertheless, albeit debate is pointing at such a harmonizing process, existing regulation initiatives still rest on a less ambitious stage. Certain legal instruments, adopted by diverse formulating agen- cies (Organization for Economic and Cooperation Development (OECD); Council of Europe, United Nations General Assembly), have achieved to enshrine widely accepted common core principles on protecting privacy that might facilitate the progressive rapprochement on international-basis. Despite the reasonability of such common approach and its suitability for tackling problems arising from and in the new space, uniformity aim on privacy issues on Internet has been intensely marked by a strong confrontation between protecting legal models, in particular the one staged by the laxer model championed by United States and the stricter one advo- cated by Europe. In the interests of achieving better understanding between the two legal systems and ensuring competitiveness in the digital economy, a bilateral agreement, the Safe Harbor Agreement, was at last reached in a middle ground form. The safe harbor, approved by the EU Commission Decision4 in 2000, is an important way for US companies to avoid experiencing interruptions in their business dealings with the EU or facing prosecution by European authorities under European privacy laws. Certifying to the safe harbor will assure that EU organizations know that a company provides "adequate" privacy protection, as defined by the European Directives. With such a compromise as a backdrop, a mainstream understanding on worldwide privacy pro- tection is conceivable whose keys are to be traced below. Regulation on privacy issues on Internet turn into a complex mechanism containing legal initiative of diverse geographical scope complemented by self-regulation and assisted by standards on privacy-enhancing technologies. Back to the analytical look at privacy and personal data regulation, three main components will be traced. First of all, scope of protection is to be delimited. In this regard, the meaning and extent of personal data notion hold a central position. Secondly, basic operating principles on privacy protection are to be related. Thirdly, diverse available techniques to enhance privacy and rule personal data issues in the course of business are identified. 3 Recommendation 3/97, “Anonymity on the Internet”, adopted by the Working Party on the Protection of Individuals with regard to the Processing of Personal Data on 3rd December 1997, XV D/5022/97 final. 4 EU Commission Decision 2000/520/EC, dated on 26th July 2000, published in O. J. L 215/7 of 25.8.2000. 14 Teresa Rodríguez de las Heras Ballell
2.3.1 Personal Data
A wide notion of personal data has been upheld, in particular by European lawmaker, in order to ensure a satis- factory scope of protection covering thereby all information which may be linked to an individual5. Legislative process distilled therefrom the current definition of personal data contained in Directive 95/46/EC with the following wording: “Personal data shall mean any information relating to an identified or identifiable natural person (...)”. Four main defining elements build the definition: “any information”, “relating to”, “an identified or identifiable”, and “natural person”. The term “any information” clearly signals the willingness of the legislator to call for a wide interpretation of personal data including information of any nature —descriptive or containing opinions or assessments—; of any sort —concerning private and family life or working relations, economic activities, and social behaviours— ; expressed in any format and stored in any medium. Information can be deemed to “relate” to an individual when it is about that individual. Establishing that link is crucial to define the scope of protection. Such a link may emerge in three possible situations: when the information is about an individual (content); when the data are used or likely to be used with the purpose to evaluate or treat in a certain way or influence the status or behaviour of an individual (purpose); when the data are likely to have an impact on a certain person’s rights and interests, taking into account all the circumstances surrounding the precise case. Information becomes sensitive for privacy purposes as far as it is related to a natural person that is identified or may be identifiable by means of certain identifiers taking into account all the factors at stake. While identifi- cation through the name is the most common occurrence in practice, other identifiers may be successfully used to single someone out. IP and e-mail addresses, or log-ins have been considered data relating to an identifiable person under certain circumstances, although such an assertion expounded in general and absolute terms is not undisputed [Guerrero Picó, 2006]. Thus, in case that identification of the data subject is not intended by the processor, appropriate state-of-the-art technical and organizational measures to prevent identification should be taken to protect the data against identification. Moreover, data must be related to a natural person. Such condition, apart from raising appealing ques- tions regarding the treatment as personal data of information about legal persons, deceased persons or unborn children, evokes a challenging matter rooted in the idea of man-machine relationship. Certain sorts of data collected in the course of the service provision are truly related to the computer, terminal equipment or access system. In addition to pure personal data (such as contact data supplied by an application form), traffic data are a valuable source to monitor user’s actions, interests and on-line behaviour, but they are purportedly identifying the machine not the individual. Ascertaining that triggers an exciting reflection about man-machine interaction. Electronic devices are today becoming true prosthesis of human being; artificial parts that, on permanent and proximity basis amplify, complement or assist individual’s actions in the digital space. These kinds of data, un- less dissociated from a person’s identity, may conduct to devise a profile including users’ interests, preferences, behaviours and private environment. Traffic data and location data are valuable assets for commercial purposes and highly prized input to fuel public control processes.
2.3.2 Principles Relating to Data Quality
Under the general premise of data minimization, data collecting, processing6 and transfer are subject to the following principles:
Fair and lawful data collecting means. • 5 COM (90) 314 final, dated on 13th September 1990, p. 19 (Commentary on Article 2). 6 Article 2 (b) of Personal Data Directive defines “processing of personal data” for the purposes of the Di- rective as “any operation or set of operations which is performed upon personal data, whether or not by automatic means, such as collection, recoding, organization, storage, adaptation or alteration, retrieval, con- sultation, use, disclosure by transmission, dissemination or otherwise making available, alignment or com- bination, blocking, erasure or destruction”. 2 Legal Framework for Personalization-Based Business Models 15
Specified, explicit and legitimate purposes, prohibiting any further data processing in a way incompatible • with those purposes or for different purposes. Adequacy, relevancy and minimization of data in relation to the purposes for which they are collected • and/or further processed. Accuracy and up-to-date state of data. • Complying with such principles implies two further conditions. On the one hand, an extensive duty to inform the subject concerned about all abovementioned particulars of the data manipulation; on the other hand, and presuming the latter condition of transparency is fulfilled, the knowledge or consent of the data subject to ensure that the processing carried out is lawful. The data subject’s consent shall mean “any freely given specific and informed indication of his wishes by which the data subject signifies his agreement to personal data relating to him processed” 7. Since the consent principle underpin the whole legal protecting system, invisible devices capturing data in Internet (spyware, adware, keystroke loggers, Javascripts, cookies, and so on) arouse serious privacy concerns. Technical functions for facilitating network operation and client-server transactions or legitimate purposes for providing value-added services within the framework of a commercial relationship initiated by the user can be proved. Accordingly, a lawful use of such devices must be based on a transparency policy intending to duly inform the user thereabout and give the opportunity to refuse their installation or deactivate their capturing functions.
2.3.3 Techniques: Licenses and Private-Enhancing Technologies
Uncertainty about applicable law in Internet has encouraged the emergence of two phenomena to alleviate distorting effect of delocalization: self-regulation and self-help. Self-regulation has shaped the form of codes of conducts and the form of license agreements. By means of those self-regulation instruments, companies adopt a privacy policy intended for ruling its activity in relation to users’ privacy and ensuring the full knowledge of the user about personal data processing applied by the company. As a powerful complement of regulation, companies and users may install self-help technical measures to support enforcement or privacy-enhancing technologies to improve protection level on personal data.
2.4 Personalization-Based Business Models
Digital economy business models can be typified in conformity with three basic strategies corresponding to three functional and activity layers. Such three-layer structure8 captures a complex conception of the digital space assuming a reciprocal complementing interaction among the three constituent parts: infrastructure, con- tent and community. Taxonomy of models proposed to clear the tangled business scene is based on identifying on which layer functions performed and activities carried on by companies are located.
7 Article 2 (h), Directive 95/46/EC of the European Parliament and of the Council, dated on 24th October 1995, on the protection of individuals with regard to the processing of personal data and on the free move- ment of such data, published in O.J. L 281/31, dated on 23rd November 1995 (hereinafter, Personal Data Directive). 8 A structure composed of three layers is not unknown within scientific literature, the parts selected to make the complex mechanism up do nevertheless differs from those proposed therein. In this regard, Benkler [2000] defines the whole communication system in accordance to three variables or layers that are strongly interconnected but independent. Following the so-called “layer approach”, in every communication three layers entwine: a physical layer (technical infrastructure composed of access systems, networks and nodes), a logical layer (software operating network functioning) and content layer (data and information). Such approach has been rescued and redefined by Lessig [2002]. 16 Teresa Rodríguez de las Heras Ballell
2.4.1 A Three-Layer Structure to Understand Digital Business Models
The infrastructure layer embraces all the functions, products and services rendering network functioning pos- sible from a technical perspective. Access providers, hosting services providers or caching services providers are the major figures playing in this instrumental stage. Insofar as their function is basically instrumental, their activity is fuelled by data not by contents, since technical services providers are devoid of ability to control and capacity to actually know the information that they transmit (mere conduit), host (hosting) or automatically and temporarily store (caching) at the request of a recipient of the service. The content layer accommodates the largest and best known Internet services covering the major part of content providers (websites, portal) and wide ranging sorts of online businesses. Content is the blood flowing through networks and nodes of the digital space’s infrastructure. Hence, content is intended to encompass any kind of information, product or services supplied on Internet or offered and provided by means thereof. Business creativity has been able to recreate on the digital stage every brick-and-mortar strategy and even open new spaces for business specifically devised for the digital world. Any attempt to describe and enumerate all on-line initiatives will suffer from an unmanageable casuism ruining any useful classification. Access to contents, supply of information, provision of services, sales and deliver of goods perfectly the fit business models competing at the content layer. Community is the most meaningful term to capture the essence of the third layer constituting the digital space. Virtual community creating models are the most sophisticated and promising business strategies. In fact, the ability to add value by exploiting community features on Internet opens challenging opportunities for social and commercial purposes. Creating communities is in itself a powerful personalizing project. A community is based on three pillars [Harvard Law Review Association, 1999]: unity of purpose emerging from a many-to-many interaction; sense of membership distilled in an enter/exit policy; and common rules governing relationships among members. Despite their origin as mechanism of social aggregation, virtual communities are skilful strategies for “networking and building business relationships” [Kannan et al., 2000]. The so-called “community factor” pervades a wide range of Internet-based relational models [Dai and Kauffman, 2002] from the most basic social networks shaped as chats, fora, bulletin boards or newsletter to the extremely sophisticated collaborative commerce projects distilled in e-marketplaces, virtual auctions, e-procurement platforms, virtual alliances or highly integral virtual business communities [Rodríguez de las Heras Ballell, 2006a]. The key factor of such networking models lies in their ability to engender, propagate and consolidate trust in market relationships. Assuming that credibility is a scarce resource in digital economy, trust-creating strategies are massively demanded by users as critical value-added services to guide their browsing and surfing.
2.4.2 Virtual Communities and Trusted Third Parties
The third category of digital business models just proposed, belonging to the layer named “community”, arouses certainly the most suggestive issues. Exploring community-based strategies leads us to the appealing concept of Trusted Third Party that distinctly vertebrates digital economy functioning and stems from the challenging versions of social-networking. Hence, theses community-creating business models are worth further analyzing to identify personalizing strategies. Three approaches are required to gain full insight of structure and functioning of community-based business models: technological, business and legal approach. From a technological perspective, a community-based business model must be necessarily built as a closed electronic environment. Unlike an open electronic environment, a closed one provides controlling capacities over membership, access, exit, members’ behaviour, content quality, security standard, data and information flow. Within the borderless space of Internet, worldwide open electronic environment par excellence, or beyond it, spaces can be closed either by physical means (virtual private networks) or by using passwords, user names or other sorts of identification mechanisms. Physical or logic closure provides control and control reduces uncertainty, insecurity and privacy risks. From a business perspective, virtual business communities would symbolise an efficient strategy of coope- tition, a neologism stemming from a fruitful combination of cooperation and competition. It would be quite 2 Legal Framework for Personalization-Based Business Models 17 reasonable to define them as the digital version of traditional fairs and market places9. Coopetition strategies are pushed by important efficiency gains achieved by the application of new technologies (electronic commu- nications effect, electronic brokerage effect and electronic integration effect [Malone et al., 1987]). Any form of multilateral electronic platform devised for trading purposes acts as an intermediary in eco- nomic sense repairing market failures by carrying on its inherent functions: aggregation, credibility, many-to- many interaction, centralized management, synergies and economies of scales. Most powerful strengths associ- ated to community-based business models are then the gains obtained from multilateral interaction managed by a centralized supervisor who, by exerting its controlling functions over the market and its participants, generates credibility and trust. E-marketplace, trading platform or virtual business community manager performs roles of a Trusted Third Party. When it controls the access to the closed environment, firstly, by selecting suitable members among applicants, and secondly, implementing a reliable access control mechanism using passwords, digital signatures or any other identification procedures, the community manager grants a sort of trust mark in benefit of the admitted participants. Third persons may rely on such trust mark insofar as the manager credibil- ity would be solidly underpinned by expertise, trustworthiness and reliability. Provided that offering, marketing and trading in the market space is ruled by the Rules Book drafted by the manager and expressly accepted by participants, the marketplace is an area of certainty regarding applicable regulation and common rules of conduct. Additionally, to the extent that the manager is performing its supervising and sanctioning functions, market participants rely on the compliance of Rules Book and the effective discouragement of deviating ac- tions within the trading scenario. Exploiting community features entails the promotion of trust in transactions and allocates Trusted Third Party’s roles on the e-marketplace’s manager, the Internet forum’s moderator, the on-line auction’s auctioneer or the virtual community’s administrator. From a legal perspective, community-shaped spaces under a central manager’s control are self-regulated environment. Structure, organization and relationships within the market are contract-based. Such a condition insulates community against the disconcerting effects arising from delocalization. Applicable legislation and internal rules are agreed by parties, provided that they are not consumers, otherwise compulsory law may annuls certain provisions derogating from protecting principles in weakest parties’ interest. Apart from those consumer-biased exceptions, business-to-business communities are underpinned by two sorts of contract-based relationships. Within the so-called intersystem dimension, communities in operation compete in the market to attract new participants. Applications of those interested in joining the community are submitted to a scrutiny prior to decide their admission or refuse their incorporation. In case of favourable result, applicants enter into a membership contract regulating their access to the system, the use of common services and their behaviour therein. Once inside the market, each competing or cooperating relationships among members are based on a contract of the pertinent nature (sales contract, hiring agreement, software licence, technology transfer, and so on). Thus, the community assembles two kinds of relationships: vertical agreements between the admitted applicants and the manager and horizontal agreements among the market participants.
2.4.3 The Logic of Personalization-Based Business Models
All business models above described are apt to implement personalizing strategies. Nevertheless, it would be wrong to think that any data manipulation or each attempt to adapt the service or the product to the user entails the launch of personalization initiative. Upon an analytical view, personalization results from completing three successive stages: controlling, profiling and personalizing. Personalization assumes prior controlling functions over data and relevant information; otherwise, person- alizing is not feasible. Merely controlling data does nevertheless not imply implementing a personalization- based strategy. Profiling abilities are decisive therefor. Comparing new collected data with existing profiles and crossing available information cobble the way to personalization. Should this three-stage process be placed on the business models’ taxonomy, meaningful results would be reached. Technical intermediaries, providing infrastructure and related services, collect and store in the course of their business personal data, traffic data and location data. Such information allows them to adapt their services 9 “Although it is Internet´s technology that makes business-to-business exchanges possible, the exchanges themselves are primarily business applications and not technological innovations” [Sculley and Woods, 2001] 18 Teresa Rodríguez de las Heras Ballell to context, technical characteristics of the accessing device, language used by the terminal equipment, installed operating system and even surrounding circumstances or user’s mood. In case that collected and processed data are related to an identified or identifiable person, provider will abide by personal data processing obligations. Their duties over collected data will be lighter if a dissociation process is applied to successfully break the connecting link between data and identifiable users. Technical services providers do then control data but do not carry on real personalization strategies. They merely provide alternative versions in conformity with certain predefined requirements. To the extent that technical intermediaries carried on controlling task over traffic and location data, certain obligations are imposed thereupon. For the purpose of criminal investigations or public defence policy, providers must retain location and traffic data during a predetermined period. Content providers of any kind are able to reach the second stage. Building profiles means a powerful strategy to sharpen the target. Personalizing access, tailoring contents, customizing online shop or providing selected value-added services are successful strategies to catch and lock-in users in the digital space rescuing them from the overwhelming sea of anonymity represented by the Web. Personalizing access, contents, products or value-added services requires, firstly, the capacity to control and process data and, secondly, the devising of profiles related to identified or identifiable persons. The whole personal data burdens are loaded on content providers. Accordingly, transparency duties, legitimate purpose principle and express consent rule must guide providers’ activity in their relations with users/subscribers/clients. If the three parts composing personal data regulation are assembled, it is quite easy to design an acceptable standard of conduct for content providers performing personalizing functions. Firstly, any invisible device intended for collecting data from users’ terminals must be avoided. Clear and full information regarding the use of those devices have to be supplied to users and an easy way to refuse their installation must be offered to users any time. Secondly, users have to be informed by the legitimate purposes of data collection (marketing, customization, security, value-added services provision, and so on). Thirdly, a reliable mechanism entitling the user to give a free consent must be enabled. The debate regarding opt-in/opt-out standard captures the concern of ensuring a real, well-informed and free consent giving. Whereas opt-in standard to obtain consent is stricter and shields users from unconscious supply of personal data, opt-out standard is endorsed by more flexibility and agility aims. The compliance vehicle of all expounded rules is the so-called “terms of use” included in the use licence agreement predisposed by provider. Further analysis thereof is traced in the next part of this chapter. Community-creating and -managing business models perform personalizing strategies in a very similar way to content providers. Nevertheless, unlike the latter, community-based business enjoy additional abilities to skilfully manage personalization. On the one hand, the extent and the depth of the knowledge about their members’ interests, preferences and behaviour is highly larger than the ones never reached by content provid- ing structures. Creating a community is indeed fuelled by managing precise data about its members. On the other hand, due to the closed nature of community-based business models, they benefit from an express and conscious acceptance by clicking or logging of any member concerning privacy policy, terms of use and rules of conduct. A self-regulated and contract-based environment enables a full and rigorous compliance of personal data regulations. Moreover, social- or business-networking enables to implement recommender-based marketing strategies through so-called “opinion leaders” and “social/business interactions”. In order to safeguard their privacy, users are entitled to opt out of sharing their information. Social-networking do not obviously offer anonymity, per- sonal data and information related to interests, preferences or hobbies are the sap of the vascular system of a community. By means of membership agreements and common policies, community equips its members with controlling devices over their data.
2.5 Legal Mechanisms to Implement Personalizing Strategies
The specific aim of this section of the chapter is to gain insight of how legal instruments help business to manage personalization. In a new space marked by the delocalization effect over traditional applicable law rules, agreements have reached a leading position to rule and manage private interests. Agreements’ power as 2 Legal Framework for Personalization-Based Business Models 19 regulation source is notably ample in business-to-business transactions where autonomy principle pervades. Hence, a B2B relationships’ environment will be assumed to test our reasoning. Prior to further enquire into legal issues arising from the use of license agreements to rule privacy matters, a suggestive consideration may be posed. The advent of digital world has entailed the passing from the age of property to the so-called “age of access” [Rifkin, 2000]. Gaining access is worthier than possessing. This change of paradigm means that products and services are defined indeed not as close items but as open “set of uses” (functionalities) in accordance to agreed license terms [Rodríguez de las Heras Ballell, 2006b]. The real value of the contract does not stem from the product or the service concerned in itself but from what one is entitled to do with or in respect of it. A music file is not a close and unique asset; on the contrary, it embodies a range of “products” depending on the available uses agreed by the parties (copy, reproduce, rent, transmit, publicly display, modify, distribute, and so on). It may be upheld that the contract is indeed the product. License agreements, site licenses or simply terms of use include provisions concerning collecting, process- ing and transferring data that fuel personalizing strategies. Privacy issues do nevertheless use to be separately dealt with under the so-called privacy policy. The acceptance of the privacy policy is anyway an express provi- sion of the license agreement. License agreements play three functions for privacy purposes: identifying the nature of collected data, fully informing of the legitimate purposes and obtaining express consent from the data subject. Despite the over- whelming success of license agreements in the business praxis, certain legal concerns have been evoked. From both procedural and substantial perspective, the validity and the enforceability of browse-agreements has been riddled with doubts. License agreements embody an appealing inherent paradox as far as they symbolize the struggle for combining the standard form contract logic of massive trade with the most demanding requirements for personalizing. In the line of standard form contract logic, users face an unavoidable alternative (take-it or leave-it) that challenges the validity of consent as well as the fairness of contractual provisions [Rodríguez de las Heras Ballell, 2006c]. As far as the consent-giving process is concerned, site licenses have inherited from their predecessor —the so-called shrink-wrap licences [Einhorn, 1998, Lemley, 1995, Baker, 1997, Puhala, 1985, Kaufman, 1989, Mi- nassian, 1997]— the controversy [Lemley, 1995, Classen, 1996, Harrison, 1999] that permeated their validity and enforceability [Kaufman, 1989, Pitet, 1997, Das, 2002], and furthermore they have had to bear the concerns arising from web agreements. In the digital world, license agreements had adopted three main formats: click-agreements, logging- agreements and browse-wrap agreements. Whereas click-agreements duly fit the way of operating business in opened electronic environment, such as a wide range of transactions entered into with online shops; accepting by logging is associated to closed spaces as e-marketplaces, electronic platform and virtual business commu- nities. In any case, consent-giving procedures are express, well-known (clicking, logging), and satisfactorily prevent users from unconscious actions. Browse-wrap agreements involve sensitive questions as far as the act of acceptance is concerned. Every website we visit has placed at the bottom of the page a link entitled “terms of use”, “legal notice” or similar expression. Under such a link, a license agreement, including the privacy policy, is hiding. All the precautions intended to ensure full knowledge, opportunity to review and free and informed consent are here undermined. The location of the link is not the most suitable, users are not expected to review the provisions and acceptance is not express but presumed from the mere use of the web. They are indeed named “click-free agreements”. Responding to an unfavorable treatment by case law10, all the suggestions are addressed to eliminate existing risks and adopt precautionary measures to ensure the obtaining of free and valid consent —better location of the link, due notice about the existence of the terms, real opportunity to review, warning about the assenting effect of mere use [Kunz et al., 2001]. As regards substantial perspective, a quick look at the license agreements drafted by the most famous online providers conveys quite illuminating ideas about how providers manage personalization and privacy. Reading,
10 A highly irregular case law proves the intense controversy embedded in the admissibility discourse of browse-wrap agreements: Register.com v. Verio Inc., 126 FSupp2d 238 (SDNY 2000); Ticketmaster v. Tick- ets.com, 2000 USDist LEXIS 4553 (CDCal March 27, 2000); Pollstar v. Gigmania Ltd., 170 FSupp2d 974 (EDCal October 17, 2000); Specht v. Netscape, 2002 USApp LEXIS 20714 (2d Cir. October 1, 2002). [Raysman and Brown, 2002, Kunkel, 2002, Grossman, 2001, Hayes, 2003] 20 Teresa Rodríguez de las Heras Ballell for instance, Google services privacy policy, at least four categories of provisions are remarkable for privacy purposes. Firstly, the kind of data collected and processed in the course of the service performance (IP, date and hour, cookie ID, URL, enquire, log) and the envisaged use thereof. Secondly, the purposes that are expected to satisfy by the data processing (improving services, offering value-added services, preventing from spam, storing backup, interacting users, marketing). Thirdly, the remainder that its data transfer policy is framed by applicable law and as far as advertisers are involved, data are facilitated in an aggregated manner. Finally, the provision concerning express consent-giving procedures —combining opt-in and opt-out standards— to cover other purposes differing from those established in the privacy policy.
2.6 Balancing Personalization Strategies and Privacy Concerns
Personalizing ability has become a powerful competitive advantage in the digital economy. Contrasting with the mainstream perception of a global, homogeneous and uniform Web, personalized spaces are emerging around virtual communities, social-networking, or simply customized relationships. Companies have become aware of the intense appeal of personalization-based strategies, recommender systems and context-aware services to make their offers different from competing products and services supplied by their rivals and provide users with appreciated value-added services suited for deploying a lock-in effect. Accordingly, many of them have embarked on ambitious projects to model access profiles, tailor information items, recommend goods and ser- vices within social-networking, sharpen ads’ target, or devise specialized business platforms (e-marketplaces). Effectiveness and competitiveness gains are expected to be high. Personalizing implies the full knowledge of users’ interests, preferences and wishes by collecting, pro- cessing and manipulating a wide range of data concerning personal details, behavior, location, previous ac- tions, and social-/business-networking membership. Therefore, personalization-based strategies, as for fuelled by sensitive personal data, transactional and behavioral information, do inevitably arouse certain legal con- cerns. According to the hypothesis proposed in this chapter, a successful personalization-based strategy entails the fulfillment of three successive phases culminating in the personalizing aim: control, profile and personal- ize. Just mentioned phases do not always process personal data in the sense provided by personal data and privacy regulations and, besides, any personal, traffic and location data processing does not always mean the performance of personalizing strategies. A more careful analysis of business scene is then due. Firstly, basics of regulations on protecting personal data and privacy have been traced. Briefly summarizing, three main constituent parts underpin the legal building on privacy protection: a wide concept of personal data as related to an identified or identifiable natural person; a principle of data minimization in accordance to legitimate and specific purposes; and a predominant position of data subject’s consent. Assembling those three elements, an useful benchmark is built to test business policies concerning manipulation of data and devising of personalization-oriented strategies (opt-in/opt-out standards, use of cookies, consent-giving procedures, nature of data collected and processed in the course of trade, data retention period). Secondly, by exploring in depth business models competing in the digital space, taxonomy has been pro- posed in accordance to the layers composing the Web: physical layer, content layer and community layer. Such a three-layer vision allows identifying three categories of business models: technical intermediaries, con- tent providers and community creators (Trusted Thirds Parties). All of them attempt to adapt their services to achieve an improvement of offer, fidelize constituencies and distinguish own services from competition. Nevertheless, adaptation does not mean personalization. Real personalizing strategies find in content providers and particularly in trading platforms, e-marketplaces, virtual communities and business-networking models the fittest spaces to develop. Thirdly, licence agreements have been identified as challenging contractual mechanisms to implement per- sonalizing strategies. Under the format of click-agreements, log-in agreements or browse-wrap agreements, on-line businesses include provisions ruling personal data and privacy matters. Through the so-called terms of use and privacy policies, they expect to successfully manage privacy issues along the fine line separating lawful personalizing initiatives from undesirable invasions of users’ private sphere. They achieve such purpose when, complying with privacy and personal data protection standards, they apply dissociation process over data to untie them from personal identities, duly inform users about collecting and processing procedures, give them a 2 Legal Framework for Personalization-Based Business Models 21 real and effective opportunity to refuse the installation or the functioning of capturing devices and obtain free consent of users to devote collected information to fulfil legitimate purposes.
References
D. C. Baker. NOTE. ProCD v. Zeidenberg: Comercial reality, flexibility in contract formation, and notions of manifested assent in the arena of shrinkwrap licences. Nortwestern University Law Review, 92:379–433, 1997. L. A. Ballesteros Moffa. La privacidad electrónica. Internet en el centro de protección. Tirant lo Blanch, 2005. Y. Benkler. From consumers to users: Shifting the deeper structures of regulation toward sustainable commons and user access. Federal Communication Law Journal, 52(3):561–579, 2000. F. Bergadamo, A. Montelero, G. Ruffo, and G. Sartor. Privacy digitale. Giuristi e informatici a confronto. G. Giappichelli Editore, 2005. H. W. Classen. Fundamentals of software licensing. IDEA, The Journal of Law and Technology, 37(1):1–85, 1996. Q. Dai and R. J. Kauffman. B2B e-commerce revisited: Leading perspectives on the key issues and research directions. Electronic Markets, 12(2):67–83, 2002. K. M. Das. Forum-selection clauses in consumer clickwrap and browsewrap agreements and the “reasonably communicated” test. Washington Law Review, 77:481–510, 2002. D. A. Einhorn. Shrink-wrap licenses: the debate continues. IDEA, The Journal of Law and Technology, 38(3): 383–401, 1998. M. Grossman. Forming contracts online. The Miami Herald, December 3rd, 2001. M. C. Guerrero Picó. El Impacto de Internet en el Derecho Fundamental a la Protección de Datos de Carácter Personal. Aranzadi, 2006. Z. M. Harrison. Just click here: Article 2B’s failure to guarantee adequate manifestation of assent in click-wrap contracts. Fordham Intellectual Property, Media & Entertainment Law Journal, 8:907–946, 1999. Harvard Law Review Association. Developments in the law: The law of cyberspace. Harvard Law Review, 112 (7):1574–1704, 1999. D. L. Hayes. The Enforceability of Shrinkwrap License Agreements On-Line and Off-Line. Fenwick & West, 2003. A. I. Herrán Ortiz. El derecho a la protección de datos personales en la sociedad de la información. Universidad de Deusto, 2003. P. Kannan, P. K. Kannan, A.-M. Chang, and A. B. Whiston. Handbook on Electronic Commerce, chapter The Internet Information Market: The Emerging Role of Intermediaries. Springer, 2000. P. M. Kaufman. The enforceability of state “shrink-wrap” license statutes in light of Vault Corp. v. Quaid Software, Ltd. Cornell Law Review, 74:222–244, 1989. C. Kuner. European Data Protection Law: Corporate compliance and regulation. Oxford University Press, 2007. R. G. Kunkel. Recent developments in shrinkwrap, clickwrap and browsewrap licenses in the United States. http://www.murdoch.edu.au/elaw/issues/v9n3/kunkel93_text.html, 2002. C. L. Kunz, M. F. del Duca, H. Thayer, and J. Debrow. Click-through agreements: Strategies for avoiding disputes on validity of assent. Business Lawyer, 57(1):401–420, 2001. M. A. Lemley. Intellectual property and shrink-wrap licenses. Southern California Law Review, 68:1239–1294, 1995. L. Lessig. The architecture of innovation. Duke Law Journal, 51(6):1781–1801, 2002. A. Lisi, editor. La privacy in Internet. Simone, 2003. T. W. Malone, J. Yates, and R. I. Benjamin. Electronic markets and electronic hierarchies. Communications of the ACM, 30(6):484–497, 1987. M. Masimini. Il diritto alla privacy. Motivi ed aspetti di una tutela problematica. Arcipelago, 2002. A. Minassian. The death of copyright: Enforceability of shrinkwrap licensing agreements. UCLA Law Review, 45:569–609, 1997. 22 Teresa Rodríguez de las Heras Ballell
C. L. Pitet. The problem with “money now, terms later”: ProCD, Inc. v. Zeidenberg and the enforceability of “shrink-wrap” software licenses. Loyola Los Angeles Law Review, 31:325–352, 1997. K. Puhala. The protection of computer software through shrink-wrap license agreements. Washington & Lee Law Review, 42:1347–1380, 1985. R. Raysman and P. Brown. Clarifying the rules for clickwrap and browsewrap agreements. New York Law Journal, 228(93), 2002. J. Rifkin. La era del acceso. La revolución de la nueva economía. Paidós (Estado y Sociedad), 2000. S. Rodotà. Tecnología y derechos fundamentales. Datospersonales.org, 8, 2004. T. Rodríguez de las Heras Ballell. El régimen jurídico de los Mercados Electrónicos Cerrados (e- marketplaces). Marcial Pons, 2006a. T. Rodríguez de las Heras Ballell. La responsabilidad por software defectuoso en la contratación mercantil. Revista Aranzadi de Derecho y Nuevas Tecnologías, 10:83–110, 2006b. T. Rodríguez de las Heras Ballell. Nueva Lex Mercatoria y contratos internacionales, chapter Las condiciones de uso de los sitios web y los browse-wrap agreements. Ediciones Jurídicas Gustavo Ibáñez, 2006c. A. B. Sculley and W. A. Woods. B2B exchanges. The Killer Application in the Business-to-Business Internet Revolution. Harper Business, 2001. 3 Technological Platforms, Convergence and Adaptation of Interactive Contents
Fabio Paternò
Istituto di Scienza e Tecnologie dell’Informazione “Alessandro Faedo” Area della Ricerca CNR di Pisa, Via G. Moruzzi 1, 56124 PISA, Italy [email protected]
Summary. Our lives are becoming a multi-device experience where people are surrounded by different types of devices (e.g. cell phones, PDAs, desktop computers, digital television sets, intelligent watches, and so on) through which they can connect to networks in different ways. Most of them are mobile personal devices carried by users moving freely about different environments populated by various other devices. Such environments raise many issues for designers and developers, such as the need to obtain interactive contents able to adapt to the interaction resources of the available devices. The objective of this book chapter is to allow readers to gain knowledge in methods and tools for the design of multi-device interactive services that can support designers and developers in addressing a number of the issues raised by multi-device environments.
3.1 Introduction
Multi-device interactive services are acquiring an increasing importance, however they raise a number of issues with regard to obtaining adaptation to the changing interaction resources. Indeed, the diversity in features of the potential devices, such as different screen size, interaction capabilities, processing and power supply, can make a user interface developed for a desktop unsuitable for a PDA and vice versa. For example, an interface layout designed for a desktop platform does not fit in the smaller screen of a PDA, or a graphic interface running on a desktop system must be transformed to a voice interface when the application migrates to a car. Thus, a user interface needs to adapt to the different features of the target platform taking into account usability principles. This chapter first provides an introduction of the main requirements that approaches to interactive content should satisfy. Next, there is a discussion of why and how model-based approaches can be useful for this purpose followed by a description of authoring environments that can be helpful for this purpose. Then, the possible solutions for adaptation at run-time are discussed with an indication of the more promising approaches. The last part is dedicated to migratory interactive services that are acquiring an increasing interest. Lastly, some conclusions are drawn along with indications of interesting possible further developments in the area.
3.2 Requirements in Device Adaptation
The device choice has an influence on the possible tasks to accomplish and how the structure of such tasks can vary in terms of possible secondary tasks, inter-task temporal relations, and content required depending on the device. One issue is that one-fits-all does not work. This means that it is not meaningful to try to support all tasks for all platforms. The big push for cross-platform design comes with the advent of mobile devices, which people can use when they are on the move. This means that there is a clear distinction regarding what it is meaningful to do with a desktop and with a mobile device. For example, people can use a desktop to compare airfares 24 Fabio Paternò and make reservations, while the mobile device can be used to check the real-time status of a particular flight. Likewise, the desktop is well-suited to reading a movie review and/or watching a trailer but not to purchasing a cinema ticket to avoid the line. There are many tasks that are not suitable at all for a mobile device. For example, currently in Europe with the advent of UMTS there is an interest in proposing football matches over phones. This seems senseless even with the last generation mobile phones, which have large displays and better connectivity. People like watching football matches but generally only when they are comfortable sitting and watching a large display in front of them. This allows them to appreciate how the footballers play, the details of the action, the tactics adopted and so on. Few people would ever do this through a mobile phone, even if it is technically possible, because of the small display and discomfort of using it on the move. Also considering that a football match lasts 90 minutes, this would be a terrible experience. What they could appreciate with a mobile device is a different task: receiving real-time updates regarding the match score. Another useful option is to have just the vocal description of the match that can be followed while driving the car or on the move. Thus, they could do this in parallel with other activities. In general, when a multi-platform application is considered, it is important to understand what type of tasks can actually be performed in each available platform. There are various possibilities: The same task can be performed on multiple platforms in the same manner (there may be only some • changes in attributes of the user interface objects from platform to platform). This is the case of tasks whose presentation remains mostly unchanged on different platforms: an example is when in a museum application textual links are provided to access general information about the museum (how to reach, timetable, etc). Same task on multiple platforms but with different user interface objects. An example of this case is high- • lighted in Fig. 3.1. In both systems users can select a geographical area (e.g. London/South East, De- von/Cornwall, etc). However, while in the desktop system a large, coloured interactive map of the museum is at the users’ disposal, in the phone, because of its limited capabilities, a text link is available for each area.
Fig. 3.1. Example of same task and different user interface objects.
Same task on multiple platforms but with different domain objects. This means that during the performance • of the same task different sets of domain objects are manipulated. An example of this is presentations of different information on a desktop system and a mobile phone for a museum application: while in the 3 Technological Platforms, Convergence and Adaptation of Interactive Contents 25
desktop system it is possible to access a wider set of domain elements (title, image type, description, author, material, and date of creation), the mobile interface supports access to only an image (which is a low resolution image just to give users a rough idea of what the work of art is), along with the indications of the title and associated museum section. Same task on multiple platforms but with different task decomposition. This means that the task is sub- • divided differently, with different sets of sub-tasks, depending on the platform. An example of this possi- bility is displayed in Fig. 3.2 that shows how differently the task access work of art is supported in a desktop and in a mobile device. In the desktop system, the users can accomplish additional sub-tasks, which are not supported in other systems. An example concerns the possibility of reading reviews of a particular work of art, which is a lengthy information-processing task that users can perform satisfactorily when sitting in front of a desktop computer, but which is not appropriate with handheld devices.
Fig. 3.2. Example of same main task and different task descompostition.
Same task on multiple platforms but with different temporal constraints. In this case the difference is in the • temporal relationships among the subtasks. With reference to the museum application, consider the case of users who wish to electronically reserve their tickets for a particular visit in order to avoid queues. In both systems they have to provide personal information. However, while in the desktop system they are free to choose the order to follow for filling in the various fields, within the phone application they are constrained by the mobile interface to follow a sequential order. 26 Fabio Paternò
Dependencies among tasks performed on different platforms. An example of this can be found when the • users have to reserve their flight tickets. Through the desktop system users can access, compare and contrast the different options about the best time for the flight ticket. Once they have selected their preferences and entered personal data, the system automatically enables their mobile phone to access real-time data regarding their flight, which ca be useful for example to know whether it is on time (see Fig. 3.3). Capturing this type of task relationships is particularly important when there is some task relevant to only a particular platform and that affects the performance of another task through a different platform. A typical situation occurs when users physically visit the museum and simultaneously annotate the most interesting works of art on the PDA. When they arrive home they would appreciate being able to receive information regarding such works first during their access to the museum web site through a desktop system.
Fig. 3.3. Example of dependencies among asks performed through different platforms.
3.3 Model-based Approaches
In this section we provide an overview of the results that can be obtained through model-based approaches when multi-device interfaces, even using different modalities, are considered, and will link up the discussion to projects currently underway. Indeed, as Myers et al. [2000] indicated while discussing the future of user interface tools, the wide platform variability encourages a return to the study of some techniques for device- independent user interface specification, ... Then, the system might choose appropriate interaction techniques taking all of these into account. The basic idea is that instead of having separate applications for each device 3 Technological Platforms, Convergence and Adaptation of Interactive Contents 27 that exchange only basic data, there is some abstract description and then an environment that is able to suggest a design for a specific device that adapts to its features and possible contexts of use. Thus, a key aspect is to be able to have different views on interactive systems, each view associated with a different abstraction level. With the support of tools, XML-based languages and transformations, it is possible to move from one level to another and tailor a description for one abstraction level to one more refined for the target interaction platform. The model-based community has long discussed such possible description levels (see for example [Szekely, 1996]). It is possible to have various viewpoints on an interactive system. Such viewpoints differ for the ab- straction levels (to what extent the details are considered) and the focus (whether the task or the user interface is considered). Such abstraction levels are:
Task and object model. At this level, the logical activities that need to be performed in order to reach the • users’ goals are considered. Often they are represented hierarchically along with indications of the temporal relations among them and their associated attributes. The objects that have to be manipulated in order to perform tasks can be identified as well. Abstract user interface. In this case the focus shifts to the user interface supporting task performance. • Only the logical structure is considered, in a modality-independent manner, thereby avoiding low-level de- tails. Interaction objects are described in terms of their semantics through interactors [Paternò and Leonardi, 1994]. Thus, it is possible to indicate, for example, that at a given point there is a need for a selection object without indicating whether the selection is performed graphically or vocally or through a gesture or some other modality. Concrete user interface. At this point, each abstract interactor is replaced with a concrete interaction • object that depends on the type of platform and media available and has a number of attributes that define more concretely how it should be perceived by the user. Final user interface. At this level the concrete interface is translated into an interface implemented by a • specific software environment (e.g. XHTML, Java, ...).
To better understand such abstraction levels we can consider an example of a task: making a hotel reserva- tion. This task can be decomposed into selecting arrival and departure dates and other subtasks. At the abstract user interface level we need to identify the interaction objects needed to support such tasks. For example, for easily specifying arrival and departure days we need selection interaction objects. When we move on to the concrete user interface, we need to consider the specific interaction objects supported. So, in a desktop inter- face, selection can be supported by a graphical list object. This choice is more effective than others because the list supports a single selection from a potentially long list of elements. The final user interface is the result of these choices and others involving attributes such as the type and size of the font, the colours, and decoration images that, for example, can show the list in the form of a calendar. Many transformations are possible among these four levels for each interaction platform considered: from higher level descriptions to more concrete ones or vice versa or between the same level of abstraction but for different type of platforms or even any combination of them. Consequently, a wide variety of situations can be addressed. More generally, the possibility of linking aspects related to user interface elements to more semantic aspects opens up the possibility of intelligent tools that can help in the design, evaluation and run-time execution. The main issue underlying the last generation of model-based approaches is the design of multi-device in- terfaces. In current practise the design of multi-platform applications is often obtained through the development of several versions of the same applications, one for each platform considered. Then, such versions can at most exchange data. This solution with no tool support is rather limited, because it implies high implementation and maintenance costs. Thus, there is a need for authoring environments able to support the development of multi-device interfaces by providing design suggestions taking into account the specific features of the devices at hand. LiquidUI is an authoring environment whose main goal is to reduce the time to develop user interfaces for multiple devices. It is based on the User Interface Markup Language (UIML) [Abrams et al., 1999], a declar- ative language that then can be transformed in Java, HTML, and WML through specific rendering software. A UIML program, with its generic vocabulary, is specific to a family of devices (such as the desktop family, the PDA family, the WAP family). There is a transformation algorithm for each family of devices. For example, 28 Fabio Paternò using a generic vocabulary for desktop applications, the developer can write a program in UIML once and have it rendered for Java or HTML. Another approach is UsiXML [Limbourg and Vanderdonckt, 2004], the semantics of the UsiXML models are based on meta-models expressed in terms of UML class diagrams, from which the XML schema definition are derived. Right now, there is no automation between the initial definition of the semantics and their derivation into XML schemas. Only a systematic method is used for each new release. UsiXML aims to address the development of user interfaces for multiple contexts of use and has the advantage of providing a graphical syntax for a majority of constituent models. However, UsiXML renderers are still at the development stage. TERESA citepP24 is intended to provide a complete semi-automatic environment supporting a number of transformations useful for designers to build and analyse their design at different abstraction levels, including the task level, and consequently generate the concrete user interface for a specific type of platform. Currently, the tool supports generation of user interface implementations in XHTML, XHTML mobile device, VoiceXML, multimodal user interfaces in X+V and Java for the digital TV. The tool is able to support different level of automations ranging from completely automatic solutions to highly interactive solutions where designers can tailor or even radically change the solutions proposed by the tool. The last version of the tool supports different entry-points, so designers can start with a high-level task models but they can also start with the abstract user interface level in cases where only a part of the related design process needs to be supported. With the TERESA tool, at each abstraction level the designer is in the position of modifying the representations while the tool keeps maintaining forward and backward the relationships with the other levels thanks to a number of automatic features that have been implemented (e.g. the possibility of links between abstract interaction objects and the corresponding tasks in the task model so that designers can immediately identify their relations). This is useful for designers to maintain a unique overall picture of the system, with an increased consistence among the user interfaces generated for the different devices and consequent improved usability for end-users. Even recent W3C standards, such as XForms (XForms, 2004), have introduced the use of abstractions similar to those considered in the model-based community to address new heterogeneous environments.
3.4 Authoring Multi-device User Interfaces
The concepts discussed in the previous section can be incorporated in authoring environments for multi-device interfaces able to deal with a variety of platforms with different modalities (such as graphical and vocal in- terfaces, digital TV, tilt-based interaction, ...). Examples of such tools will be discussed (e .g. Multimodal TERESA). Once we have identified the tasks that are meaningful to support for each platform then we have to identify how they performance can vary according to the target platform. The idea is to first identify potential interactors in terms of their semantics (how they can change the state of the application) and then to generate corresponding interfaces depending on the platform features. All devices belonging to a given platform will receive an interface with consistent implementation. An abstract user interface is structured into presentations and connections indicating how it is possible to move from one presentation to another. Each presentation is structured into interactors and composition operators. We have defined a number of composition operators, which aim to capture communication effects that often designers want to achieve when they structure their user interfaces. The purpose of the composition operators is to indicate how to put together interactors. Each composition operator is associated with a com- munication goal. Depending on such goals, different implementation techniques will be used to support the composition operator. Figure 3.4 shows an example of a Web page taken from a frequently accessed Web site. We can note how the designer used various techniques to highlight groups of related interface elements. On the top there are elements that are ordered according to the potential user interest. Some elements are grouped using implementation techniques such as same background, same structure, bullets and so on. There are elements that are related to the rest of the Web site, such as the search element. Other elements are highlighted using large image and fonts because they are considered important. In general, the composition operators can involve several interactors or even compositions of interactors. In addition, their definition is modality-independent. They are: 3 Technological Platforms, Convergence and Adaptation of Interactive Contents 29
Grouping (G): indicates a set of interface elements logically connected to each other. • Relation (R): highlights a relation (usually one-to-many) among some elements; one element has some • effects on a set of elements. Ordering (O): some kind of ordering among a set of elements can be highlighted. • Hierarchy (H): different levels of importance can be defined among a set of elements. • There are different types of interaction elements depending on the type of task supported. We have selection elements (to select between a set of elements), edit (to edit an object), control (to trigger an event within the user interface, which can be useful to activate either a functionality or the transition to a new presentation). There are different types of only_output elements (text, object, description, feedback) depending on the type of output the application provides to the user: a textual one, an object, a description, or a feedback about a particular state of the user interface.
Fig. 3.4. Web page with indication of some associated communication goals.
The Multimodal TERESA authoring environment allows designers and developers to start from two pos- sible points: the task model description or the abstract interface description. In both cases they have to specify the target platform (in the current tool version either multimodal desktop or multimodal PDA). If they start with the task model then the tool automatically generates the corresponding abstract interface. As it can be seen in Fig. 3.5, the main area is mainly divided into four parts: the top-left dedicated to the list of presentations com- posing the user interface, the bottom-left indicating the connections defining how it is possible to move from one presentation to another, the top-right indicating the abstract description of the currently selected presenta- tion and the bottom-right part displays the description of the possible concrete implementation of the currently selected element in the abstract part. The concrete part has three tabbed panes, one for the concrete graphical attributes, one for the concrete vocal attributes and one to specify how to compose the multimodal attributes. For example, if we consider the single selection interactor used to indicate the time of a cinema reservation, then the tool as a first suggestion for an implementation in a graphical+vocal desktop interface would propose that the input be equivalent (either graphical or vocal) and the prompt and feedback both redundant. Then, in the vocal section there would be an indication of the corresponding label, and the associated vocal message and feedback, in addition to the definition of the possible choice elements. The graphical part indicates the interaction technique for implementation (e.g. a radio-button), and the corresponding label and elements. The 30 Fabio Paternò tool keeps information in the graphical and vocal parts consistent. So, if the designer indicates five possible choice elements in the vocal part, then this is indicated when the graphical part is accessed as well. Likewise, in the case of a text output, if the corresponding multimodal property is complementarity, then different texts can be specified for vocal and graphical rendering, while if the multimodal attribute is redundant, then the text modified in either part will be updated for the other one as well.
Fig. 3.5. The MultiModal TERESA Environment.
3.5 Run-time adaptation to the device
This section is dedicated to run-time support for multi-device environments. Different types of solution and associated software architectures will be first introduced. Issues and solutions for automatic transformation from desktop interfaces to different platforms (especially mobile ones) will be discussed, showing how presentation, navigation and content can be transformed and the usability issues to address in this process. I will show examples of results that can be obtained by tools provided by main software companies such as Google, Nokia, Microsoft, Opera, along with research results from various groups (including mine). The increasing availability of mobile devices has stimulated interest in tools for adapting the large amount of existing Web applications originally developed for desktop systems into versions that are accessible and usable for mobile devices. This adaptation process implies transforming various aspects:
the presentations, the perceivable aspects, including choice of media and interaction techniques, layout, • graphical attributes, etc; the dynamic behaviour, including navigation structure, dynamic activation and deactivation of interaction • techniques; the content, including text, labels, images. • In carrying out this adaptation it is important to consider the main characteristics of the target platform, in this section the mobile device will be mainly considered. By platform we mean a group of devices that share similar characteristics (such as the desktop, the mobile, the vocal, ...). For example, one aspect to consider is that often mobile devices have no pointing device, thus users have to navigate through 5-way keys, which 3 Technological Platforms, Convergence and Adaptation of Interactive Contents 31 allow them to go left, right, up, down and select the current element. There are also softkeys, which are used to activate commands, but their number and purpose vary depending on the device. In addition, text input is slow and users often have to pay to access the information, and thus prefer short sessions. Regarding the devices, the need for describing their features derives from the increasing availability of various types of interactive devices. Thus, in order to have applications able to adapt to their features there should be a way to represent them. This is particularly important in the area of mobile phones, in which the possible characteristics are the most variable. The generic Composite Capabilities/Preference Profiles (CC/PP) framework (www.w3.org/2001/di/) provides a mechanism through which a mobile user agent —a client, such as a browser, that performs rendering within a mobile device— can transmit information about the mobile de- vice. It is based on RDF and aims to provide a firm foundation for UAProf. The user agent profile (UAProf; www.openmobilealliance.org/release_program/uap_v20.html) is an application of the CC/PP framework. It in- cludes device hardware and software characteristics, information about the network the device is connected to, and other attributes. It is possible ot identify a device through the header of HTTP requests. All the devices complying with UAProf have a CC/PP description of their characteristics in a repository server, which can be queried for knowing them. The description of the devices is in the Resource Description Framework (RDF), language XML-based. When a mobile device sends a request, ity also informs about the URL where its profile is through a specific field in the request called X-Wap-Profile. For example, the x_wap_profile for a Nokia N9500 is http://nds1.nds.nokia.com/uaprof/N9500r100.xml. In this area another proposal is WURFL (Wireless Universal Resource File) [Passani, 2006], which is an XML configuration file which contains information about capabilities and features of several wireless devices. The main scope of the file is to collect as much information as we can about all the existing wireless devices that access WAP pages so that developers will be able to build better applications and better services for the users. This proposal aims to support Web applications for mobile devices. The goal is programmatically abstract away devices differences, avoid that we need to modify applications whenever a new device ships, avoid that we need to track new devices that ship (particularly those in uninteresting markets). The basic idea is a global database of all devices and their capabilities. Starting from the assumption that browsers are different, but they also have many features in common with one another; browsers/devices coming from the same manufacturer are most often an evolution of the same hardware/software. In other words, differences between, for example, a Nokia 7110 and a Nokia 6210 are minimal; devices from different manufacturers may run the same software. WURFL has created a compact, small, and easy to update matrix. The WURFL is based on the concept of family of devices. All devices are descendent of a generic device, but they may also descend from more specialized families (such as those who use the same browser). Its goal is to overcome some limitations of the UAProf standard developed in the OMA. Uaprof seems to rely too much on someone else’s setting up the infrastructure to request profiles. There are cases of manufacturers just associating the profile of a different phones into a new one. The WURFL can be installed in any site and does not need to access device profiles from a repository on the net. In general there are various approaches to authoring multi-device interfaces: Device-specific authoring which means that a specific version for each target platform is developed sepa- • rtely. One example is the Amazon Web site, which has a separate version for mobile devices. This approach is clearly expensive in terms of time and effort. Multiple-device authoring is similar to the previous one but there is a single application that has separate • parts for different platforms. An example is the use of CSS depending on the media. Single authoring. In this case, only one version of the application is built, which is then adapted to various • target platforms. There are two main possibilities for this purpose: either to include the authors’ hints or to specify an abstract description, which is then refined according to the target platform. Interesting contributions in this area are PIMA [Banavar et al., 2004] and SUPPLE [Gajos et al., 2005]. Automatic re-authoring. Here, there is an automatic transformation of a version for a given platform, • usually the desktop, into a version for the target platform. Recently, particular attention has been paid to the latter category, with the aim of obtaining a solution that does not require particular effort in terms of time but is still able to produce meaningful results. Various solutions have been proposed in this category, a first distinction can be made depending on where the re- authoring process occurs: the client device, the application server or an intermediate proxy server. The last 32 Fabio Paternò solution seems particularly interesting because performing the transformation on the client device can raise performance issues with limited capabilities devices, while a solution on the application server would require duplication of installations in all the applications of interest. The feasibility of proxy-based solutions is also shown by its widespread use in tools, such as Google for mobile devices (www.google.com/xhtml) [Kamvar and Baluja, 2006], which converts the Web pages identified by the search engine into versions adapted for mobile devices. Indeed, depending on user agent in http request Google search redirects to www.google.com/xhtml in the case of mobile devices. This version (see Fig. 3.6) has radio buttons instead of tabs to access the various sections. There is no advertisement in the XHTML version, the pieces of text for each link are smallers in the XHTML version, the results for the XHTML version do not contain link cached or similar pages and don’t indicate the size of the page, and the user can access only the previous and next result page.
Fig. 3.6. The two user interfaces for the Google search engine.
In addition, we think that effective solutions for transforming desktop Web sites for mobile access should be based on semantic aspects. Unfortunately, the semantic Web so far has mainly focused on the data semantics through the use of ontologies and languages that allow for more intelligent processing. It would also be impor- tant to consider the semantics of interaction, which is related to the tasks to support in order to reach the users’ goals. Several solutions for automatic re-authoring from desktop-to-mobile have been proposed in recent years. The simplest one just proposes resizing the elements according to the size of the target screen. However, they often generate unusable results with unreadable elements and presentation structures unsuitable for the mobile device. Some solutions such as AvantGo (www.avantgo.com) translate elements and images into other formats, and compress and convert images to match the device characteristics but suffer from similar limitations as the former approach. Thus, research work has focused on transformations able to go further, to modify both the content and structure originally designed for desktop systems to make them suitable for display on small screens. Even in this case various possibilities have been explored. The most common transformation supported by current mobile devices is into a single column (the narrow solution): the order of the content follows that of the mark-up file starting from the top, the images are scaled to the size of the screen, and the text is always visible and the content compacted without blank spaces. It eliminates scrolling in one dimension, though it greatly increases the amount of scrolling in the other dimension. For example Opera SSR (Small Screen Ren- dering, www.opera.com/products/smartphone/ smallscreen/) uses a remote server to pre-process Web pages before sending them to a mobile device, Web content is compressed to reduce the size of data transfers (see Fig. 3.7). In general, in this solution content requiring a good deal of space such as maps and tables can become 3 Technological Platforms, Convergence and Adaptation of Interactive Contents 33 unreadable; and often it is difficult to understand that the corresponding desktop page has changed because the initial part of several desktop pages is indistinguishable.
Fig. 3.7. The example of transformation with Opera SSR.
Various approaches have considered the application of information visualization [Spence, 2001] techniques to address these issues. Fish-eye representations have been considered, for example Fishnet [Baudisch et al., 2004], which is a fisheye Web browser that shows a focus region at a readable scale, while spatially com- pressing page content outside the focus region. However, generating fish-eye representations in mobile devices can require excessive processing. Overview + detail splits a Web page into multiple sections and provides an overview page with links to these sections. The overview page can be either a thumbnail image, or a text sum- mary of the Web page. Within this approach various solutions have been proposed. Smartview [Milic-Frayling and Sommerer, 2002] is a thumbnail view of the original Web page in zoom-out, fitting the screen horizontally. The approach partitions the page in logical regions; when one is selected, content is presented inside the screen space in a detailed view. In Gateway [MacKay et al., 2004] the detailed view uses a focus-plus-context tech- nique, enlarging the selected region above the detailed view. Summary Thumbnails [Lam and Baudisch, 2005] uses the same thumbnail approach but the texts are summarized enabling good legibility (fonts are enlarged to a legible size and characters are cropped from right to left until the sentence fits in the available area). The main issue with this type of approach is that it works well in some cases, less in others because they mainly focus on the transformation of specific elements (for example Summary Thumbnail mainly works on text snip- pets). Another contribution in this area is MiniMap [Roto et al., 2006] a browser for Nokia 6600 mobile phones 34 Fabio Paternò developed at Nokia Research. It removes the need for horizontal scrolling to read text and provides enough contextual information to give an idea of the page structure and the current location without destroying the original page layout. The user interface is organised in such a way that text size should not exceed the screen space and provides an overview+detail representation. The overview is given by an area dedicated to showing where the current mobile page is located in the original desktop page. However, this solution is effective only with mobile devices with relatively large screens. We believe that model-based approaches [Paternò, 1999, Szekely, 1996] can provide a more general solu- tion to adaptation to the interaction device. They are based on the use of logical descriptions that capture the main semantic aspects of the user interface and hide low-level details. Some first studies of how to apply them in this context have already been proposed [Bandelloni et al., 2004, Eisenstein et al., 2001, Florins and Vander- donckt, 2004]. These interesting works were useful to provide solutions to specific issues raised by supporting interactions of mobile users, but they did not address the issue of providing a general solution for taking Web sites originally developed for desktop systems and dynamically transforming them into accessible and usable versions for mobile devices while users are accessing them. Another interesting application of model-based approaches in this area is PUC [Nichols et al., 2002] which dynamically generates user interfaces for mobile devices able to control a domestic appliance starting with its logical description. A novel solution [Bandelloni et al., 2004] for a different domain (Web applications), which is characterised by wider variability in terms of content and tasks to support, is based on the use of a migration/proxy server able to support both adaptation and state preservation across multiple devices.
3.6 Migratory User Interfaces
This section discusses how mobile users can be supported in multi-device environments. To this end, distributed and migratory interfaces are introduced. One important aspect of ubiquitous environments is to provide users with the possibility to freely move about and naturally continue the interaction with the available applications through a variety of interactive devices (i.e. cell phones, PDAs, desktop computers, digital television sets, intelligent watches, and so on). Indeed, in such environments one big potential source of frustration is that people have to start their session over again from the beginning at each interaction device change. Migratory interactive services can overcome this limitation and support continuous task performance. This implies that interactive applications be able to follow users and adapt to the changing context of use while preserving their state. Migratory interfaces are interfaces that can transfer among different devices, and thus allow the users to continue their tasks. This definition highlights important concepts: task performance continuity, device adaptation and interface usability. Task performance continuity means that when migration occurs users do not have to restart the application on the new device, but they can continue their task from the same point where they left off, without having to re-enter the same data and go through the same long series of interactions to get to the presentation they were accessing on the previous device. General solutions for migratory interactive services can be obtained by means of addressing three aspects: adapt and preserve the state of the software application parts dedicated to interacting with end users; support mechanisms for application logic reconfiguration; and define suitably flexible mechanisms from the underlying network layers. There are many applications that can benefit from migratory interfaces. In general, applications that require time to be completed (such as games, business applications) or applications that have some rigid deadline and thus need to be completed wherever the user is (e.g. online auctions). Other applications that can benefit from this flexible reconfiguration support are those that have to provide users with continuous support during the whole day through different devices (for example, in the assisted living domain). ICrafter [Ponnekanti et al., 2001] is a solution to generate adaptive interfaces for accessing services in in- teractive spaces. It generates interfaces that adapt to different devices starting with XML-based descriptions of the service that must be supported. However, ICrafter is limited to creating support for controlling interactive workspaces by generating UI for services obtained by dynamic composition of elementary ones and does not provide support for continuity of task performance across different devices. Aura [Garlan et al., 2002] provides support for migration but it is obtained by changing the application depending on the resources available in 3 Technological Platforms, Convergence and Adaptation of Interactive Contents 35
Fig. 3.8. A migration example scenario. the device in question, while there is a need for solutions able to generate interfaces of the same application that adapt to the interaction resources available. Bharat and Cardelli [1995] addressed the migration of entire applications (which is problematic with limited-resource devices and different CPU architectures or operating systems), while we focus on the migration of the UI part of a software application. Kozuch and Satyanarayanan [2002] identified a solution for migration based on the encapsulation of all volatile execution state of a virtual machine. However, their solution mainly supports migration of applications among desktop or laptop systems by copying the application with the current state in a virtual machine and then copy the virtual machine in the target device. This solution does not address the support of different interaction platforms supporting different interaction resources and modalities, with the consequent ability to adapt to them. Chung and Dewan [1996] proposed a specific solution for migration of applications shared among several users. When migration is trig- gered the environment starts a new copy of the application process in the target system and replay the saved sequence of input events to the copy to ensure that the process will get the state where it left off. This solution does not consider migration across platforms and consequently does not support run-time generation of a new version of the UI for a different platform. A discussion of some high-level requirements for software architec- tures in multi-device environments is proposed in [Balme et al., 2004] without presenting a detailed software architecture and implementation indicating a concrete solutions at these issues. In this section, we also introduce a specific architectural solution, based on a migration/proxy server, able to support migration of user interfaces associated with applications hosted by different content servers. At the HHIS Laboratory of ISTI-CNR, Ban- delloni et al. [2004] found a solution based on the use of pre-computed interfaces for different platforms that are dynamically activated, which then evolved in a new solution [R.Bandelloni et al., 2007], implemented in an engineered prototype, supporting also migration through different modalities (such as voice). Especially in heterogeneous environments (namely, environments in which different types of devices exist), the concept of migratory user interfaces raises a number of design issues that should be appropriately analysed and addressed in the attempt to identifying an effective migration architecture/process. A suitable framework for migration should consider at least the dimensions described hereafter. Such dimensions are:
Activation type: how the migration is triggered. The simplest case is on demand, in which the user actively • selects when and how to migrate. Otherwise, in automatic migration, it is the system that activates the device change (depending on e.g. mobile device battery consumption level, device proximity, etc). Type of migration: This dimension analyses the extent of migration, as there are cases in which only a • portion of the interactive application should be migrated. A number of migration types can be identified: Total migration allows basically the user to change the device used to interact with the application. 36 Fabio Paternò
– Partial migration is the ability to migrate a portion of the UI (the remaining portion remains in the source device). – In the distributing migration the user interface is totally distributed over two or more devices after migration. – The aggregating migration performs the inverse process: the interface of multiple source devices are grouped in the user interface of a single target device. – The multiple migration occurs when both the source and the target of the migration process are multiple devices. Number/combinations of migration modalities. This dimension analyses the modalities involved in the • migration process. Mono-modality means that the devices involved in the migration adopt the same modal- ity interaction. Trans-modality means that the user can migrate changing the interface modality. An exam- ple of migration from graphical interface to vocal interface is the case of users navigating the Web through a PDA or Desktop PC and afterwards migrate the application to a mobile phone supporting only vocal interaction. Lastly, with multi-modality the migratory interface contemporaneously supports two or more interaction modalities at least in one device involved in the migration. Work in this area often has mainly focused on graphical interfaces, investigating how to change the graphical representations depending on the size of the screens available. Type of interface activated. This dimension specifies how the user interface is generated in order to be • rendered on the target device(s). With precomputed user interfaces the UI has been produced in advance for each type of device. Thus, at runtime there is no need for further adaptation to the device but only for adaptation of the state of the UI. On the contrary, if a runtime generation of user interfaces is considered, the migration engine generates the UI according to the features of the target device when migration occurs. In an intermediate approach the migration engine adapts dynamically to the different devices using some templates previously created. Granularity of adaptation. The adaptation process can be affected at various levels: the entire application • can be changed depending on the new context or the UI components (presentation, navigation, content). Migration time. The device change that characterise migration can occur in different temporal evolutions: • – Continuous: The user interface migration occurs immediately after its triggering. – Postponed: The request of migration can be triggered at any time but it actually occurs after some time. This can happen when the target device is not immediately available (for example when migrating from the desktop in the office to the vocal device in the car, which has to be first turned on by the user when he enters the car). How the UI is Adapted. Several strategies can be identified regarding how to adapt user interfaces after a • migration process occurs: – Conservation. This strategy maintains the arrangement and the presentation of each object of the user interface: one possible example is the simple scaling of the user interface to different screen sizes. – Rearrangement. In this case, all the UI objects are kept during the migration but they are rearranged according to some techniques (e.g. using different layout strategies). – Increase. When the UI migrates from one device with limited resources to one offering more capabil- ities, the UI might be improved accordingly, by providing users with more features. – Reduction. This technique is the opposite of increase and it can be applied when the UI migrates from desktop to mobile device because some activities that can be performed on the desktop might result unsuitable on a mobile device. – Simplification. In this case, all the user interface objects are kept during the migration but their repre- sentation is simplified, for example, different resolutions are used for figures or figures are substituted with textual descriptions. – Enhancement. This technique represents the opposite of simplification (e.g. a textual description might be substituted with multimedia information). The impact of migration on tasks depends on how the user interface is adapted because reduction and • increase can produce some change on the range of tasks supported by each device. Differently, conservation and rearrangement do not produce any effect on the set of tasks supported. Then some possible cases are: (i) after a partial or distributing migration some tasks can be performed on two or more devices in the 3 Technological Platforms, Convergence and Adaptation of Interactive Contents 37
same manner (task redundancy), which means, for instance, that the decomposition of the different tasks into subtasks is unchanged, as well as the temporal relationships (sequencing, concurrency, etc) occurring among them; (ii) after a partial or distributing migration a part of a task can be supported on one device and the other part/s is/are available on different devices (task complementarity). Additional cases are when the number of task supported (iii) increases (task increase) or (iv) decreases (task decrease) after migration. Obviously, a final case might be identified when the migration has no impact on tasks, as they remain substantially unchanged. Context model. During adaptation of the UI the migration process can consider the context in terms of • description of device, user and environment. Generally, the context dimension that is most taken into ac- count with the aim of producing usable UI is the device and its properties, together with the surrounding environment. Context reasoning. A context modelling process begins by identifying the context of interest. This depends • on the specific task which should be associated with the context of interest. The context of interest can be a primitive context, which can directly be captured by employing a sensor; or it can be a higher-level context, which is a result of manipulation of several primitive contexts. If a context of interest is a higher-level context, a reasoning scheme is inadvertently required, which can be either a logic-based reasoning scheme or a probabilistic reasoning scheme. A logic-based reasoning scheme considers a primitive context as a factual data while a probabilistic reasoning scheme does not. Depending on the nature of the sensed data available and the way the data are manipulated, ignorance can be classified as follows: – Incompleteness refers to the fact that some vital data about the real-world situation to be reasoned about is missing; the available data, however, are considered to be accurate. – Imprecision refers to inexact data from sensors. Inexact data arises due, partly, to the physical lim- itations of the sensing elements employed. Different sensors have different resolution, accuracy, and sensing range. Besides, the performance of physical sensors can be influenced by external factors such as surrounding noise or temperature. – Uncertainty refers to the absence of knowledge about the reliability of the data sources —this knowl- edge might be information about the parameters listed above to determine (characterise) the degree of imprecision incorporated in sensory data. Implementation environment. The migration process can involve different types of applications. Probably • due to their diffusion, the most recurrently considered applications are web-based systems (static/dynamic pages), but also other applications (Java, Microsoft .NET, etc) can be considered. Architecture. With regard to the architecture of the migration support environment there are different • strategies, for example: proxy-based, in which there is an intelligent unit managing all migration requests and sending all data to target devices; and peer to peer, where the devices directly communicate and nego- tiate the migration parameters.
3.7 Conclusions
This chapter has discussed a number of issues related to technological platforms, convergence and adaptation of interactive contents and it has described the state of art in this area. Particular attention has been dedicated to solutions for adaptation to the device. Migratory interactive services has been described along with an indication of the relevant design dimensions. This is an area that is acquiring an increasing importance, with a strong technological push. One important research area for the next years is dedicated to making the transformation rules driving adaptation accessible and modifiable by end users in order to allow their customization for specific preferences.
References
Abrams, M., Phanouriou, C., Batongbacal, A., Williams, S., and Shuster, J. (1999). UIML: An appliance- independent XML user interface language. In 8th International World Wide Web Conference, Toronto, Canada. 38 Fabio Paternò
Balme, L., Demeure, A., Barralon, N., Coutaz, J., and Calvary, G. (2004). CAMELEON-RT: a software archi- tecture reference model for distributed, migratable, and plastic user interfaces. Lecture Notes in Computer Science, 3295:291–302. Banavar, G., Bergman, L. D., Gaeremynck, Y., Soroker, D., and Sussman, J. B. (2004). Tooling and system support for authoring multi-device applications. Journal of Systems and Software, 69(3):227–242. Bandelloni, R., S.Berti, and Paternò, F. (2004). Mixed-initiative, trans-modal interface migration. Lecture Notes in Computer Science, 3160:216–227. Baudisch, P., Lee, B., and Hanna, L. (2004). Fishnet, a fisheye web browser with search term popouts: A comparative evaluation with overview and linear view. In International Working Conference on Advanced Visual Interfaces, Gallipoli, Italy. Bharat, K. A. and Cardelli, L. (1995). Migratory applications. In 8th ACM Symposium on User Interface Software and Technology, Pittsburgh, Pennsylvania (USA). Chung, G. and Dewan, P. (1996). A mechanism for supporting client migration in a shared window system. In 9th ACM Symposium on User Interface Software and Technology, Seattle, Washington (USA). Eisenstein, J., Vanderdonckt, J., and Puerta, A. R. (2001). Applying model-based techniques to the development of UIs for mobile computers. In International Conference on Intelligent User Interfaces, Santa Fe, New Mexico (USA). Florins, M. and Vanderdonckt, J. (2004). Graceful degradation of user interfaces as a design method for multi- platform systems. In International Conference on Intelligent User Interfaces, Funchal, Madeira (Portugal). Gajos, K., Christianson, D., Hoffmann, R., Shaked, T., Henning, K., Long, J. J., and Weld, D. S. (2005). Fast and robust interface generation for ubiquitous applications. Lecture Notes in Computer Science, 3660:37–55. Garlan, D., Siewiorek, D., Smailagic, A., and Steenkiste, P. (2002). Project Aura: Toward distraction-free pervasive computing. IEEE Pervasive Computing, 21(2):22–31. Kamvar, M. and Baluja, S. (2006). A large scale study of wireless search behavior: Google Mobile search. In SIGCHI Conference on Human Factors in Computing Systems, Montréal, Canada. Kozuch, M. and Satyanarayanan, M. (2002). Internet Suspend/Resume. In 4th IEEE Workshop on Mobile Computing Systems and Applications, New York (USA). Lam, H. and Baudisch, P. (2005). Summary thumbnails: Readable overviews for small screen web browsers. In SIGCHI Conference on Human Factors in Computing Systems, Portland, Oregon (USA). Limbourg, Q. and Vanderdonckt, J. (2004). UsiXML: A user interface description language supporting multiple levels of independence. In Workshop on Engineering Advanced Web Applications, in conjunction with the 4th International Conference on Web Engineering, Munich, Germany. MacKay, B., Watters, C., and Duffy, J. (2004). Web page transformation when switching devices. Lecture Notes in Computer Science, 3160:228–239. Milic-Frayling, N. and Sommerer, R. (2002). Smartview: Enhanced document viewer for mobile devices. Technical report, Microsoft Research, Cambridge (UK). Myers, B., Hudson, S., and Pausch, R. (2000). Past, present, future of user interface tools. ACM Transactions on Computer-Human Interaction, 7(1):3–38. Nichols, J., Myers, B. A., Higgins, M., Hughes, J., Harris, T. K., Rosenfeld, R., and Pignol, M. (2002). Gener- ating remote control interfaces for complex appliances. In 15th ACM Symposium on User Interface Software and Technology, Paris, France. Passani, L. (2006). The Wireless Universal Resource File. http://wurfl.sourceforge.net/. Paternò, F. (1999). Model-based Design and Evaluation of Interactive Applications. Spriger-Verlag. Paternò, F. and Leonardi, A. (1994). A semantics-based approach to the design and implementation of interac- tion objects. Computer Graphics Forum, 13(3):195–204. Ponnekanti, S. R., Lee, B., Fox, A., Hanrahan, P., and Winograd, T. (2001). ICrafter: A service framework for ubiquitous computing environments. Lecture Notes in Computer Science, 2201:56–75. R.Bandelloni, Paternò, F., Santoro, C., and Scorcia, A. (2007). Web user interface migration through different modalities with dynamic device discovery. In 2nd International Workshop on Adaptation and Evolution in Web Systems Engineering, Como, Italy. Roto, V., Popescu, A., Koivisto, A., and Vartiainen, E. (2006). Minimap: a web page visualization method for mobile phones. In SIGCHI Conference on Human Factors in Computing Systems, Montréal, Canada. 3 Technological Platforms, Convergence and Adaptation of Interactive Contents 39
Spence, R. (2001). Information Visualization. ACM Press. Szekely, P. (1996). Retrospective and challenges for model-based interface development. In 2nd International Workshop on Computer-Aided Design of User Interfaces, Namur, Belgium.
4 Filtering Techniques for Selection of Contents and Products
Qing Li1, Yuanzhu Peter Chen2, and Zhangxi Lin3
1 School of Economic Information Engineering, Southwestern University of Finance and Economics, China. [email protected] 2 Department of Computer Science, Memorial University of Newfoundland, Canada. [email protected] 3 The Rawls College of Business Administration, Texas Tech University, USA [email protected]
Summary. With the development of e-commerce and the proliferation of easily accessible information, filter- ing systems (also known as recommender systems) have become a popular technique to prune large informa- tion spaces so that users are directed to products that best meet their needs and preferences. In this chapter, we present an overview of the field of filtering techniques and describe the current filtering methods that are usually classified as content-based, collaborative, and hybrid filtering approaches. We also analyze various limitations of current filtering techniques and discuss the challenges for building a large-scale and practical filtering system. In particular, content-based filtering methods, which select the appropriate information for a user by matching user preferences, are first introduced. Second, we look into the memory-based and model-based collaborative filtering approaches to elaborate on the collaborative mechanism, which utilizes the opinions of other users to make recommendations for personalized information filtering. Third, hybrid filtering techniques are introduced, which combines the above filtering methods. At last, we address the challenges for building a large-scale and commercial recommender system. These challenges include scalability of filtering algorithms, data sparsity, effectiveness of recommender systems and so forth.
4.1 Introduction
Information filtering, also known as information recommendation, has an obvious appeal in an environment where the amount of on-line information vastly outstrips any individual’s capability to survey. In our daily life, critique experts perform information filtering on our behalves. For instance, we often check ranking lists for bestselling movies and listen to movie critics. However, such manually abstracted information is not responsive and comprehensive enough compared to the volumous raw information generated everyday. Thus, information filtering technologies offer a powerful solution to this problem. Information filtering has found its way into the activities in specific areas of current social life. For example, filtering systems can continuously provide financial investors with relevant information, such as news, financial reports, and so forth, for better analysis and decision making. With the help of a filtering system, a novice can obtain suggestions from a professional in a community network. In reality, some of the on-line vendors have already incorporated more or less filtering capabilities into their commercial promotion such as Amazon, Yahoo! and CDNow. Information recommendation focuses on providing an individual with tailored information. Different from search engines which target at immediate information acquisition, information recommendation is more of a long-term information delivery service. Products, including books, movies, music, and so forth, are often selected and recommended to users automatically according to their preferences by such systems. Techniques applied in information filtering are generally categorized as content-based filtering (CBF), col- laborative filtering (CF), and hybrid filtering (HF) according to their operation mechanism [Li and Kim, 2003]. 42 Qing Li, Yuanzhu Peter Chen, and Zhangxi Lin
In particular, content-based filtering utilizes the summarized content information of products for individual users. In contrast, collaborative filtering applies preference induction based on the similarities among multiple users. Hybrid filtering utilizes content-based and collaborative filtering for more flexible and accurate recom- mendation. In this chapter, we present a categorized overview of the field of information filtering and analyze various limitations of these filtering techniques. First, content-based filtering methods, which select the appropriate information for a user by matching the user preferences, are introduced in Section 4.2. Second, we look into the memory-based and model-based collaborative filtering approaches to elaborate on the collaborative mechanism in Section 4.3. In Section 4.4, hybrid filtering techniques are introduced as extension to the above approaches. To conclude, we address the challenges for building a large-scale and commercial recommender system.
4.2 Content-Based Filtering
Content-based filtering selects the appropriate information for a user by matching the user preference against an information repository (Figure 4.1). Typically, in a content-based filtering system, we construct and maintain a profile for each user to model the user’s preferences. It can be generated explicitly or implicitly. For instance, the system may have a user explicitly describe his/her preferences, or it builds a user profiles based on the user’s browsing behaviors (e.g. clicks and reading time) [Wittenburg et al., 1995]. On the other hand, the system searches the information repository to recommend products by matching the user profiles against products in the repository. Although details differ from system to system, content-based filtering systems share in common a means for creating user profiles, and a means for comparing products to user profiles for recommendation.
User Interface
Interested Documents
A Browsed Documents
B Matched documents
User Profile
Information Repository
Fig. 4.1. Content-based filtering.
4.2.1 Construction of User Profiles
A content-based filtering system usually processes textual products such as news articles and web pages. For non-textual ones (e.g. multimedia files) the textual descriptions of these items in the file metadata are utilized. For instance, a music filtering system typically utilizes the textual descriptions such as a music genre, piece 4 Filtering Techniques for Selection of Contents and Products 43 title, artist name, and album title to build user profiles. A user profile can be explicitly customized by a user or implicity constructed by the system based on the user’s behaviors. In user customization, a user is provided with an interfaces as a step in registration, where possibly check boxes are used for a user to select from a known set of attributes, e.g. name of favorite singers and genre of favorite books or music. Alternatively, the interface may allow a user to enter texts in an edit box to express his or her interests, e.g. scenario of interested movies, name of musician. Obviously, user customization puts a big burden on end users. Because not many users would like to make this effort at the registration. In addition, users are reluctant to revise their profiles frequently as their preferences change. Due to the limitation of manual user customization, automatic construction of user profiles is usually adopted in content-based filtering systems. In this case, the system first selects interested products based on a user’s browsing behaviors. Second, a user profile, typically represented as a set of keywords, is generated based on the textual contents of interested items. For instance, the Fab system [Balabanovic and Shoham, 1997], a web page recommender, represents a web page preferred by the users with the 100 most “important” words therein. A user profiles is generated as a set of words using these summary words of the preferred pages. The “importance” of a word can be determined by a weighting metric. The most accepted one is the normalized TF-IDF metric [Baeza-Yates and Ribeiro-Neto, 1999]. The weight of term i in document j is denoted by wi,j :
(1 α)fi,j N wi,j = α + − log , (4.1) „ maxifi,j « × ni where N is total number of documents that can be recommended to a user, ni is the number of documents containing term i, and α is a weighting factor usually set to 0.5. fi,j is the frequency of word i in document j; N maxi fi,j is the maximum frequency among all words that appear in document dj ; and the factor log is a ni variant of the inverse document frequency, an indication of the rareness of term i in the collection of documents to be recommended.
4.2.2 Selection of Recommended Products
Once a user profile is constructed, we should compare the profile against an information repository to select relevant products. There are several techniques to evaluate the similarity between a user profile and a product using Boolean methods [Anick et al., 1989, Lee, 1994, Verhoeff et al., 1961], vector-space methods [Salton and Buckley, 1988], probabilistic models [Robertson and Sparck Jones, 1976], neural networks [Kim and Raghavan, 2000, Wilkinson and Hingston, 1991], fuzzy set models [Ogawa et al., 1991], etc. The most accepted is the vector-space approach using cosine similarity. As mentioned above, a user profile u and a product p can be represented as TF-IDF vectors u =< w1,u, w2,u, ...wM,u > and p =< w1,p, w2,p, ...wM,p > where M is the total number of wordsWin the system. Essentially, a user profileW or a product is a document, and each component of the vector can be calculated by Equation 4.1. The cosine similarity between a profile and an item is denoted by s(u, p):
M wi,u wi,p s(u, p) = i=1 × . (4.2) PM 2 M 2 i=1 wi,u i=1 wi,p q q For example, if a user has read a large numberP of news articlesP related to financial risk investment, then a content-based recommender will provide other relevant articles to him/her. By processing the articles read by this user, a user profile can be constructed. Intuitively, an article interested by the user would have a high frequency for the finance-related terms, which is reflected by this user’s profile. Consequently, a recommender system using the above cosine similarity measure will assign a high similarity value to the articles containing more finance-related terms.
4.2.3 Limitations
Although content-based filtering has proven to be effective in recommending textual items relevant to a topic, it also has its limitations [Adomavicius and Tuzhilin, 2005]: 44 Qing Li, Yuanzhu Peter Chen, and Zhangxi Lin
Content-based filtering more than often provides recommendation in a literal sense, because all the infor- • mation is selected and recommended based on textual contents. Even though a product was essentially useful, it might be under-valued because of the ambiguous and thus misleading appearance of the textual contents. In addition, it is indistinguishable in quality of the recommended products. This is because the term vector for a product simply captures the frequency of each term in an article, and a poorly worded article can well have an equal or even higher similarity value than a finely written one. Content-based filtering generally works well with sufficient textual information. However, other multimedia • files such as images, audio and video streams are not applicable if the metadata do not have enough textual information. A new user of such a system must face a “slow start” because he/she has to rate relatively large number • of items before a recommender can induce his/her preferences and can serve him/her with reliable recom- mendation. Content-based filtering usually focuses on recommending the most similar to users. However, in some cases • the most similar product is effective redundant, such as two different news articles describing the same event. For example, Daily-learn [Billsus and Pazzani, 2000] discards items not only if they are different from a user’s preferences but also overly similar to the previously recommended ones.
4.3 Collaborative Filtering
Unlike content-based filtering, collaborative filtering is a technology wherein peer opinions are employed to predict the interests of others. The idea of collaborative filtering has been around for about 20 years. Gold- berg and his colleagues first applied collaborative filtering to recommender systems [Goldberg et al., 1992, Terry, 1993]. There, binary preference indicators and custom texts were applied to identify like-minded users manually. Subsequently, GroupLens [Resnick et al., 1994] and Ringo [Shardanand and Maes, 1995], devel- oped independently, were proposed for automated prediction. Collaborative filtering has prevailed not only in research but also in the industry. There are plenty of collaborative recommender systems designed for differ- ent purposes. These include the MovieLens system for movies, Jeter system for jokes [Gupta et al., 1999], Flycasting for online radio [Hauver and French, 2001], VERSIFI (formerly AdaptiveInfo.com) for news, etc. Most collaborative filtering techniques falls into one of two categories: memory-based and model-based methods. Memory-based methods store massive raw user ratings in a database which can be uploaded into computer memory at once to compute preference similarities among users. Here, a user rating corresponds to a point in a metric space. In contrast, model-based algorithms build abstract models upon raw ratings and recommend products using these light-weight models. Both approaches have been shown to be effective in collaborative filtering.
4.3.1 Memory-Based Collaborative Filtering
Memory-based collaborative filtering techniques have been deployed widely at many commercial web sites because not only are they simple and conceptually intuitive, but also they are deemed sufficiently accurate for many real-world applications. One of the earliest memory-based collaborative filtering systems [Shardanand and Maes, 1995] is devised by the GroupLens Research Lab, University of Minnesota. It adopts the Pearson correlation coefficient method to compute the similarities among users which are later used to group the like- minded users. Based on the opinions of like-minded users about a certain product, the recommender can infer the preference of a user who does not have a rating for this product. Generally, a memory-based collaborative filtering algorithm consists of two steps:
1. Calculate user similarities to find the nearest neighbors of the target user in the metric space. One of the most accepted metrics to compute similarity is the Pearson Correlation Coefficient. Consider two users x and y and the intersection of their rated products contains m elements, i.e. they share m co-rated prod- ucts. Let Rx,i (respectively Ry,i) denote the rating on product i by user x (respectively user y), which is typically in 1, 2, 3...5 . As a shorthand, we use R¯x (respectively R¯y) to denote the average rating value { } 4 Filtering Techniques for Selection of Contents and Products 45
of user x (respectively user y) over all the products. The similarity between user x and y is denoted by Sp(x, y):
m cov(x, y) ((Rx,i R¯x) (Ry,i R¯y)) Sp(x, y) = = i=1 − × − , (4.3) σxσy Pm ¯ 2 m ¯ 2 i=1(Rx,i Rx) i=1(Ry,i Ry) q − q − 2. Predict a user’s rating on a yet-to-rat productPbased on the ratings ofPthe above selected near neighbors. The prediction of user x on product i, denoted by Px,i, in the simplest form could be computed as the average ratings of these neighbors on this product. However, the most common approach [Shardanand and Maes, 1995] is based on a weighted average of rating deviations of like-minded users. Formally,
n (Ry,i R¯y) Sp(x, y) ˆ y=1 Px,i = Rx + n − × , (4.4) P Sp(x, y) y=1 | | P 0 where n is the number of the selected nearest neighbors, and Rˆx is user x s average rating value over all products that he/she has rated so far. The following is a simple example of applying the above memory-based algorithm for recommendations. Four users have rated a subset of six pieces of music using values in 1, 2, ..., 5 as shown in Table 4.1. { }
Table 4.1. User-product table. Kate Luke Mike Nancy Music 1 1 4 2 Music 2 5 2 4 Music 3 3 5 Music 4 2 5 Music 5 4 1 Music 6 2 5
The higher the score, the more preference the user. The recommendation problem is to estimate the users’ ratings for the unrated products in the table, i.e. to fill in the empty cells. To predict Kate’s rating on music 6, we first compute the similarity between Kate and Luke using Equation 4.3.
¯ (1+5+2+4) RKate = 4 = 3 (4+2+5+1) RLuke = 4 = 3 Sp(Kate, Luke) = = (1−3)×(4−3)+(5−3)×(2−3)+(2−3)×(5−3)+(4−3)×(1−3) √(1−3)2+(5−3)2+(2−3)2+(4−3)2×√(4−3)2+(2−3)2+(5−3)2+(1−3)2 = 0.8 − This means that Kate and Luke have dissimilar appreciation in music. Similarly, we can compute the metric between Kate and Mike, which is 1.0. This indicates that they have a very similar taste in music. Note that there is no need to compute the value for Kate and Nancy since they have no co-rated music pieces and Nancy dose not have a reference rating for music 6. Next, we select the 2-nearest neighbors of Kate using these similarity values, i.e. Luke and Mike. Using the rating of Mike and that of Luke for music 6, we obtain an estimated rating of Kate for this piece:
¯ ¯ (RLuke,6−RLuke)×Sp(Kate,Luke)+(RMike,6−RMike)×Sp(Kate,Mike) P , = Rˆ + Kate i Kate |Sp(Kate,Luke)|+|Sp(Kate,Mike)| (2−3)×(−0.8)+(5−3)×1 = 3 + |1|+|−0.8| = 4.56 Therefore, the recommender can recommend music 6 to Kate since she is likely to appreciate it based on the prediction score 4.56. 46 Qing Li, Yuanzhu Peter Chen, and Zhangxi Lin
There are a series of further research on memory-based collaborative filtering with some adjustments. For instance, a variety of similarity metrics are adopted instead of the original Pearson correlation, such as the mean-squared difference [Shardanand and Maes, 1995], vector similarity [Breese et al., 1998] or probabilistic similarity [Pennock et al., 2000]. Although details of these techniques are different, all of them share the same intuition that users tend to have similar opinions on a product if they have similar preferences on other products. These memory-based filtering techniques are referred to as user-based approaches. In contrast, item-based collaborative filtering, e.g. Sarwar et al. [2001] assume that a user would also be interested in products similar to the ones that he/she has high ratings for. For instance, if a user has bought many action movies before, he/she usually has a strong intention to buy more action movies. Compared to user-based approaches, item- based methods show more advantages from a theoretical point of view. First, human beings can be emotional and may change their preferences and interests over time or under different circumstances. On the contrary, the features of merchandises are fairly stable. Therefore, the similarity between items often involves a higher system overhead. Second, the construction of user profiles is a little difficult and usually adds the burden of users to input. Furthermore, users are sometimes reluctant to reveal their private information through user profiles.
4.3.2 Model-Based Collaborative Filtering
In contrast to memory-based collaborative filtering, a model-based method constructs an abstract model for underlying user preferences from which predictions are inferred. Several approaches have been proposed and studied in previous work [Popescul et al., 2001, Huang et al., 2004, Hofmann, 2003, Jin et al., 2003b, Nas- raoui and Pavuluri, 2004]. All of these share the same idea of clustering, i.e. partitioning users or products into groups of similar tastes or attributes for recommendations. Examples include the Bayesian clustering (BC) model [Breese et al., 1998], Bayesian network (BN) model [Breese et al., 1998], and Aspect model [Hofmann, 2003]. The basic idea of the BC model is to assume that ratings are observations of the multinomial mixture model with parameters, and the model is estimated using the Expectation Maximization (EM) algorithm. The BN model aims at capturing item dependency, where each item is a node and dependency structure is learned from observables. The Aspect model is based on a latent cause model that introduces the notion of user com- munities or groups of items. With the successful application of the aspect model, various probabilistic models have been developed by researchers to complement its shortcomings. The flexible mixture model [Si and Jin, 2003] extends clustering algorithms for collaborative filtering by clustering both users and items simultane- ously without assuming that each user and item should only belong to a single cluster. The preference-based graphic model [Jin et al., 2003a] addresses the behaviorial distinction between users. That is, users with similar interests in items may have very different rating habits; some users tend to assign a higher rating to all items than other users. Popescul et al. [2001] extend the Aspect model by incorporating the content information of items in order to improve the quality of recommendation and to overcome rating data sparsity. Hofmann [2004] further extends the Aspect model to Probabilistic Latent Semantic Analysis (PLSA), which supports numerical ratings instead of categorical ratings (e.g. binary ratings). Pennock et al. [2000] propose a method combining both memory-based and model-based approaches and demonstrate that this combined approach provide better recommendations than pure memory-based or model- based collaborative approaches. Yu et al. [2002] also propose a probabilistic approach to collaborative filtering combining both approaches. In particular, it first learns a probabilistic model for each user’s preferences using active learning. Next, a mixture model of user profiles is applied for recommendations.
4.3.3 Limitations
Since collaborative filtering uses other users’ ratings, it can deals with any kind of content for universal rec- ommendation, even those that are quite different from what have been rated in nature. However, collaborative filtering has its own limitations :
Sparsity problem. In any recommender system, the number of ratings collected is usually very small • compared to the number of predictions to make. Effective prediction of ratings from a small number of examples is critical in this sense. One solution to this problem is to adopt user profile information or item 4 Filtering Techniques for Selection of Contents and Products 47
attributes when calculating user or item similarities. Alternatively, the sparsity problem can be addressed by applying an associative retrieval framework and related spreading activation algorithms to explore transitive associations among consumers through their past transactions and feedback [Huang et al., 2004]. Cold start problem. It is hard for pure collaborative filtering to recommend a new item where there is no • historical data and hence no user opinions about it. Music 5 in Table 4.2 is an example. The same holds for new users to the system.
Table 4.2. User bias problem. Item Jack Oliver Peter Tom Genre Music 1 5 4 3 Rock Music 2 4 4 3 Rock Music 3 4 3 Rock Music 4 4 3 Country Music 5 Rock Music 6 4 Rock
Ignorance problem. Collaborative filtering can be ignorant to certain hidden information (e.g. item cat- • egory) that could be used for better system performance. For instance in Table 4.2, collaborative filtering only utilities the rating columns but ignores the genre columns in practice. collaborative recommenders are based on the parts of ratings from users which may mislead the inferences made by systems. Here, Music 3 and 4 have the same ratings, and thus have the same chance to be recommended to user Jack by item-based filtering. However, if the system knows that Music 1,2 and 3 fall in the rock category and Music 4 in the country category, it can recommend Music 3 to Jack with a higher priority, provided that he is known to better appreciate rock music.
4.4 Hybrid Filtering
Although collaborative filtering is able to improve the quality of content-based recommendations based on the ratings of different users, it completely ignores other information that may be available from the semantic con- tents. On the other hand, content-based filtering does not suffer from this problem. As such, a hybrid approach which combines content-based and collaborative filtering helps to alleviate this problem. Hybrid recommender systems generally are sub-categorized into three different models. The first is the simple combination model, which combines results from the collaborative and content- based filters as shown in Fig. 4.2. For example, ProfBuilder [Wasfi, 1998] recommends web pages using both content-based and collaborative filters, each of which creates a separate recommendation list. Claypool et al. [1999] describe a hybrid approach for online news delivery, combining the two predictions using an adaptive weighted average. As the number of users accessing an item increases, so does the weight of the collaborative component. However, Claypool et al. do not clearly specify how to determine the weights of collaborative and content-based components as the system continues to operate. The second is the sequential combination model as shown in Fig. 4.3. In this model, user profiles are first constructed by a content-based filtering algorithm. Next, a collaborative filtering algorithm is applied to make predictions based on these user profiles, such as RAAP [Delgado et al., 1998] and Fab filtering [Balabanovic and Shoham, 1997]. RAAP uses a content-based filtering technique to help a user to classify domain specific information found on the Web, and also recommends URLs to other users with similar interests. To deter- mine the similarity of interests among users, a scalable Pearson correlation algorithm based on the web page categories is used. The Fab system also uses content-based techniques rather than user ratings to create user profiles. Hence the quality of predictions is completely dependent on the content-based filtering techniques, and inaccurate profiles can cause inaccurate correlations among users, thus yielding poor predictions. 48 Qing Li, Yuanzhu Peter Chen, and Zhangxi Lin
Rating Collaborative Matrix Filter Combination
Products Content-based Filter
Fig. 4.2. Simple combination.
Content- User Collaborative Products based Filter Profiles Filter
Fig. 4.3. Sequential combination.
The third is the synthetic combination model, in which both semantic contents and user ratings are applied to make recommendations; these include the probabilistic model by Popescul et al. [2001], Ripper system by Basu et al. [1998], and the music recommender by Li et al. [2004, 2007]. Ripper is a machine learning system using a combination of content information and rating data in an effort to produce better recommendations. Good et al. [1999] combine personal information filtering agents and user ratings to make recommendations. Popescul et al. [2001] provide a probabilistic model for unified collaborative and content-based recommenda- tions. Li et al. [2004, 2007] present a music recommender system based on an item-based probabilistic model, where items are classified into groups and predictions are made for users considering the Gaussian distribution of user ratings. Audio features and user ratings are utilized together for classification and recommendation.
Rating Matrix Combination Filter Products
Fig. 4.4. Synthetic combination.
4.5 Challenges for the Current Filtering Techniques
Recommender systems discussed so far can be extended in several ways to improve the scalability, data sparsity tolerance, and effectiveness of recommender systems. Such more advanced recommender systems can provide better recommendation capabilities. Here, we review some preliminary efforts in these directions and also identify various future research opportunities.
4.5.1 Scalability
Existing recommender systems can deal with thousands of consumers within a reasonable amount time, but the demand of modern e-commerce systems is to handle millions of consumers. On the other hand, more and more 4 Filtering Techniques for Selection of Contents and Products 49 products are available for consumers to select at an exponentially increasing speed. A robust and successful recommender system should have the ability to process large number of recommended items for numerous consumers. One possible approach to the scalability problem is to divide products or users into different groups. Sarwar et al. [2002] partition the users in a system using a clustering algorithm and threat each partition as subsystem. In particular, it first executes a clustering algorithm to create user groups. Next, for a given user, the selection of its nearest neighbors is restricted to groups that are close to this user. As a further extension, Xue et al. [2005] introduces a smoothing-based method where smoothed clusters are used for neighbor selection. Traditional filtering algorithms usually are implemented in one or several central servers which play a management role to collect information from individual users and make recommendations. However, ideas from the community of distributed computing can be borrowed for building more scalable filters. Han et al. [2004a,b] provide a distributed filtering algorithm, PipeCF, using a peer-to-peer (P2P) overlay based on Distributed Hash Tables (DHTs). However, the performance of DHTs under the dynamic conditions common for P2P systems is unknown. Tveit [2001] proposes a P2P collaborative system for mobile commerce based on Gnutella. Peers are represented as software assistant agents interfacing with a mobile customer. When a peer receives a query rating vector, it calculates the similarity to the rating vector carried by the message and caches the message at the node. If the similarity is greater than a threshold, the cached voting vector is sent back to the peer which had issued the query rating vector. Otherwise, the query rating vector is broadcast further to its neighbors. It is noticed that this system has the communication overhead as any P2P network with an unstructured overlay. Kim et al. [2008] provide a distributed collaborative filtering algorithm which is based on the Gnutella schema. It applies a self-organizing protocol using friend lists to create shortcuts among peers, which provides a logical network on top of Gnutella’s unstructured overlay. This feature makes it suitable work over any straightforward use of the Gnutella schema such as Tveit.
4.5.2 Data Sparsity
Data sparsity refers to the difficulty that most users rate only a small number of products and hence a very sparse user-item matrix is available. Consequently, the accuracy of the method is often quite poor. In addition, because of data sparsity, a recommender system may have insufficient ratings to make a recommendation. In particular, for memory-based filtering algorithms, user similarity is only defined between customers who have rated at least two products in common. Many pairs of customers have no correlation at all. In practice, many commercial recommender systems are used to evaluate large product sets (e.g. Amazon.com recommends books and CDnow recommends music albums). In these systems, even active customers may have rated well under 1% of the products (1% of 2 million books is 20,000 books — a large set on which to have an opinion). Accordingly, the system may not be able to provide a particular user with a non-trivial portion of all items in the system. To overcome data sparsity, Brand [2003] and Sarwar et al. [2000] apply the Singular Value Decompo- sition (SVD) technique to reduce the dimensionality of the customer-product ratings matrix. Others rely on the estimation of missing ratings based on the obtained ones. Xue et al. [2005] introduce a smoothing-based method to estimate unrated products. Hu and Lu [2006] also apply a hybrid predictive algorithm with smooth- ing. Their algorithm uses item-based methods to provide the basis for data smoothing and builds a prediction model based on both users’ and items’ aspects for better robustness against sparsity and for higher prediction accuracy. Wang et al. [2006] reformulate the memory-based collaborative filtering problem under a generative probabilistic framework, treating individual user-item ratings as predictors of missing ratings. The final rating is estimated by fusing predictions from three different sources. The complete model is more robust against data sparsity because different types of ratings are used in concert while additional ratings from similar users towards similar items are employed as a background model to smooth the predictions.
4.5.3 Effectiveness of Recommendation
The problem of developing good metrics to measure the effectiveness of recommendation has been extensively addressed in the recommender systems literature. Some examples of this work include [Herlocker et al., 1999, 50 Qing Li, Yuanzhu Peter Chen, and Zhangxi Lin
2004, Yang and Padmanabhan, 2001]. In most of the recommender systems literature, the performance evalua- tion of recommendation algorithms is usually done in terms of coverage and accuracy. The coverage measure is the percentage of items for which a recommender system is capable of providing predictions. The accuracy measures can take two forms. The statistical accuracy metric compares the estimated ratings against the actual ratings including Mean Absolute Error (MAE), root mean squared error, and correlation between predictions and ratings. Alternatively, it may also determine how well a recommender system can make predictions of high-relevance items. In this sense, an accuracy measure includes classical information retrieval measures of precision, recall, F-measure (a harmonic mean of precision and recall) in recommender systems [Adomavicius and Tuzhilin, 2005]. Despite their popularity, these empirical evaluation measures have some limitations. These measures are typically performed on test data that the users chose to rate. However, items that users choose to rate are likely to constitute a skewed sample, e.g. users may rate mostly the items that they are interested in. Understand- ably, it is expensive and time-consuming to conduct controlled experiments with users in the recommender systems settings. In addition, although crucial for measuring the accuracy of recommendations, the technical measures mentioned earlier often do not adequately capture “usefulness” and “quality” of recommendations. For example, as Wade [2003] observes for a supermarket application, recommending obvious items (such as milk or bread) that the consumer will buy anyway will produce high accuracy rates; however, it will not be very innovative to the consumer. Developing and studying the measures that would remedy the limitations described here indicates an interesting and important research direction.
4.5.4 Further Investigation
Despite the efforts in the fields mentioned above, other important issues have been explored include user pri- vacy, trustworthiness and context-aware recommendation. One of user concerns to use recommender systems freely and comfortably is user privacy. Users are usually • reluctant to disclose their private information such as purchase, reading, browsing records. However, most current filtering algorithms need to obtain user private information for further analysis and recommendation services. Some work [Canny, 2002, Polat and Du, 2003] has studied on how to protect user privacy in recommender systems. Current filtering techniques assume that user ratings are trustable and treat all users equally. However, some • authors [Kautz et al., 1997, Zacharia et al., 2000, Massa and Avesani, 2004, O’Donovan and Smyth, 2005] argue that the opinions of experts should be more emphasized than that of novices. Context-aware recommendation [Brunato and Battiti, 2003, Byung Kwon, 2003, Lemire et al., 2005, Chen, • 2005, Yu et al., 2006] study on how to apply contextual information to improve a system performance and provide on-demand services in a certain circumstance. For instance, a context-aware restaurant rec- ommender system provides a traveller with recommended restaurants which are close to his/her current location.
4.6 Conclusion
Research in recommender systems has made significant progresses over last decades when numerous content- based, collaborative, and hybrid methods have been proposed and several commercial systems have been devel- oped. However, despite all of these advances, current recommender systems still deserve further improvements to make them more effective in a broader range of applications.
References
G. Adomavicius and A. Tuzhilin. Toward the next generation of recommender systems: a survey of the state-of- the-art and possible extensions. IEEE Transactions on Knowledge and Data Engineering, 17(6):734–749, 2005. 4 Filtering Techniques for Selection of Contents and Products 51
P. G. Anick, J. D. Brennan, R. A. Flynn, D. R. Hanssen, B. Alvey, and J. M. Robbins. A direct manipula- tion interface for Boolean information retrieval via natural language query. In 13th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, Cambridge, Massachusetts (USA), 1989. R. Baeza-Yates and B. Ribeiro-Neto. Modern information retrieval. Addison-Wesley, 1999. M. Balabanovic and Y. Shoham. Fab: Content-based, collaborative recommendation. Communications of the ACM, 40(3):66–72, 1997. C. Basu, H. Hirsh, and W. Cohen. Recommendation as classification: Using social and content-based informa- tion in recommendation. In 15th National Conference on Artificial Intelligence, Madison, Wisconsin (USA), 1998. D. Billsus and M. J. Pazzani. User modeling for adaptive news access. User Modeling and User-Adapted Interaction, 10(2):147–180, 2000. M. Brand. Fast online SVD revisions for lightweight recommender systems. In 3rd SIAM International Con- ference on Data Mining, San Francisco, California (USA), 2003. J. S. Breese, D. Heckerman, and C. Kadie. Empirical analysis of predictive algorithms for collaborative filtering. In 14th Conference on Uncertainty in Artificial Intelligence, Madison, Wisconsin (USA), 1998. M. Brunato and R. Battiti. PILGRIM: A location broker and mobility-aware recommendation system. In 1st IEEE International Conference on Pervasive Computing and Communications, Dallas, Texas (USA), 2003. O. Byung Kwon. “i know what you need to buy”: Context-aware multimedia-based recommendation system. Expert Systems With Applications, 25(3):387–400, 2003. J. Canny. Collaborative filtering with privacy. In IEEE Symposium on Security and Privacy, Oakland, California (USA), 2002. A. Chen. Context-aware collaborative filtering system: Predicting the user’s preference in the ubiquitous com- puting environment. In Extended Abstracts of the Conference on Human Factors in Computing Systems, Portland, Oregon (USA), 2005. M. Claypool, A. Gokhale, T. Miranda, P. Murnikov, D. Netes, and M. Sartin. Combining content-based and collaborative filters in an online newspaper. In ACM SIGIR Workshop on Recommender Systems, Berkeley, California (USA), 1999. J. Delgado, N. Ishii, and T. Ura. Content-based collaborative information filtering. Lecture Notes in Computer Science, 1435:206–215, 1998. D. Goldberg, D. Nichols, B. M. Oki, and D. Terry. Using collaborative filtering to weave an information tapestry. Communications of the ACM, 35(12):61–70, 1992. N. Good, J. B. Schafer, J. A. Konstan, A. Borchers, B. M. Sarwar, J. L. Herlocker, and J. T. Riedl. Combining collaborative filtering with personal agents for better recommendations. In 16th National Conference on Artificial Intelligence, Orlando, Florida (USA), 1999. D. Gupta, M. Digiovanni, H. Narita, and K. Goldberg. Jester 2.0: A new linear-time collaborative filtering algo- rithm applied to jokes. In Workshop on Recommender Systems, in conjunction with ACM SIGIR, Berkeley, California (USA), 1999. P. Han, B. Xie, F. Yang, and R. Shen. A scalable P2P recommender system based on distributed collaborative filtering. Expert Systems with Applications, 27(2):203–210, 2004a. P. Han, B. Xie, F. Yang, J. Wang, and R. Shen. A novel distributed collaborative filtering algorithm and its implementation on P2P overlay network. In 8th Pacific-Asia Conference on Advances in Knowledge Discovery and Data Mining, Sydney, Australia, 2004b. D. B. Hauver and J. C. French. Flycasting: using collaborative filtering to generate a playlist for online radio. In 1st International Conference on Web Delivering of Music, Florence, Italy, 2001. J. L. Herlocker, J. A. Konstan, and J. T. Riedl. An algorithmic framework for performing collaborative filter- ing. In 22nd Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, Berkeley, California (USA), 1999. J. L. Herlocker, J. A. Konstan, L. G. Terveen, and J. T. Riedl. Evaluating collaborative filtering recommender systems. ACM Transactions on Information Systems, 22(1):5–53, 2004. 52 Qing Li, Yuanzhu Peter Chen, and Zhangxi Lin
T. Hofmann. Collaborative filtering via Gaussian probabilistic latent semantic analysis. In 26th Annual Inter- national ACM SIGIR Conference on Research and Development in Information Retrieval, Toronto, Canada, 2003. T. Hofmann. Latent semantic models for collaborative filtering. ACM Transactions on Information Systems, 22 (1):89–115, 2004. R. Hu and Y. Lu. A hybrid user and item-based collaborative filtering with smoothing on sparse data. In 16th International Conference on Artificial Reality and Telexistence, Hangzhou, China, 2006. Z. Huang, H. Chen, and D. Zeng. Applying associative retrieval techniques to alleviate the sparsity problem in collaborative filtering. ACM Transactions on Information Systems, 22(1):116–142, 2004. R. Jin, L. Si, and C. X. Zhai. Preference-based graphic models for collaborative filtering. In 19th Conference on Uncertainty in Artificial Intelligence, Acapulco, Mexico, 2003a. R. Jin, L. Si, C. X. Zhai, and J. Callan. Collaborative filtering with decoupled models for preferences and ratings. In 12th ACM International Conference on Information and Knowledge Management, New Orleans, Louisiana (USA), 2003b. H. Kautz, B. Selman, and M. Shah. Referral Web: Combining social networks and collaborative filtering. Communications of the ACM, 40(3):63–65, 1997. B. M. Kim, Q. Li, A. E. Howe, and Y. P. Chen. Collaborative web agent based on friend networks. Applied Artificial Intelligence, 22, 2008. M. Kim and V. V. Raghavan. Adaptive concept-based retrieval using a neural network. In ACM SIGIR Workshop on Mathematical/Formal Methods in Information Retrieval, Athens, Greece, 2000. J. H. Lee. Ranking documents in thesaurus-based Boolean retrieval systems. Information Processing and Management, 30(1):79–91, 1994. D. Lemire, H. Boley, S. McGrath, and M. Ball. Collaborative filtering and inference rules for context-aware learning object recommendation. International Journal of Interactive Technology and Smart Education, 2 (3):179–188, 2005. Q. Li and B. M. Kim. Clustering approach for hybrid recommender system. In IEEE/WIC International Conference on Web Intelligence, Halifax, Canada, 2003. Q. Li, B. M. Kim, D. H. Guan, and D. Whan Oh. A music recommender based on audio features. In 27th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, Sheffield, United Kingdom, 2004. Q. Li, S. H. Myaeng, and B. M. Kim. A probabilistic music recommender considering user opinions and audio features. Information Processing and Management, 43(2):473–487, 2007. P. Massa and P. Avesani. Trust-aware collaborative filtering for recommender systems. In International Con- ference on Cooperative Information Systems, Agia Napa, Cyprus, 2004. O. Nasraoui and M. Pavuluri. A context ultra-sensitive approach to high quality web recommendations based on web usage mining and neural network committees. In 3rd International Workshop on Web Dynamics, in conjunction with WWW, New York, USA, 2004. J. O’Donovan and B. Smyth. Trust in recommender systems. In 10th International Conference on Intelligent User Interfaces, San Diego, California (USA), 2005. Y. Ogawa, T. Morita, and K. Kobayashi. A fuzzy document retrieval system using the keyword connection matrix and a learning method. Fuzzy Sets and Systems, 39(2):163–179, 1991. D. M. Pennock, E. Horvitz, S. Lawrence, and C. L. Giles. Collaborative filtering by personality diagnosis: A hybrid memory-and model-based approach. In 16th Conference on Uncertainty in Artificial Intelligence, Stanford, California (USA), 2000. H. Polat and W. Du. Privacy-preserving collaborative filtering. International Journal of Electronic Commerce, 9(4):9–35, 2003. A. Popescul, L. H. Ungar, D. M. Pennock, and S. Lawrence. Probabilistic models for unified collaborative and content-based recommendation in sparse-data environments. In 17th Conference on Uncertainty in Artificial Intelligence, Seattle, Washington (USA), 2001. P. Resnick, N. Iacovou, M. Suchak, P. Bergstrom, and J. T. Riedl. GroupLens: an open architecture for collab- orative filtering of netnews. In ACM Conference on Computer-Supported Cooperative Work, Chapel Hill, North Carolina (USA), 1994. 4 Filtering Techniques for Selection of Contents and Products 53
S. E. Robertson and K. Sparck Jones. Relevance weighting of search terms. Journal of the American Society for Information Science, 27(3):129–46, 1976. G. Salton and C. Buckley. Term-weighting approaches in automatic text retrieval. Information Processing and Management, 24(5):513–523, 1988. B. M. Sarwar, G. Karypis, J. A. Konstan, and J. T. Riedl. Application of dimensionality reduction in recom- mender system-a case study. In International Workshop on Web Mining for E-Commerce – Challenges and Opportunities, Boston, Massachusetts (USA), 2000. B. M. Sarwar, G. Karypis, J. A. Konstan, and J. Reidl. Item-based collaborative filtering recommendation algorithms. In 10th International World Wide Web Conference, Hong Kong, China, 2001. B. M. Sarwar, G. Karypis, J. A. Konstan, and J. T. Riedl. Recommender systems for large-scale e-commerce: Scalable neighborhood formation using clustering. In 5th International Conference on Computer and Infor- mation Technology, Shanghai, China, 2002. U. Shardanand and P. Maes. Social information filtering: Algorithms for automating “word of mouth”. In SIGCHI Conference on Human factors in Computing Systems, Denver, Colorado (USA), 1995. L. Si and R. Jin. Flexible mixture model for collaborative filtering. In 12th International Conference on Machine Learning (ICML), Washington, D.C. (USA), 2003. D. B. Terry. A tour through Tapestry. In Proceedings of the Conference on Organizational Computing Systems, Milpitas, California (USA), 1993. A. Tveit. Peer-to-peer based recommendations for mobile commerce. In 1st International Workshop on Mobile Commerce, Rome, Italy, 2001. J. Verhoeff, W. Goffman, and J. Belzer. Inefficiency of the use of Boolean functions for information retrieval systems. Communications of the ACM, 4(12):557–558, 1961. W. Wade. A grocery cart that holds bread, butter and preferences. New York Times, 2003. J. Wang, A. P. de Vries, and M. J. T. Reinders. Unifying user-based and item-based collaborative filtering approaches by similarity fusion. In 29th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, Seattle, Washington (USA), 2006. A. M. A. Wasfi. Collecting user access patterns for building user profiles and collaborative filtering. In 4th International Conference on Intelligent User Interfaces, Los Angeles, California (USA), 1998. R. Wilkinson and P. Hingston. Using the cosine measure in a neural network for document retrieval. In 14th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, Chicago, Illinois (USA), 1991. K. Wittenburg, D. Das, W. Hill, and L. Stead. Group asynchronous browsing on the World Wide Web. In 4th International World Wide Web Conference, Boston, Massachusetts (USA), 1995. G. R. Xue, C. Lin, Q. Yang, W. S. Xi, H. J. Zeng, Y. Yu, and Z. Chen. Scalable collaborative filtering using cluster-based smoothing. In 28th Annual International ACM SIGIR Conference on Research and Develop- ment in Information Retrieval, Salvador, Brazil, 2005. Y. Yang and B. Padmanabhan. On evaluating online personalization. In Workshop on Information Technology and Systems, Montréal, Canada, 2001. K. Yu, X. Xu, J. Tao, M. Ester, and H. P. Kriegel. Instance selection techniques for memory-based collaborative filtering. In 2nd SIAM International Conference on Data Mining, Arlington, Virginia (USA), 2002. Z. Yu, X. Zhou, D. Zhang, C. Y. Chin, X. Wang, and J. Men. Supporting context-aware media recommendations for smart phones. IEEE Pervasive Computing, 5(3):68–75, 2006. G. Zacharia, A. Moukas, and P. Maes. Collaborative reputation mechanisms for electronic marketplaces. De- cision Support Systems, 29(4):371–388, 2000.
5 Personalization and Contextualization of Media
Marcus Specht1, Andreas Zimmermann2 and Tim de Jong1
1 Educational Technology Expertise Centre, Open Universiteit Nederlands Valkenburgerweg 177, 6419 AT Heerlen {marcus.specht,tim.dejong}@ou.nl 2 Fraunhofer-Institute for Applied Information Technology FIT Schloss Birlinghoven, 53754 Sankt Augustin, Germany Andreas.Zimmermann@fit.fraunhofer.de
Summary. Personalization allows users to obtain information that is adapted to their needs, goals, knowl- edge, interests or other characteristics. User models deliver the main parameters for selecting and adapting information presentation to the individual user. Nevertheless the individual user is always in a current context and contextualization of information systems is an important research topic in several areas of research like context-aware systems, ubiquitous computing, or applications areas contextualized learning support. Contextu- alization complements personalization so that environmental states or the context of use can also be taken into account. The paper will show how personalization and contextualization approaches are related in different phases of the adaptation process and how the approaches can be used in a complementary way for personalized and contextualized information systems. The analysis will guide the reader through the phases of informa- tion acquisition in adaptive information systems, the representation of personal and contextual information, and the selection adaptive methods applicable based on personal profiles and on context information available. Furthermore application examples will be described which integrate personalization and contextualization in ubiquitous computing for learning and information delivery. The paper will present an integrated approach for describing and planning personalized and contextualized information systems which allow the individual to be- come a digital nomad making use of ubiquitous computing facilities while still having a personal information space. The strengths and weaknesses of today’s approaches will be discussed.
5.1 Introduction
In the last decade personalization has been an issue of discussions in different research fields [Kobsa and Wahlster, 1989, Brusilovsky, 1996, Kobsa, 1996, Kim, 2002]. Recently also for a variety of products personal- ization is an essential added value and more and more personalized services allow for personalized user portals in the Web 2.0. In the past personalization was mostly interpreted as an adaptation of a machine service to a human user based on a profile of a user. Examples for adaptive methods and personalization can be found in different research areas as intelligent tutoring systems [Carbonell, 1970, Clancey, 1987, Anderson et al., 1989], adaptive user interfaces [Carroll and Carrithers, 1984, Brusilovsky et al., 1995], adaptive hyperme- dia [Brusilovsky, 1996], intelligent multimedia or intelligent agents [Feiner and McKeown, 1995, Rickel and Johnson, 1997]. In the last years new research approaches like context-aware computing [Dey and Abowd, 1999, Gross and Specht, 2001, Oppermann and Specht, 2000, Schilit et al., 1994, Zimmermann et al., 2005] extended the idea of personalization for not only adapting to a personal profile of a user (preferences, knowledge, or cognitive characteristics) but also the context of use [Oppermann and Specht, 2000] became more and more important. In the effort to create added value in ubiquitous environments, the current context of use becomes more im- portant as the computer disappears. From our point of view, it is not enough to supply content or services that 56 Marcus Specht, Andreas Zimmermann and Tim de Jong consider single environmental or user characteristics; we need to identify approaches for the integration and interpretation of different sensing components for modeling the user and the context more appropriately. The combination of context sensor data with user modeling appears to be one of the key challenges for personalized information systems of the future. Jameson [2001] describes how modeling the user and modeling the context of the user can both lead to more valid assumptions about the current needs and goals of a user. In this sense, delivering information that is to the point has two main facets: Personalization allows users to obtain information that is adapted to their needs, goals, knowledge, interests • or other characteristics. User models deliver the main parameters for selecting and adapting information presentation to the individual user. Contextualization complements personalization so that environmental states or the context of use can also • be taken into account for adaptation. Furthermore the combination of approaches for user modeling and user model inference can deliver much more valid data about users when they are combined with contextual information. Personalization and contextualization can be seen as specialized forms of adaptation. At the core of adaptive systems are adaptive methods, which take an adaptation mean as a certain user characteristic or a part of the current user context and adapt different adaptation targets to this adaptation mean [Leutner, 1992]. From the background of user modeling and user adapted interaction different phases of adaptation are often differentiated ranging from user model acquisition, user model representation, information inference, adaptation, and evaluation of adaptation. In this chapter we will outline some works from the background of adaptive systems based on user modeling and show how this work is related to context-aware systems. As one result it can be seen that personalization and contextualization can be seen as complementary concepts, especially in newer approaches of ubiquitous computing for contextualized services [Zimmermann et al., 2005]. In the following we will also give some examples for the relation of personalization and contextualization and apply our analysis in the field of contextualized media for learning. Contextualized media for learning is a relatively new research area that combines the latest developments in ubiquitous and context aware computing with pedagogical approaches relevant to structure more situated and context aware learning support. Searching for different backgrounds of mobile and contextualized learning authors have identified the relations between existing educational paradigms and new classes of mobile applications for education [Naismith et al., 2004]. Furthermore best practices of mobile learning applications have been identified and discussed in focused work- shops [Stone et al., 2002, Tatar et al., 2003]. We consider this field as an ideal evaluation area for the integration of personalization and contextualization as it also strongly integrates theoretical backgrounds about cognition in context and therefore emphasis the differences between personalization and contextualization and added values of the complementary use in educational artefacts.
5.2 Adaptation: Personalization and Contextualisation
In adaptive educational hypermedia a variety of research work about questions on how to adapt media to individuals and groups of learners has been done [Leutner, 1992, Brusilovsky, 1996, Weber and Specht, 1997, de Bra and Calvi, 1998, Specht, 1998]. From our point of view an adaptation process in contextualized learning applications needs to consider four key questions: 1. Why does the system adapt? 2. How does the system gather the information to adapt to? 3. What parts or components of the learning process are adapted? 4. What information does the system use for adaptation? Answering these four questions requires a certain amount of adaptation knowledge. The quality of the adaptation knowledge is a crucial factor for making the personalization and contextualization of applications a benefit rather than a hindrance [Bellotti and Edwards, 2001]. Basically, the idea of adaptation stems out of system dynamics in which self regulative systems have to adapt to changing conditions in an environment or internal and external feedback loops. The application of 5 Personalization and Contextualization of Media 57 adaptation in computer systems tries to mainly adapt the interaction and the facilities of a system to the changing needs of a user and the differences between users. So on the one hand changing parameters in an environment and associated conditions can trigger the execution of an adaptation. Many characteristics might be taken into account as catalysts for an adaptation process. They can be clustered into three main categories: Inter-individual differences address varieties among several users along manifold dimensions. Physi- • ological characteristics like disabilities are of major concern for application designers if they want to have their system accepted by a large community. The consideration of user preferences like language, colour schemes, modality of interaction, menu options or security properties, and numberless other per- sonal favourite preferences is one of the most popular sources of adaptation and can be reused in different applications. An important variable for the content adaptation is the awareness of the user’s interests and disinterests. The user’s interest is a confounded variable that filters perception and the acceptance of adap- tation. Additionally, a broad spectrum of personality characteristics like emotions, self-confidence, moti- vation, beliefs, which are difficult to assess automatically have an impact on the adaptation process and its evaluation. The same holds for the user’s level of factual, general and domain-specific knowledge (e.g. beginners or experts), which is a valuable source of adaptive operations [Specht and Oppermann, 1998]. Intra-individual differences consider the evolution and further development of a single user, as well as • the task over time. A static system falls short of changing user requirements as the user’s activities and goals evolve. In an extreme case users are overstrained by the system in the beginning and perceive the same system as cumbersome and restricted as the user’s expertise increases. In the same manner, the need for a higher flexibility of computer systems is pushed by the changing of the tasks to be accomplished with such a system. Environmental differences result from the mobility of computing devices, applications and people, which • leads to highly dynamic computing environments. Unlike desktop applications, which rely on a carefully configured and largely static set of resources, ubiquitous computing applications are subject to changes in available resources such as network connectivity and input and output devices. Moreover, they are fre- quently required to cooperate spontaneously and opportunistically with previously unknown software ser- vices in order to accomplish tasks on behalf of users. Thus, the environment surrounding an application and its user is a major source to justify adaptation operations.
Contextualized learning applications mainly focus on the pedagogical models behind the adaptation [Pask, 1964, Frank, 1965, Salomon, 1975]. Classical educational hypermedia system mainly adapted according for compensation of knowledge deficits, ergonomic reasons, or adaptations to learning styles for an easier intro- duction into a topic. Location-based services are an example for the type of adaptation mainly to environmental differences. In those applications the individual possibility for encoding and decoding information is one in- teresting aspect for better understanding artifacts in context. Furthermore the authentic collaboration on a topic is another example in principle all differences from inter-, intra-, and environmental differences play a role for best learning support. As for example co learners can be selected according to the learning task of an individ- ual, his/her level of the domain, and the current location of persons or the availability of local communication facilities. Applications from the field of contextualised learning can make use of context in different ways and the role context plays in these applications is important. Four roles of context can be identified: Filter. The context serves as a basis of the preparation of a selection of choices. The usage of context for • decision support configures queries carried out on possibly large data and information sets and generates a specific subset from this set. Trigger. The context initiates an activity based on the occurrence of some significant event, i.e. changes in • the context. The usage of context as a trigger can cause various actions, depending on the application. The context as a trigger either activates or deactivates an action. Annotation. The context is used for the generation of metadata in order to tag this information to physical • or virtual objects. The captured context can be matched against a filter context in a later retrieval. Reflection. The context serves as a pure information source. This usage of context comprises the display • and recording of context information as well as the exploitation of context information for the parameterised execution of actions (e.g. reconfiguration). 58 Marcus Specht, Andreas Zimmermann and Tim de Jong
An obvious observation of these four roles of context clarifies that a single contextualised learning applica- tion can make use of context in more than one way. The context of such an application takes over several roles either in parallel, temporally interlinked or in a combination of both. Each role of the context might emphasise another part of the context description. Some contextualised learning applications allow the learner to take over control and play the role of the context. Then, the user pretends to be in a certain context [Brown, 1998] and filters, triggers, annotates or reflects through the manual specification of context information.
5.3 Contextualisation complements Personalisation
Inter- and intra-individual differences particularly refer to adaptation effects based on changes in characteris- tics of users. Jameson [2003] defines all systems that automatically perform an adaptation to the individual user in some nontrivial way as user-adaptive systems. User-adaptive systems emphasise dialog systems and demand for a detailed, explicit model of the user to adapt the behaviour of the system. User modelling compo- nents are constructed in such a way that they are tightly intertwined with the application. Prominent functions for user-adaptive systems are product recommendation, interface adaptation, learner support, or information retrieval [Kröner, 2001]. The e-commerce sector prefers the more popular term personalization focusing on systems that tailor products, services and information to better cater to each individual user. Their primary driver for using personalization is the belief that they can establish a stronger relationship with a customer through treating them individually (see Amazon3 or BroadVision4 as examples). If now environmental differences are integrated with inter- and intra-individual differences into a coherent whole, the notion of context becomes apparent. The notion of context varies across many different research areas. In general, context is understood as a knowledge type that facilitates a selective and focused processing of information. In context-aware computing a variety of notions and interpretations has developed over the years. Zimmermann et. al distinguish between definitions by synonym or definitions by example which mainly name and describe certain context parameters as location, identity, time, temperature, noise, as well as beliefs, desires, and commitments and intentions [Zimmermann et al., 2007]. Furthermore they introduce an operational definition of context describing following main categories of context information:
Individuality context includes information about objects and users in the real world as well as information • about groups and the attributes or properties the members have in common. Time context: from simple points in time to ranges, intervals and a complete history of entities. • Location contexts are divided into quantitative and qualitative location models, which allow to work with • absolute and relative positions. Activity context reflects the entities goals, tasks, and actions. • Relations context captures the relation an entity has established to other entities, and describes social, • functional, and compositional relationships.
5.4 Acquisition of Context Information
The realization of adaptation mechanisms in contextualized learning applications requires the coding of knowl- edge for observing the learners’ context. Traditionally and factually, the learners’ interaction with the appli- cation is the most important source of adaptivity. In here lies the biggest inter-individual and intra-individual varieties and in here lies the biggest need for adaptation during the usage of an application. Generally, over a certain time period the application needs to detect and record relevant indicators regarding context information, implicitly or explicitly [Jameson, 2003]. The knowledge of how to acquire context includes the selection of the appropriate acquisition method to collect information about learners to adapt to. These methods supply information that is provided explicitly by the user, implicitly through observations or from external sources:
3 http://www.amazon.com 4 http://www.broadvision.com 5 Personalization and Contextualization of Media 59
Questionnaires or forms. Questionnaires and forms present an effective means to capture information • explicitly provided by the learner. The learner fills in questionnaires and delivers a variety of valuable information to the application. This sort of input events oftentimes enables the application to learn about the learner’s properties, preferences, interests, task, etc. because they are explicitly specified by the learner. Tests. Like questionnaires and forms, tests are filled in by the learners in order to make information avail- • able to the contextualized learning application. In general, tests serve as means of the determination of the learner’s knowledge level, competence, and expertise. In addition, third party evaluation, assessments and results of tests (e.g. through tutors in learning systems) can be supplied to the application. Sensors. As additional sources of information and technical components, sensors measure physical or • chemical properties of the learners and their environments. Sensors quantify temperature, humidity, pres- sure, sound, lightness, magnetism, acceleration, force, and many more other properties of the environment. Furthermore, sensors measure the learner’s blood pressure, body temperature or eye movements as well as the location of the learner in- and outdoor. Queries. Furthermore, there exists a multiplicity of external sources like databases, pools or external appli- • cations that manage valuable information (e.g. learner profiles, weather forecasts). Information cannot be gained exclusively through direct interaction of the learner with the application or through observation of the learner. Additional queries to external information sources may augment the application with a variety of valuable information.
The careful selection of appropriate acquisition methods requires a lot of knowledge because one acquisi- tion method might deliver valuable information for one domain, but might be useless in another. The determination of the learner’s knowledge exemplifies possible difficulties in the selection process: us- ing a questionnaire the learner in fact specifies her knowledge explicitly. However, this input may be afflicted with several drawbacks because the learner might have limited self-assessment capabilities and may overes- timate her knowledge. On the other hand, if the system uses sensors for the observation of the learner, the determination of her knowledge might result from weak interpretation methods like “the longer the learner accesses a webpage the more knowledge she will gain”. Any implementation of the acquisition procedure will always be subjective and fail in covering the entire spectrum of learner personalities.
5.5 Targets of Adaptation
Targets of adaptation are parts of the application that is adapted by the adaptive method. An adaptation of these application parts may be immediately visible or noticeable to the learner or the adaptation effects may be hidden from the learner and display their impact at a later point in time. Contextualized learning applications consider five properties or parts of the system that can be tailored to the context and the learner: 1. Human-computer interaction, i.e. the modality needed to enter commands or data, and receive informa- tion and services. 2. Information presentation, i.e. the methods and coding required for receiving and displaying content (front-end). 3. Functionality, i.e. the features needed to perform tasks (back-end) and the ability of the application to solve a single task or a set of tasks. 4. Information and service selection, i.e. the information content, density and depth as well as the function- ality and complexity of necessary services. 5. Knowledge base, i.e. the collection, organization, and retrieval of the knowledge about and the model of the learner, the context and the application. Examples for adaptation targets taken from the domain of contextualized learning application can be the pace of the instruction [Tennyson and Christensen, 1988, Leutner, 1992] that can be modified based on diagnos- tic modules embedded in the learning process or adaptation of content presentations, the sequencing of contents and others. Taking into account contextual information enables new interesting applications as the selection of information based on locations or even the streaming of media to a mobile user in changing contexts. 60 Marcus Specht, Andreas Zimmermann and Tim de Jong 5.6 The Adaptation Process in Contextualized Learning Applications
The way the system reacts or adapts is based on model assumptions on user needs, heuristics or ontological models of the application domain. In most adaptive educational hypermedia applications a learner model is the basis for the adaptation of the previously given parameters of the learning process. Nevertheless there are a several examples where the adaptation takes place not only to the learner knowledge, preferences, interests, cognitive capabilities, but also to tasks and learner goals. In contextualized learning the information used for adaptation is extended by the environmental parameters. The inference methods of the adaptive system can gain precision from the additional information of environmental sensors. The adaptation process controls the way how the application behavior is adapted according to changes in the context. Dieterich et al. [1993] assume four phases of one cycle of the adaptation process: Initiating the process cycle, proposal of alternatives, deciding on one alternative and the execution of an adaptation. In the particular case of contextualized learning applications, this four-phase process needs to be framed by two extensions: identification of the need for adaptation and retaining the adaptation after it has been executed. In summary, the six phases of the adaptation process cycle can be described as follows:
Identify. Before an adaptation process cycle can be initiated, the need for an adaptation has to be identified • based on changes in the context. Initiate. If the need for adaptation is sufficient either the user or the system initiates the adaptation process • cycle. Propose. After the process cycle is initiated, one or more alternatives for the adaptation method need to be • proposed. Decide. During the decision phase one alternative of the set of adaptation methods proposed in the preced- • ing step is selected. Execute. Then, the selected adaptation method is performed using a set of adaptation operations that are • conducted by either the user or the system. Retain. After the execution of the selected adaptation method, the adaptation process cycle finishes with • the retaining of the adaptation for measuring the success of the process cycle or for a later inspection.
As basic knowledge-based systems, rule-based systems present a popular way to deterministically express adaptation knowledge because of their simple implementation. The interpretation of rules leads to decisions that determine the role of the context during the course of the adaptation process cycle. As a basic classification schema the introduced questions can describe a variety of adaptation methods, an overview can be seen in Table 5.1.
5.7 Personalized and Contextualized Media Applied in Learning Support
Situated learning as introduced by Wenger and Lave [1991] states the importance of knowledge acquisition in a cultural context and the integration in a community of practice. Learning in this sense must not only be planned structured by a curriculum but also by the tasks and learning situations and the interaction with the social environment of the learner. This is often contrasted with the classroom-based learning where most knowledge is out of context and presented de-contextualized. On the one hand the process of contextualization and de- contextualization might be important for abstraction and generalization of knowledge on the other hand in the sense of cognitive apprenticeship [Collins et al., 1989] it is reasonable to guide the learner towards appropriate levels and context of knowledge coming from an authentic learning situation. From a constructivist point of view not only knowledge is always contextualized and but also the construc- tion of knowledge, e.g. learning is always situated within its application and the community of practice [Mandl et al., 1995]. Stein defines four central elements of situated learning where the content emphasizes higher order thinking rather than acquisition of facts, the context for embedding the learning process in the social, psy- chological, and material environment in which the learner is situated, the community of practice that enables reflection and knowledge construction, and the participation in a process of reflecting, interpreting and nego- tiating meaning [Stein, 2003]. From the perspective of situated learning several requirements for new learning 5 Personalization and Contextualization of Media 61
Adaptive methods What is adapted? To which features? Why? Learning goal Learner Didactical reasons Content Preferences Preference model • Teaching method • Usage • Compensation of deficits • • Previous knowledge • Reduction of deficits Teaching style • Professional background • • Ergonomic reasons Media selection Knowledge • • Sequence Interests Efficiency • • Time constraints Goals • Effectiveness • • Help Task • Acceptance • • Complexity • Presentation • ... • Hiding • Dimming • Annotation • Extensions in contextualized learning Presentation Context sensors Situation 3D sound User location Authenticity of learning situa- • Augmented reality displays • Time • tions • Distribution to different con- • Light Situated collaboration • texts • Noise • Active construction of knowl- • Other user’s locations • edge •
Table 5.1. An overview for classification of adaptive methods.
tools can be stated like: use authentic problems, allow multiple perspectives, enable learning with peers and so- cial interaction within communities, enable active construction and reflection about knowledge. A shift towards a new tradition of online learning is described by Herrington et al. [2002]. Moreover the idea of situated learning is also closely related to the ideas of blended learning and learning on demand especially in educational systems for adults and at the workplace [Oppermann and Specht, 2006]. An important point that is not taken into account by a lot of new approaches for delivering personalized learning on demand is the aspect that the need (demand) for knowledge and learning arises in a working context with the motivation for solving specific problems or understanding problem situations. This notion of personalized learning on demand in the workplace exemplifies the potential of contextualized learning in the workplace. Learners that identify a problem in a certain working situation are highly motivated for learning and acquiring knowledge for problem solving. They have a complex problem situation as a demand, which can be used for delivering learning content adapted to their situation. Furthermore not only the delivery of content into a certain context or practice is needed but also interaction facilities must be provided which allow an appropriate interaction and cooperation with educational systems. The contextualization of the learning on demand can not only be seen from the point of view of an ac- tual problem or learning situation but also in a longer lasting process of learning activities that are integrated. Different learning activities are combined in blended learning approaches where the preparation for a task, up- dates on base knowledge, then the application in an actual working situation and the documentation of problem solutions and the reflection about one’s activities evaluates that process. In the literature for mobile and contextualized learning support different examples can be found for utilizing those context parameters to enable learning. Several projects have looked at how to use and enrich contents with contextual metadata [EQUATOR Project, 2003, Specht and Kravcik, 2006]. Mostly interesting new approaches 62 Marcus Specht, Andreas Zimmermann and Tim de Jong in context-aware systems see the main strength in combining different context parameters for user support. Even more new approaches tend to combine different forms of metadata about learning objects and media to allow for flexible and contextualized learning support. In the MACE project the combination of various types of content, usage, social and contextual metadata enables users to develop multiple perspectives and navigation paths that effectively lead to experience multiplication for the learner himself [Stefaner et al., 2007]. Identity context is also often combined with other forms of context. One specific example of such a combination is given by Ogata and Yano [2004b] who present CLUE, a system for learning English in real-world situations. CLUE uses (i) a learner profile to adapt the learning content to the learner’s interest and (ii) location information to link objects/locations to suitable English expressions, i.e. appropriate learning content. Likewise, Bo [2002] combines a user profile and user position, to facilitate personalised and location-based information delivery. AwarePhone [Bardram and Hansen, 2004] uses several context-parameters at the same time. First of all, location is used to locate fellow employees within the hospital. Second, a calendar artefact is used to capture and share time context and also indicate the activity of a user at a certain moment. The activity is furthermore given by a shared status message. The combination of these three context parameters lead to what the writers call “context-mediated social awareness”. Environmental context information is used in several systems, most notably QueryLens [Konomi, 2002] which focuses on information sharing using smart objects. Moreover, the TANGO system presented by Ogata and Yano [2004a] and the Musex system [Yatani et al., 2004] detect objects in the vicinity by using RFID tags. Moop [Mattila and Fordel, 2005] couples a GPS location to observations/information gathered in the field for later analysis in the classroom. Wallop [Farnham et al., 2004] allows its users to discover social relationships and provides social awareness by storing and analysing social context information; to derive the social context communication patterns system interactions and co-occurrence information were analysed.
5.8 Conclusion
The chapter has illustrated the common principles of personalized and contextualized media with the concept of adaptive methods and their parameters. As adaptation mean a system can focus on intra-individual, inter- individual differences or environmental changes and differences. It has been shown that both the personalization of systems to inter- and intra-individual differences as also the contextualization to environmental changes can be used in a complementary way to create new forms of adaptation with effective learning support. In the adaptation process personalization and contextualization differ mostly in the form triggers for adaptive methods are used. In user-adaptive systems mostly users take decisions or even sometimes identify the adaptive method to be instantiated or applied. In context adaptive systems the available information resources for adaptation become highly complex and that leads on the one hand to a need for more transparent user interfaces for contextualization interfaces as also the need for more semiautomatic processes of information aggregation and validation for adaptive contextualized support. In that sense it also becomes important that contextual information can not only be used for filtering and triggering adaptation but can play a role for annotation and reflection for supporting learning in the context of use. In the last part several examples from the field of mobile and ubiquitous learning support have been pre- sented that use context parameters for a more effective learning and build on personalized learning support.
References
J. R. Anderson, F. G. Conrad, and A. T. Corbett. Skill acquisition and the LISP tutor. Cognitive Science, 13: 467–505, 1989. J. E. Bardram and T. R. Hansen. The AWARE architecture: Supporting context-mediated social awareness in mobile cooperation. In ACM Conference on Computer Supported Cooperative Work, Chicago, Illinois (USA), 2004. V. Bellotti and W. K. Edwards. Intelligibility and accountability: Human considerations in context aware systems. Human Computer Interaction, 16(2-4):193–212, 2001. 5 Personalization and Contextualization of Media 63
G. Bo. MOBILEARN: Project final report. http://www.mobilearn.org/download/results/ /pub- lic_deliverables/MOBIlearn_Final%20Report.zip, 2002. P. J. Brown. Triggering information by context. Personal Technologies, 2(1):1–9, 1998. P. Brusilovsky. Methods and techniques of adaptive hypermedia. User Modeling and User-Adapted Interaction, 6(2-3):87–129, 1996. P. Brusilovsky, M. Specht, and G. Weber. Towards adaptive learning environments. In Herausforderungen eines globalen Informationsverbundes für die Informatik (GISI), Zürich, Switzerland, 1995. J. R. Carbonell. AI in CAI: An artificial intelligence approach to computer-aided instruction. IEEE Transactions on Man-Machine Systems, 11(4):190–202, 1970. J. M. Carroll and C. Carrithers. Training wheels in a user interface. Communications of the ACM, 27(8): 800–806, 1984. W. J. Clancey. Knowledge based tutoring: The GUIDON program. MIT Press, 1987. A. Collins, J. S. Brown, and S. E. Newman. Knowing, Learning and Instruction, chapter Cognitive Appren- ticeship: Teaching the Craft of Reading, Writing, and Mathematics. Lawrence Erlbaum Associates, 1989. P. de Bra and L. Calvi. The New Review of Hypermedia and Multimedia, volume 1, chapter AHA! An open Adaptive Hypermedia Architecture. Taylor Graham Publishers, 1998. A. K. Dey and G. D. Abowd, editors. Towards a Better Understanding of Context and Context-Awareness, Karlsruhe, Germany, 1999. H. Dieterich, U. Malinowski, T. Kühme, and M. Schneider-Hufschmidt. Adaptive User Interfaces, chapter State of the Art in Adaptive User Interfaces. North-Holland, 1993. EQUATOR Project. The EQUATOR interdisciplinary research collaboration. http://www.equator.ac.uk/, 2003. S. Farnham, S. U. Kelly, W. Portnoy, and J. L. K. Schwartz. Wallop: designing social software for co-located social networks. In 37th Annual Hawaii International Conference on System Sciences, Big Island, Hawaii (USA), 2004. S. K. Feiner and K. R. McKeown. Intelligent Multimedia Interfaces, chapter Automating the Generation of Coordinated Multimedia Explanations. MIT Press, 1995. H. Frank. Lehrautomaten für die einzel- und gruppenschulung. In Lehrmaschinen in kybernetischer und pädagogischer, Sicht, Germany, 1965. T. Gross and M. Specht. Awareness in context-aware information systems. In Mensch und Computer, Bad Honnef, Germany, 2001. J. Herrington, R. Olivier, T. Herrington, and H. Sparrow. Towards a new tradition of online instruction: Using situated learning theory to design web-based units. In 17th Annual ASCILITE Conference, Auckland, New Zealand, 2002. A. Jameson. Modelling both the context and the user. Personal Technologies, 5(1):29–33, 2001. A. Jameson. Systems that adapt to their users: An integrative overview. In 9th International Conference on User Modelling, Johnstown, Philadelphia (USA), 2003. W. Kim. Personalization: Definition, status, and challenges ahead. Journal of Object Technology, 1(1):29–40, 2002. A. Kobsa. Generic user modeling systems. User Modeling and User-Adapted Interaction, 11(1):49–63, 1996. A. Kobsa and W. Wahlster. User Models in Dialog Systems. Springer Verlag, 1989. S. I. Konomi. QueryLens: Beyond ID-based information access. In 4th International Conference on Ubiquitous Computing, Göteborg, Sweden, 2002. A. Kröner. Adaptive Layout of Dynamic Web Pages. Akademische Verlagsgesellschaft Aka GmbH, 2001. D. Leutner. Adaptive Lehrsysteme; Instruktionspsychologische Grundlagen und experimentelle Analysen. Beltz, 1992. H. Mandl, H. Gruber, and A. Renkl. Information und Lernen mit Multimedia, chapter Situiertes Lernen in multimedialen Lernumgebungen. Psychologie Verlags Union, 1995. P. Mattila and T. Fordel. MOOP - using m-learning environment in primary schools. In 4th World Conference on mLearning, Cape Town, South Africa, 2005. L. Naismith, P. Lonsdale, G. Vavoula, and M. Sharples. Literature review in mobile technologies and learning. Technical Report NESTA Futurelab Series, report 11, University of Birmingham, 2004. 64 Marcus Specht, Andreas Zimmermann and Tim de Jong
H. Ogata and Y. Yano. Context-aware support for computer-supported ubiquitous learning. In 2nd IEEE International Workshop on Wireless and Mobile Technologies in Education, Taoyuan, Taiwan, 2004a. H. Ogata and Y. Yano. Knowledge awareness map for computer-supported ubiquitous language-learning. In 2nd IEEE International Workshop on Wireless and Mobile Technologies in Education, Taoyuan, Taiwan, 2004b. R. Oppermann and M. Specht. A context-sensitive nomadic exhibition guide. In 2nd International Symposium on Handheld and Ubiquituous Computing, Bristol, United Kingdom, 2000. R. Oppermann and M. Specht. Engaged Learning with Emerging Technologies, chapter Situated Learning in the Process of Work. Springer Verlag, 2006. G. Pask. Teaching machines and programming, chapter The adaptive teaching system. Pergamon Press, 1964. J. Rickel and W. L. Johnson. Integrating pedagogical capabilities in a virtual environment agent. In 1st Inter- national Conference on Autonomous Agents, Marina del Rey, California (USA), 1997. G. Salomon. Adaptiver Unterricht, chapter Heuristische Modelle für die Gewinnung von Interaktionshypothe- sen. Kösel, 1975. B. N. Schilit, N. I. Adams, and R. Want. Context-aware computing applications. In Workshop on Mobile Computing Systems and Applications, Santa Cruz, California (USA), 1994. M. Specht. Adaptive Methoden in computerbasierten Lehr/Lernsystemen. University of Trier, 1998. M. Specht and M. Kravcik. Authoring of learning objects in context. International Journal on E-Learning, 5 (1):25–33, 2006. M. Specht and R. Oppermann. The New Review of Hypermedia and Multimedia, volume 4, chapter ACE, Adaptive courseware environment. Taylor Graham Publishers, 1998. M. Stefaner, E. Dalla Vecchia, M. Condotta, M. Wolpers, M. Specht, M. Apelt, and E. Duval. MACE - enriching architectural learning objects for experience multiplication. Lecture Notes in Computer Science, 4753:322– 336, 2007. D. Stein. Situated learning in adult education. http://ericacve.org/docs/situated195.htm, 2003. A. Stone, G. Alsop, J. Briggs, and C. Tompsett. M-learning and e-learning: a review of work undertaken by the learning technology research group. In European Workshop on Mobile and Contextual Learning, London, United Kingdom, 2002. D. Tatar, D. Roschelle, P. Vahey, and W. R. Penuel. Handhelds go to school: Lessons learned. Computer, 36 (9):30–37, 2003. R. D. Tennyson and D. L. Christensen. Instructional Designs for microcomputer courseware, chapter MAIS: An intelligent learning System. Lawrence Erlbaum Associates, 1988. G. Weber and M. Specht. User modeling and adaptive navigation support in WWW-based tutoring systems. In 6th International Conference on User Modeling, Chia Laguna, Italy, 1997. E. Wenger and J. Lave. Situated Learning: Legitimate peripheral participation. Cambridge University Press, 1991. K. Yatani, M. Sugimoto, and F. Kusunoki. Musex: a system for supporting children’s collaborative learning in a museum with PDAs. In 2nd IEEE International Workshop on Wireless and Mobile Technologies in Education, Taoyuan, Taiwan, 2004. A. Zimmermann, A. Lorenz, and M. Specht. Personalization and context-management. User Modeling and User-Adapted Interaction, 15(3-4):275–302, 2005. A. Zimmermann, A. Lorenz, and R. Oppermann. An operational definition of context. In 6th International and Interdisciplinary Conference on Modeling and Using Context, Roskilde, Denmark, 2007. 6 Contextualized Recommender Systems: Data Model and Recommendation Process
Wolfgang Woerndl and Johann Schlichter
Technische Universität München Boltzmannstr. 3, 85748 Garching b. München, Germany {woerndl,schlicht}@in.tum.de
Summary. Recommender systems filter information items according to users’ interests and preferences. Sys- tems thereby utilize explicit ratings of items or implicitly observed user behaviour. In addition, it is important to tailor information access with regard to users’ current context, especially in mobile and ubiquitous application domains. Although the integration of context into recommender systems has not been investigated thoroughly so far, several approaches and applications do exist. We give a comprehensive overview about the state-of-the-art of contextualized recommender systems in this chapter. First we explain how context can be modeled, gener- ated and managed. Then, we discuss how to introduce context into the data model of recommender systems. We argue to consider context as an additional dimension to the traditional user-item ratings matrix of collaborative filtering. Context attributes can be utilized in different ways in various recommendation processes. We provide an overview of principal approaches that exist in the research literature. The discussion is illustrated by several application examples. We conclude with a summary and an outlook about future research prospects.
6.1 Introduction
The sheer volume of data that exists in our networked society today makes it difficult for users to find and access relevant information. Personalization of content is a technique to reduce the omnipresent information overload. Thereby, information is customized according to users’ needs and preferences. Often, recommender systems using collaborative or content-based filtering are applied. In a mobile or ubiquitous scenario, informa- tion personalization is even more important, because of the limitations of mobile devices regarding displays, input capabilities, bandwidth etc. For example, users cannot browse through many Web search results on the small screen of a mobile device in contrast to working on a desktop PC. It would be more desirable to pre-filter results to adapt to the user’s current needs, if possible. Another promising and possibly complementary approach is to utilize context. For instance, a user travel- ling with a PDA or a smartphone needs access to current train schedules, the weather report at her destination or a recommendation for a restaurant in her vicinity, that is currently open and caters to the users prefer- ences [Schwinger et al., 2005]. Hence while on the one hand personalized and context-adapted information access is important in mobile settings or scenarios with non-standard user interfaces, on the other hand obtain- ing relevant information is even more difficult. A lot of work has been done with regard to context-aware applications and personalization research in general. For example, Baldauf et al. [2007] present a recent overview on context-aware systems, however without mentioning context-aware recommender systems. Adomavicius and Tuzhilin [2005] outline the state- of-the-art of recommender systems and also dedicate one chapter to context-awareness where they mainly explain their multi-dimensional recommender approach [Adomavicius et al., 2005]. Other papers consider some kind of context attributes as part of a user and/or item model, but these systems cannot easily be applied to other application areas and it is difficult to derive general principles for contextualized recommender systems. The integration of context into recommender systems has not been investigated thoroughly up to now [Anand and 66 Wolfgang Woerndl and Johann Schlichter
Mobasher, 2005], but several approaches and applications do exist. The goal of this chapter is to present an overview over concepts and principles of contextualized recommender systems. Firstly, we explain how context can be modelled and acquired (Section 6.2). Then, we discuss in Section 6.3 how to introduce context into the data model of recommender systems. We argue to consider context as an additional dimension to the traditional user-item ratings matrix of collaborative filtering. Context attributes can be utilized in different ways in various recommendation processes including collaborative filtering, context- based filtering and hybrid recommenders (Section 6.4). We provide an overview of some principal approaches we have developed in our group [Woerndl and Schlichter, 2007] and also ideas available in research literature. The discussion is illustrated by application examples where possible. We finally conclude with a brief summary in Section 6.5.
6.2 Context: Definition, Model and Acquisition
In a recent survey on web personalization, Anand and Mobasher [2005] list context-awareness as an issue that has to be addressed appropriately if personalization systems are to find wider acceptance. They state that “context as a concept has rarely been incorporated into personalization research”. Furthermore, they identify that “one of the reasons for this is that it is hard to arrive at a consensus of what defines context let alone modelling the concept”. Therefore, we will first define what we mean by context, and also present how (high level) context can be modelled and acquired for example in mobile scenarios we are working on.
6.2.1 Context Definition and Modelling
In our work we follow the context definition from Dey et al. [2001]: “Context is any information that can be used to characterize the situation of entities (i.e. whether a person, place or subject) that are considered relevant to the interaction between a user and an application, including the user and the application themselves”. That means, context is very dynamic and transient. For example in a mobile application domain, a context model could include location, movement, lighting condition or current availability of network bandwidth of mobile devices. As far as context modelling is concerned, several approaches for representing context information have been developed [Strang and Linnhoff-Popien, 2004, Baldauf et al., 2007]. Key-value pairs can be used as a simple data structure. Markup scheme models consist of hierarchically organized markup tags with attributes and content. An example is W3C’s standard Composite Capabilities/Preference Profile (CC/PP) [Klyne et al., 2004] for representing device capabilities. Object oriented or graphical models, e.g. using the Unified Mod- elling Language (UML), provide more powerful means to represent context types and describing attributes. In addition, ontologies are a very promising instrument for modelling contextual information due to their high and formal expressiveness and the possibilities for applying ontology reasoning techniques. Ontology-based models can thereby represent complex relationships between (context) concepts.
6.2.2 Acquisition of High Level Context from Sensor Data
Most of the context is available as low level sensor data that is unsuitable to be utilized in recommender or personalization systems. These applications need higher level context. For example, instead of exploiting GPS coordinates directly, it is more desirable to be able to use names of places or nearby landmarks. Therefore, we explain below how to acquire high level context from sensor data.
Process Model
Most approaches in the literature are based on a similar process for the generation of high level context, even if they specialize in different ways. For example, [Himberg et al., 2003] show that data mining techniques are 6 Contextualized Recommender Systems: Data Model and Recommendation Process 67 suitable for identifying high level context, and [Korpipää et al., 2003] use Naïve Bayes networks to classify context. The following process model consists of the following four tasks and is similar to what is used in other approaches [Woerndl et al., 2007a]:
1. Acquisition of raw data. 2. Preprocessing: aggregation and feature extraction. 3. Interpreting the data. 4. Utilization in applications.
Note that this is not a one-time process but an iterative activity to generate context. The first step is to acquire raw data from sensors, e.g. for location, acceleration, temperature, sound/voice, etc. For each sensor, a different software component has to be developed. Thereby, it is important to consider measurement ranges and error rates to get useful results. Thus, the acquisition of raw data allows for a perception of the physical environment and establishes the foundation for the other tasks. The second step is preprocessing of the (raw) data. This step varies significantly in the existing literature, but some kind of preprocessing is done in about every approach. One reason is that the sheer amount of data usually is huge after acquiring raw data, and it cannot be handled directly, especially in our domain of mobile devices. Therefore, aggregation and feature extraction has to be applied. One example is to use sliding windows to extract features from sensor data streams. Thereby, an algorithm examines only a subset of the whole data set at one time, and calculates characteristics such as mean values or standard deviation. The extraction window is then moved along the data stream. The size and overlap of the windows have an influence on the necessary memory and processing requirements, which is especially an issue for mobile devices. The third step in the process model is interpretation which realizes the transition from lower-level data to context information that is useful in applications. The interpretation may take several steps, dependent on the used methods, e.g. supervised or unsupervised learning algorithms. After interpreting the data, the fourth step is to utilize the generated high level context in applications, or to visualize results for users.
Implementation Example
We have designed and implemented software components that realize the model explained above [Woerndl et al., 2007a]. The architecture consists of a client-side part, running on a PDA or smartphone. After acquisition, the data is transferred to a server-based application for the preprocessing and interpretation. Users can control the context acquisition via an interface on the client. The client also allows for entering annotations for a supervised training algorithm. We have tested the feasibility of our approach by experimenting with several sensors, including an acceleration sensor to infer the activity a user is doing, for example “running” or “sitting”. Figure 6.1 shows the visualization of the acceleration sensor information using real world data. The aggre- gated values of the three axis of the acceleration sensor are plotted at the top, and the inferred and actual user activity are displayed at the bottom of the screen. We have not done a formal study regarding the accuracy of the results yet, but our tests indicate that the algorithm is performing reasonably well and can be used to develop an application which monitors, stores and analyzes user activities. Some problems emerged when activities are somewhat related. For example, it is difficult to distinguish between “sitting” and “sitting in a car”. Therefore at timestamp 950, the algorithm first inferred that the user is driving, even though that was not the case.
6.3 Context in Recommender Systems Data Model
After explaining what we mean by context, and how (high level) context can be acquired, we now focus on the integration of context into recommender systems. The discussion is split into first the data model of recom- mender systems (Section 6.3) and then the various recommender system processes (Section 6.4). 68 Wolfgang Woerndl and Johann Schlichter
Fig. 6.1. Inferring user activity from an acceleration sensor.
6.3.1 Traditional Data Model
A recommender system tries to predict the relevance of information items for a user. This is traditionally based on information about the user, meta data associated with items and/or implicit or explicit ratings for items made by a group of users. The standard data model can be summarized as follows [Anand and Mobasher, 2005]. Given is a set of n items I = ij : 1 j n, and a set of m users U = uk : 1 k m. Each user is described by a t-dimensional vector of attrib≤ utes,≤ which is also called (user) profile.≤So the≤ profile contains attribute-value pairs such as “language = EN”. The items have associated meta data, in form of a s-dimensional vector (item description). For example, an item may be a particular restaurant (in a restaurant guide application), and the item description may contain “location = (longitude, latitude)”, and also the price range and opening hours of a restaurant to be utilized in a tourist guide. Thus items and users are technically modelled in the same way with respect to data structures. In addition, often users’ ratings of items are used. Technically, the rating is a function r : U I S that maps a user instance (and therefore the values of the user profile attribute values) and an item×instance−→ (or the values of the corresponding item meta data attribute values) to some appropriate mapping result-space S, e.g. a rating value in [0, 1]. In other words, a rating ri,s expresses how much a user uj likes an item is. Ratings can be obtained using either implicit or explicit user profiling methods. An example for explicit ratings would be filling out a Web form by a user to express which type of restaurants she liked. Implicit ratings can be acquired by observing the user, for example the consideration of eating at a particular place results in a positive rating, without any additional user input. The goal of the personalization process is to recommend items ij to an active user ua that would be of interest to the user. How this is done varies among the different recommendation techniques and is covered in Section 6.4 of this chapter. Since recommenders usually operate on the user-item matrix with ratings as ele- ments, one major problem is the sparseness in this matrix, e.g. there are often large numbers of users and items but only few ratings that can be exploited. Another drawback is the cold start problem because recommenda- tions for new users or items are hard to generate, depending on the used technique. 6 Contextualized Recommender Systems: Data Model and Recommendation Process 69
6.3.2 Integrating Context
The question discussed in this section is how to incorporate context information into the data model of recom- mender systems.
Context Versus User Profile
A first idea would be to incorporate context information as part of the user profile. However, our assumption is that it is important to distinguish between user profile and context attributes and treat them as separate entities in the recommender data model. While the user profile is rather static and somewhat longer lasting, context is highly dynamic and transient. Depending on the application area, user models can also be highly dynamic, but from our point of view a user profile contains information such as preferences, interests or knowledge (e.g. the user’ss favourite type of restaurant), whereas the context model describes the current environment in which the user operates in (e.g. her current location). Profile information can be implicitly observed or explicitly provided by the user. The profile is stored permanently on a profile/identity server and can be manually modified by the user if needed, for example if user preferences or interests change. On the other hand, context information is neither stored permanently nor manually entered by the user. In our model, the context is captured and analyzed when a user wants to generate a recommendation or implicitly or explicitly submits a rating for an item. In the latter case, the captured context model is stored together with the rating to be used later. However, it is possible to transfer (inferred, high level) context information to the profile. For instance in the acceleration sensor example from Section 6.2.2, the current activity information such as “sitting in a car” at one particular moment is context, while derived information such as “the user drives from A to B every working day” could be part of a user profile.
User-Item-Context Matrix
As far as the context model is concerned (Section 6.2.1), most available contextualized recommender systems only use a rather simple model so far, such as attribute value pairs. More formally, the context C can be described as a vector of context attributes, analogous to the user profile and the item meta data [Chen, 2005]. The similarity or identity of context values is not as clear as with users or items and it may depend on the actual application domain. For example, for a tourist guide the same context might be constituted by a range of GPS coordinates. The integration of context extends the domain of the rating function to U I C. In other words, context adds another dimension to the item-user matrix (Figure 6.2), as introduced by×Adoma× vicius et al. [2005]. This is also of importance with regard to the sparseness of ratings, because R may only be defined for a small part of possible contexts, e.g. ratings may be valid in one particular context only. Which context attributes are actually modelled and used in systems is largely dependent on the requirements of the application domains.
Spreading Activation Framework
As an alternative to the explained user-item-context matrix, Kovács and Ueno [2006] present CASAN nets, which are an extension of the classical associative spreading activation network formalism. In CASAN nets, content nodes are not just connected by plain directed weighted links but also have associated link type and context nodes. Their spreading activation algorithm incorporates mechanisms to handle these two new functions of nodes. However, it is not clear how these CASAN nets can really be exploited in a practical recommendation process.
6.4 Context in Recommender Systems Process
After explaining alternatives for integrating context into the data model of recommender systems, we will now cover how recommender system processes can be improved by utilizing context. The discussion is structured 70 Wolfgang Woerndl and Johann Schlichter
Fig. 6.2. The user-item-context matrix. according to the most important types of recommenders: collaborative filtering (Section 6.4.1), context-based recommender (Section 6.4.2) and hybrid approaches (Section 6.4.3).
6.4.1 Collaborative Filtering
In comparison to non-collaborative, individual recommender techniques, traditional collaborative filtering (CF) does take other users into account and tries to automate the process of word of mouth. This is usually done by recommending items to an active user that have been rated highly by users who are similar to the active user. Two basic steps are needed to generate recommendations: neighbourhood creation and ratings prediction. Neighbourhood creation means selecting users that are similar to the active user. In most approaches, sim- ilar users are users that have rated items in a similar way as the active user in the past. This user-user similarity usually is calculated by using Pearson correlation, cosine-based approach or other likewise metrics [Adomavi- cius and Tuzhilin, 2005]. Subsequently, a set S of n users who are most similar to the active user is computed. For predicting the ratings of some items for the active user, the mean average of ratings given by the users in S for an item is determined. This average can be weighted according to the actual user similarity of a user to the active user and also adjusted if some users are always giving higher-than-average (or lower) ratings. Finally, a set of items with the highest predicted ratings is presented to the active user as recommendations.
Contextualized Neighbourhood Creation and Ratings Prediction
When introducing context into CF, the main idea is to weigh ratings in both algorithms according to context similarity. For the user-user metrics the deviation of ratings of users in comparison to the ratings of the active users is computed. Thereby no additional information has an influence while considering the context of ratings appears to be an important factor. For example, if one user A rated restaurants when having lunch, and another user B rated the restaurant when dining in the evening, the current time noon or evening when the active user asks for a recommendation may play a significant role when determining the neighbourhood of similar users. Context can be incorporated by weighing the ratings according to context similarity with the current context. Therefore a contextualized CF algorithm should store a snapshot of the current context with every rating made. As for the ratings prediction, the option is also to apply a higher weight to ratings that were made in a similar context in comparison to the current context of the active user, analog to applying weights to user similarity. For instance, we assume A, B as the set of users who are similar to the active user. A user A rated an item { } as 0.2 on a scale from 0 to 1 in a context c1, the other user gave a rating 0.8 in context c2. A contextualized 6 Contextualized Recommender Systems: Data Model and Recommendation Process 71
CF algorithm would compare the current context of the active user to c1 and c2 and then compute the predicted rating according to the context similarity, e.g. 0.7 if the current context of the actice user is similar to c2. Chen presents a context-aware collaborative filtering system based on past experiences of users [Chen, 2005]. The scenario is ubiquitous computing to assist users in different situations. To incorporate context, the approach weighs the current context of the active user against the context of each rating with respect to their relevance to locate ratings given in a similar context. Thereby, Chen uses a modified Pearson correlation coefficient to measure the correlation between context variables such as location, temperature and time [Chen, 2005]. One major problem of this approach is the availability of ratings in comparable contexts. Collaborative filtering needs a sufficient number of ratings to produce satisfying results. This sparseness problem is an issue with CF in general, but this problem is aggravated when integrating context information. In addition, contextualized CF needs a context similarity function as we outlined above, which may be difficult to formalize. One possible solution is to move from simple context models to context ontologies (Section 6.2.1). Thereby, the relationship between objects in the context model can be utilized to improve recommender sys- tems. For example in a mobile recommender: if there are not enough ratings for a particular location, e.g. a street, the ontological information that “street” is part of “town” could be used to increase the available of relevant ratings in a certain context. Another general solution to cope with the disadvantages including the new item and new user problems of CF is to combine CF with other recommendation techniques to form a hybrid recommender. Contextualized hybrid recommenders are covered in Section 6.4.3. The approach by Herlocker and Konstan [2001] was one of the first examples of a contextualized recom- mender system. They argue that adding content-independent context can improve collaborative filtering. The context is knowledge about a user’s task which consists of a list of example items. Therefore, this approach is still two-dimensional, i.e. it uses only user and item dimensions, and does not explicitly model any additional contextual information.
Social Context
Context does not always have to be physical context such as location and environment conditions. Equally important is the social context [Woerndl and Groh, 2007], e.g. the social network of a user, buddy lists, past interactions with other users etc. It seems obvious that social context is important for collaborative filtering, because it does resemble to the word of mouth principle of CF. For example in social groups, group members and opinion leaders have great impact on the advice seeking and advice evaluation process of other group members in domains related to users’ tastes where personal opinions figure to play a major role. To use social models in recommender system, Groh and Ehmig [2007] replace the neighbourhood creation based on ratings similarity in collaborative filtering by a social neighbourhood of users, e.g. all friends or all friends plus friends of friends. They demonstrate that their social recommender outperforms traditional CF in their scenario of recommending clubs. In this approach, ratings are not used at all for the neighbourhood creation. Another option is to weigh the ratings with a social context similarity score which should be computed at the time a rating is made, to account for changing social environments of users.
6.4.2 Content-Based Recommender Systems
In content-based recommender systems the approach is to generate recommendations based on the analysis of items previously rated by the active user and generating a user profile based on the meta data of these items. The profile is then used to predict a rating for previously unseen items [Anand and Mobasher, 2005]. In other words, content-based techniques are individual and do not take other users into consideration. Only one row of the user-item matrix is used, the ratings of the active user. Matching between items can be done in different ways, for example by performing text analysis to find similar items or by identifying and applying rules. In general, we need a function F that maps (U I) I, i.e. the user-item matrix with a user’s past ratings and a particular item, to a result space S, e.g. [0, 1].× For×example, if user A liked Chinese restaurants in the past, F would produce a high value for new restaurants that have “Chinese” in their item description. Moving to a contextual recommender, this function has to consider the context, i.e. F : (U I) I C S. For × × × −→ 72 Wolfgang Woerndl and Johann Schlichter example, if user A previously liked Chinese restaurants only for dinner, recommendations may change when she is looking for a quick lunch. There are different approaches to actually realize a contextualized content-based recommender. The most prominent example for a contextual content-based recommender system is the multi-dimensional approach by Adomavicius et al. (see below). Other options include rule-based systems (see below) or statistics-based approaches. An example for the latter is the mobile opportunistic planning study by Horvitz et al. [2007].
Multi-Dimensional Recommender Systems
Adomavicius et al. [2005] enhance the user-item data model to a multidimensional setting where the additional dimensions constitute different context attributes such as location and time. In general, the recommendation process can then be generalized to a multidimensional utility function u : D1 D2 . . . Dn R. Then, the recommendation problem is defined by selecting certain “what” dimensions× and× certain× “for −whom→ ” dimensions, and recommending the “what” tuples for each “for whom” tuple that maximizes the utility function u [Adomavicius and Tuzhilin, 2005]. Adomavicius et al. also propose a reduction-based approach with the goal to reduce the dimensions of the utility function. For example, if a person wants to see a movie on a Saturday night, the approach would only consider ratings of movies seen on weekends and thereby eliminate the “time of the week” dimension. Thereby, the n-dimensional cube of ratings is projected to the user-item matrix and any standard algorithm including CF can be applied [Adomavicius and Tuzhilin, 2005].
Point-of-Interest Trigger
The PoiAppRecommender is an example for a content-based context-aware recommender using rules or trig- gers that we have designed and implemented. This recommender is integrated in a play.tools framework which supports the development of innovative, mobile applications for devices such as smartphones [Woerndl et al., 2007b]. The realized acquisition of high level context from sensors is also integrated in this framework and can be used for personalized applications. Part of the framework is a deployment server where developers of mobile applications can register their services and users can browse and search for relevant and interesting gadgets. Users can access the deployment server and download client modules on their mobile devices. One problem for users is to find interesting and with regard to their current context relevant applications. The PoiAppRecommender does not recommend points-of-interest (POIs) but recommends mobile appli- cations based on POIs in the vicinity of the user using triggers. A software developer or administrator can select among types of points-of-interest (such as restaurant, museum or train station) and specify within which radius of an actual POI an application is recommended. This is done when registering the application with the mentioned deployment server. When making a recommendation, the system then retrieves the current user position (using a GPS-enabled mobile device), determines POIs in the vicinity and generates a recommendation based on this context infor- mation. For example, an administrator can specify that her mobile train table application shall be recommended when the user is near a train station. After applying the trigger rules, our approach uses collaborative filtering to rank found items according to user ratings of applications in a second step. User ratings are collected implicitly by automatically recording when a user installs an application within our framework on her mobile device. It is also optionally possible for users to explicitly rate applications after usage. The ratings are stored together with context information (time, location, used device, ...) to capture the situation when a rating was made as motivated above.
6.4.3 Hybrid Recommender
Hybrid recommender systems combine different approaches with the goal of utilizing the strengths of one al- gorithm while avoiding its weaknesses by applying a second approach [Burke, 2002]. Combination strategies are: weighted (the scores of several recommendation techniques are combined), cascading (one recommender 6 Contextualized Recommender Systems: Data Model and Recommendation Process 73 refines the recommendations given by another), switching (switching between appropriate recommendation techniques), mixed (system presents results some several recommenders at the same time), feature combina- tion (features from different recommendation data sources are thrown together into a single recommendation algorithm) or feature augmentation (the output from one technique is used as an input feature to another).
Contextualized Hybrid Recommender Systems
When integrating context, one option is to combine several algorithms to one hybrid context-aware recom- mender system. In principle, all of the above combination strategies could be used, but it is particularly promis- ing to use a hybrid recommender to reduce the mentioned complexity of the user-item-context matrix (Sec- tion 6.3.2). Thereby, a first algorithm RS 1 would operate on two dimensions and compute an intermediary result set of items by taking only user and context attributes into account. In a second step, RS 2 further filters and ranks the results by considering the third dimension, for example by utilizing item meta data. Thus, the used combination strategy is cascading or feature augmentation. Figure 6.2 illustrates this principle.
Fig. 6.3. Hybrid contextual recommender.
Recommending Mobile Applications
This principle of combining several recommenders into one context-sensitive system has been realized in the already mentioned play.tools framework for the development of mobile applications (Section 6.4.2) [Woerndl et al., 2007b]. First of all, a switching strategy is used to select among different available recommender modules. Users can thereby choose between the following content-based and collaborative filtering components:
LocationAppRecommender. • CFAppRecommender. • PoiAppRecommender (see Section 6.4.2). • The recommender components itself are also context-sensitive and hybrid. The LocationAppRecommender selects mobile applications that were used in a similar context (most importantly, a nearby location) by other users in a first step. In a second step, the intermediary results are ranked according to ratings (positive/negative), i.e. taking other users into account. Thus, this recommender is an implementation of a cascading hybrid rec- ommender. The first step of the LocationAppRecommender can be seen as a contextualized collaborative rec- ommender where the similarity of users is calculated according to context similarity only, and not according to previously made user ratings. Therefore, this recommender is in principle related to the social recommender explained in Section 6.4.1 [Woerndl and Groh, 2007]. The CFAppRecommender applies collaborative filtering in our domain of mobile applications using the Taste library [Lemire and Maclachlan, 2005]. Taste provides a set of components from which one can construct a customized recommender system. The CFAppRecommender analyzes the item-user data without considering 74 Wolfgang Woerndl and Johann Schlichter context in the first step and filters the intermediary results, i.e. recommended mobile applications, according to context attributes such as devices capabilities (display size and input capabilities) and currently available network bandwidth in the second step.
6.5 Conclusion
In this chapter, we have identified some general ideas for utilizing context to improve recommender systems. First, we have discussed how to integrate context in the data model of recommender systems. In the main part of the chapter, we have explained contextualized recommender processes, grouped according to the different recommendation techniques that are mainly available today. We have also illustrated the ideas by outlining context-aware recommender applications including a hybrid recommender system which recommends mobile applications to users derived from what other users have installed in a similar physical context (location, cur- rently used type of device, etc). When evaluating contextual recommender systems, a problem is the availability of relevant data sets. Ex- isting data sets (e.g. from the GroupLens project1) are not associated with contextual information. Furthermore, since context is very dynamic it is hard to judge when a predicted recommendation is actually good a bad. A rating given for an item in a certain context may not be valid in another context. Therefore, it seems that an eval- uation of context-aware recommender systems should not only involve computed accuracy metrics [Herlocker et al., 2004], but also some kind of user studies. Contextual recommender systems seem to be especially promising in very dynamic environments. This includes mobile and ubiquitous scenarios as motivated in the introduction. Thereby, user interfaces are crucial and developing innovative user interfaces for contextualized personalization applications is a major future re- search issues. This eminently applies for non-standard application domains. One example is ad-hoc networking of cars, where information like rating of points-of-interest in the vicinity can be exchanged between vehicles and utilized for collaborative, context-aware applications [Woerndl and Eigner, 2007].
References
G. Adomavicius and A. Tuzhilin. Towards the next generation of recommender systems: A survey of the state- of-the-art and possible extensions. IEEE Transactions on Knowledge and Data Engineering, 17(6), June 2005. G. Adomavicius, R. Sankaranarayanan, S. Sen, and A. Tuzhilin. Incorporating contextual information in rec- ommender system using a multidimensional approach. ACM Transactions on Information Systems, 23(1): 103–145, 2005. S. S. Anand and B. Mobasher. Intelligent techniques for web personalization. Lecture Notes in Artificial Intelligence, 3169:1–36, 2005. M. Baldauf, S. Dushdar, and F. Rosenberg. Survey on context-aware systems. International Journal of Ad Hoc and Ubiquitous Computing, 2(4):263–277, 2007. R. Burke. Hybrid recommender systems: Survey and experiments. User Modeling and User-Adapted Interac- tion, 12(4):331–370, 2002. A. Chen. Context-aware collaborative filtering system. In 1st International Workshop on Location- and Context- Awareness, pages 244–253, Oberpfaffenhofen, Germany, 2005. A. K. Dey, G. D. Abowd, and D. Salber. Conceptual framework and a toolkit for supporting the rapid prototyp- ing of context-aware-applications. Human-Computer Interaction, 16(2-4):97–166, 2001. G. Groh and C. Ehmig. Improving recommender systems via social networks. In 1st International Conference on Supporting Group Work, Sanibel Island, Florida (USA), 2007. J. L. Herlocker and J. A. Konstan. Content-independent task-focused recommendation. IEEE Internet Com- puting, 5(6):40–47, 2001.
1 http://www.grouplens.org/ 6 Contextualized Recommender Systems: Data Model and Recommendation Process 75
J. L. Herlocker, J. A. Konstan, L. G. Terveen, and J. T. Riedl. Evaluating collaborative filtering recommender systems. ACM Transactions on Information Systems, 22(1):5–53, 2004. J. Himberg, J. A. Flanagan, and J. Mäntyjärvi. Towards context awareness using symbol clustering map. In 4th International Workshop on Self-Organizing Maps, pages 249–254, Kitakyushu, Japan, 2003. E. Horvitz, P. Koch, and M. Subramani. Mobile opportunistic planning: Methods and models. Lecture Notes in Artificial Intelligence, 4511:228–237, 2007. G. Klyne, F. Reynolds, C. Woodrow, H. Ohto, J. Hjelm, M. H. Butler, and L. Tran. Composite Capa- bility/Preference Profiles (CC/PP): Structure and Vocabularies 1.0, W3C Recommendation, 2004. URL http://www.w3.org/TR/CCPP-struct-vocab/. P. Korpipää, M. Koskinen, J. Peltola, S. M. Mäkelä, and T. Seppänen. Bayesian approach to sensor-based context awareness. Personal and Ubiquitous Computing, 7(2):113–124, 2003. A. I. Kovács and H. Ueno. Recommending in context: A spreading activation model that is independent of the type of recommender system and its contents. In 2nd International Workshop on Web Personalisation, Recommender Systems and Intelligent User Interfaces, Dublin, Ireland, 2006. D. Lemire and A. Maclachlan. Slope one predictors for online rating-based collaborative filtering. In 5th SIAM International Conference on Data Mining, Newport Beach, California (USA), 2005. SIAM. W. Schwinger, C. Grün, B. Pröll, W. Retschitzegger, and A. Schauerhuber. Context-awareness in mobile tourism guides - a comprehensive survey. Technical report, Technical University of Wien, 2005. T. Strang and C. Linnhoff-Popien. Context modeling survey. In 1st International Workshop on Advanced Context Modelling, Reasoning and Management, pages 34–41, Nottingham, United Kingdom, 2004. W. Woerndl and R. Eigner. Collaborative, context-aware applications for inter-networked cars. In 5th Interna- tional Workshop on Distributed and Mobile Collaboration, Paris, France, 2007. W. Woerndl and G. Groh. Utilizing physical and social context to improve recommender systems. In Workshop on Web Personalization and Recommender Systems, in conjunction with the IEEE/WIC/ACM International Conference on Web Intelligence, Silicon Valley, California (USA), 2007. W. Woerndl and J. Schlichter. Introducing context into recommender systems. In Workshop on Recommender Systems in e-Commerce, in conjunction with the 22nd Conference on Artificial Intelligence, Vancouver, Canada, 2007. W. Woerndl, C. Schueller, and T. Rottach. Generating high level context from sensor data for mobile appli- cations. In IADIS International Conference on Wireless Applications and Computing, Lisbon, Portugal, 2007a. W. Woerndl, C. Schueller, and R. Wojtech. Hybrid recommender system for context-aware recommendations of mobile applications. In 3rd IEEE International Workshop on Web Personalisation, Recommender Systems and Intelligent User Interfaces, Istanbul, Turkey, 2007b.
7 Enhanced Media Descriptions for the Automatic Adaptation of Audiovisual Content Retrieval
Iván Cantador, Fernando López, Jesús Bescós, Pablo Castells, and José M. Martínez
Escuela Politécnica Superior, Universidad Autónoma de Madrid (Spain) {ivan.cantador,f.lopezDELIT,J.Bescos,pablo.castells,JoseM.Martinez}@uam.es
Summary. The continuous growth of the audiovisual content body worldwide poses new challenges to content retrieval and delivery technologies, calling for new solutions that cope with the current scale and complexity of content corpora, technological infrastructures, and user audiences. This chapter focuses on a set of initiatives and achievements addressing problems that involve the effective selection of content items and their automatic adjustment to fit a range of support infrastructures. Starting out from a proper description of multimedia content, the adaptive effects discussed here comprise low to high-level adaptation methods, from media transcoding and transmoding, scalable bit-stream modification, or video skimming and summarization, to the ranking of content units according to background user interests in different scenarios (e.g. presence vs. absence of an explicit user query, single vs. multiple users, etc).
7.1 Introduction
The continuous growth of audiovisual content in digital format available worldwide poses new challenges to content retrieval technologies, calling for new solutions that cope with the current scale and complexity of content corpora, technological infrastructures, and user audiences. Two major problem areas to be addressed in facing such challenges involve (i) the effective selection of content items when the scale of the retrieval space surpasses the capacity of users to query the corpus, or even browse search result lists; and (ii) the automatic adjustment of multimedia content to fit a wide variety of support infrastructures (terminals, networks, codecs, players, etc), while making the most of the available delivery channel capabilities. Addressing such problems implies work at the level of the identification, representation, dynamic analysis, and effective introduction of the contextual conditions that are relevant for the content retrieval and delivery processes, in order to best suit the situation at hand, in a way that optimizes the effectiveness in terms of user satisfaction. This involves building up a system awareness of dynamic conditions such as user interests and preferences of different kinds (high-level and low-level, broad and specific, explicit and implicit, content- related, source-related, etc), device and network capabilities (screen resolution, network bandwidth, etc). A proper description of multimedia content itself is needed, ranging from signal-level (colour, bitrate, etc) to syn- tactic (objects, motion, visual attention, etc) and semantic-level descriptions (topics, domain concepts, events, semantic relations). This chapter focuses on a set of initiatives and achievements addressing such problems, resulting in different forms of personalized retrieval and dynamic adaptation (previous or simultaneously to delivery, not covered here) of multimedia content. The comprehensive view on multimedia adaptation provided here comprises low to high-level adaptation methods from the ranking of content units according to background user interests in different scenarios (e.g. presence vs. absence of an explicit user query, single vs. multiple users, etc) to media adaptation techniques to different usage scenarios (terminals, networks, user preferences, etc). The chapter is structured as follows. The next section addresses the different aspects of content that need to be explicitly modelled in order to enable the adaptation techniques discussed thereafter. Section 7.3 focuses 78 Iván Cantador, Fernando López, Jesús Bescós, Pablo Castells, and José M. Martínez on the personalization of content retrieval, the different sides and problems involved, and describes specific proposals in a semantic-based approach. After this, Section 7.4 the adaptation of content at signal-level, that is, customizing it in a content-agnostic way to the different terminals, networks and user preferences (from the media presentation point of view). Finally, some concluding remarks are provided in Section 7.5.
7.2 Content Modelling
Content access and delivery involve different processes or phases, such as query, selection (filtering), linking (recommendation), semantic adaptation, and signal-level adaptation (these last two steps could be combined in one), each involving different technologies. The automatic customization problem can thus be addressed at the different phases, motivating diverse requirements and approaches for each step of the chain. A common key aspect to all of them is the need for appropriate descriptions of the content to which the adaptation strategies apply, at the proper level and providing the relevant details needed to handle the content pieces, as required by the adaptive operations. Considering the specifics of multimedia content, the following aspects need to be taken into account to this respect:
Representation model: The syntax and semantics associated to the bitstream that represents the signal • used to deliver and present the content. Description model: The syntax and semantics associated to what is present in the content from a signal- • level point of view (e.g. colour, shape, texture, motion for visual content) and from a mid-level point of view (generic objects and events), as well as the descriptive information required from an archival point of view. High-level semantics model: The syntax and semantics associated to the high-level interpretation and • meaning of what is present in the content, generally linked to a particular application domain.
The requirements and available technologies related to each of the above description levels are discussed in the next sections.
7.2.1 Content Representation
There are different specifications for the structuring of the bitstream used to represent media content. Although proprietary formats were in use in the past, nowadays most applications work with standardized formats defin- ing both the syntax and the semantics of the bitstream, from a purely signal point of view, and mainly describing a compressed representation of the media, with a good subjective quality even in the presence of heavy com- pression. Most of the wide-spread standards for audiovisual content belong to the JPEG family for images (the well- known JPEG format available in all digital cameras, and the new JPEG2000 providing much better quality for the same bitrate), and to the MPEG family for video and audio: from MPEG-2 used in DVDs and Digital Television, and MP3 for audio on the Internet, to heavily improved versions in the form of MPEG-4 Advanced Audio Coding (AAC) and MPEG-4 Advanced Video Coding (AVC) also known as H.264.
7.2.2 Content Description
Besides the content representation of the signal, it is required to have a description of the content, that is, metadata (data about the content essence). These descriptions range from the classical archival data (title, author, release date, abstract, keywords, etc) to content-based descriptions of what is present in the content (e.g., a goal in a soccer video, nudity in a film, an interview by the anchorman in a news program). There are several standards for archival metadata (e.g., EXIF for images, ID3 for audio files, Dublin Core for generic digital content), whilst only a small number of standards, mostly broadcasters-oriented, deal to some extent with content-based metadata (e.g., SMTPE Metadata Dictionary, EBU P/Meta, TV-Anytime). 7 Enhanced Media Descriptions for Audiovisual Content Retrieval 79
The MPEG-7 standard [Manjunath et al., 2003] is the most comprehensive multimedia content description specification, covering both archival metadata as well a large number of metadata for content-based description. MPEG-7 provides description tools for low-level audiovisual descriptions (e.g. colour, shape, audio spectrum), mid-level (e.g. face description, musical timbre), structural level (e.g. shot and scene decomposition, musical movements), and even tools for semantic and linguistic description. The main drawback of the MPEG-7 specifi- cations is their complexity, mainly due to the fact that MPEG-7 was aimed to be a generic standard, not focused on a particular application domain. In order to help reducing this complexity, and following previous MPEG standards, profiles (subset of tools) have been defined for different applications domains.
7.2.3 Usage Context Description
By usage context we mean the terminal capabilities, network conditions and media related user preferences (e.g. preference for an image slide rather than a video, or a travelling video instead of an image that does not fit in the screen) that are active in each media consumption session. All this information is required in order to make decisions about what kind of adaptation should be performed over a media for a specific session. There are several standards dealing with some of these descriptions (e.g., UAProf, CC/PP), but the Digital Item Adaptation (DIA) part of MPEG-21 [Burnett et al., 2006, Vetro and Timmerer, 2005] is the one covering the more complete set of such descriptions, as well as additional tools for media adaptation techniques.
7.2.4 High-level Semantics
The descriptions discussed above convey a lot of useful information for handling and reformatting content at the signal-level, as described later in Section 7.4. However, from the point of view of cognitive user needs, higher-level information is required. At the end of the consumption chain, users of multimedia content are concerned with the information they are going to get by viewing a specific document, which is a major source of relevant input for personalizing the choice of documents to be delivered, or fragments to be viewed within. High-level semantics aim to describe what objects appear in a scene (people, cars, buildings, trees, roads), what type of scene is displayed (tennis, beach, restaurant, indoors/outdoors), what happens in it (a person enters the room, walks, sits down, smiles, eats), and so forth. The type of entities, events, and subjects that may appear in a multimedia document are virtually anything, which makes it hard to provide a general framework supporting the needed descriptions in a general way. While it is indeed impossible to model the world as a whole (although some attempts have been made [Lenat and Guha, 1990]), a partial and feasible approach is possible for restricted domains, as supported by ontology- based knowledge technologies [Staab and Studer, 2004]. Ontology-driven representations, and in particular the ones supported by W3C standards such as RDFS [Brickley and Guha, 2004] and OWL [McGuinness and van Harmelen, 2004], have nice properties such as being formal, non-ambiguous, rich (in direct proportion to the human effort invested), and enabling automatic inference based on Description Logics [Baader et al., 2003]. In the techniques discussed in the next section, the proposed representation to describe the meanings within content consists of a list of domain concepts that appear or happen in the content [Castells et al., 2007]. Concepts are associated to content by semantic annotation links, which may include time stamps when concepts occur in specific content segments, and one or several weight values indicating the strength of the link. The weights can reflect the importance of the concept in the content, the certainty that the concept actually occurs (e.g. when it has been recognized by an automatic content analysis technique), or any other relevant numeric measures. Ontology-based annotations can be created manually, or obtained through automatic means [Dasiopoulou et al., 2005], e.g. by extracting concepts from manual textual annotations, spotting words or known sounds in audio tracks, detecting objects and events in visual scenes, or by coordinating several of these approaches. This representation goes beyond the simpler and currently dominating forms of free annotation, commonly consisting of plain string keywords or arbitrary text sentences. It provides a unified, unambiguous representa- tion of the semantic space for annotation, hence a solid ground upon which powerful adaptation techniques can be devised, capable of making sense of the high-level semantic descriptions. Additionally, all the aforemen- tioned facilities supported by ontology-based technologies are available to the advantage of the development of personalization strategies. 80 Iván Cantador, Fernando López, Jesús Bescós, Pablo Castells, and José M. Martínez 7.3 Personalized Content Retrieval
In general terms, personalizing the selection of content for user access involves knowing something about the user beyond her last single request, and taking advantage of this knowledge in order to improve the system response to the actual user needs, in terms of the utility of the retrieved content for each individual user. Room for such improvement exists increasingly often, to varying degrees, in common retrieval scenarios, either be- cause the request is vague, in the sense that too much content matches it for the user to process (e.g. as happens most of the time in large-scale content spaces), or because there is no explicit request (e.g. the user is not in- stantly aware of new content of interest). In addition to the requirements in the scope of content description, the problems to be addressed to realize this view can be related to the areas of representation, elicitation, and exploitation of user preferences.
7.3.1 User Preference Modelling
Modelling user preferences is central to content personalization technologies, as it forms the basis on which the system may automate decisions on behalf of users. User profiles for multimedia retrieval provide a logical view of known or predicted user particularities and needs with respect to content features and formats, in a suitable form for use by the algorithms that implement the automatic adaptation strategies. A variety of structures and paradigms have been used in research and industry for the representation of relevant user information in this context. Among standardization efforts, the MPEG-21 DIA stands out as one of the most exhaustive speci- fications covering, among other more media-related features, information such as demographic data, general preferences, audiovisual presentation preferences, disabilities, focus of attention, etc. The Multimedia Descrip- tion Schemes in MPEG-7 provides additional structures for storing usage history data [Manjunath et al., 2003]. MPEG-21 describes the user information which, along with content descriptions in MPEG-7, provides a suit- able basis for multimedia-specific content adaptation in the delivery phase, as will be described in Section 7.4. Related to content selection, the usage environment data described in MPEG-21 is generally not enough (or not relevant) to make far-reaching predictions of what content choice a user might enjoy. To this respect, it is more important to know what the content is about, and what subjects, concepts or objects the user is interested in. A common approach to model this kind of preference is the vector-space model, in which user interests are represented as a vector of weights defining the intensity of interest for different things. The space of such things, in terms of which user interests are described, can correspond to elements of diverse nature, such as individual documents, categories from a taxonomy, or even plain words [Chirita et al., 2005, Jeh and Widom, 2003, Micarelli and Sciarrone, 2004]. In our research, we have explored the potential of enhanced representations of meanings as the foundation of this preference space, beyond prior simple approaches, as a means to enable improvements in the reach and accuracy of personalization [Cantador et al., 2008, Gauch et al., 2004]. More specifically, we have developed ontology-based preference models, where the preference space is based on domain ontologies [Staab and Studer, 2004], in such a way that user interests are represented by a weight for each domain concept in the ontology. These weights can be derived from the observation of user activity (queries and accessed content) over a period of time, and the repeated occurrence of concepts related to the objects involved in this activity. This is of course a problem on its own; the reader is referred to e.g. [Cantador et al., 2008, Mylonas et al., 2008] for further discussion. The advantage of using ontologies as the representational grounding for the semantic space lies in the precise and detailed information that is made available for the system e.g. to match user preferences to con- tent descriptions, enabling substan-tially more elaborate strategies than simpler representations support. For instance, the system can derive new user preferences from the initial ones based on the extra knowledge sup- plied by an ontology, using the formal inference mechanisms supported by ontology standards. This is further explained when we discuss the use for context modelling and content filtering in the next sections. An alternative (or complementary) view on the personalized recommendation of content is the social one, in which users are not just analyzed in isolation, but in comparison to other users. A well-known approach in this direction is the so-called collaborative filtering strategy, which is based on the general assumption that users with common traits may enjoy the same or related content [Adomavicius and Tuzhilin, 2005]. This principle raises the problem of measuring the similarity between users, which can be done based on a comparison of the 7 Enhanced Media Descriptions for Audiovisual Content Retrieval 81 profile information (e.g. semantic preferences, demographic information), or the history of common choices. The similarity of picked contents can be in turn measured by functions other than straight equality, where again, a rich semantic representation of content enables the detection of indirect similarities between items, that would not be found by checking for plain coincidence. For in-stance, a user who liked the films Interiors and Annie Hall, and a user who liked Broadway Danny Rose and Manhattan would be considered candidates for a possible affinity based on indirect evidence, since they both enjoyed films directed by Woody Allen, even if they did not watch the same films. This is possible if a domain ontology on cinema is available to the system, where movies, directors, actors and so on are defined as interrelated domain concepts and individuals.
7.3.2 Context Modelling
In order to properly customize content, knowledge about the context in which the content is sought and con- sumed by a particular user at a given time, is relevant as well. Research in this area has commonly considered user preferences and terminal and network capabilities (which we discuss later in this chapter) as part of the context model. Beyond this, other elements addressed in the literature include the rendering context (e.g. noise, illumination), location, time, meteorological data, etc. Besides the user preference information described in the previous section, in our research on content re- trieval adaptation we consider three additional contextual dimensions, namely (i) the dynamic user focus, con- sisting of a weighted, semantically coherent set of domain concepts that have been involved, in some way or other, in an ongoing user session [Vallet et al., 2007]; (ii) the semantic context of meanings, defined as the domain concepts closely related (through semantic relations explicitly defined in an ontology) to a given set of concepts [Mylonas et al., 2008]; and (iii) the social context of a user, consisting of the sets of users related to her in different possible ways [Cantador and Castells, 2006]. The semantic context of a concept is given by the concepts around it in the semantic network defined by a domain ontology, based on the relations that interlink the concepts. Semantic paths provide a basis to define distance measures between concepts, upon which we build fuzzy contextual supersets for a given set of concepts [Vallet et al., 2007]. This step is key in our approach to handle the social context, in which further similarities between users are found by comparing the semantic context of their preferences [Cantador and Castells, 2006]. This is applied in a similar way in order to take advantage of the live user focus to contextualize persistent user preferences, and enable a more accurate personalization of retrieval results [Mylonas et al., 2008], as will be described in the next section.
7.3.3 Content Filtering
Based on the relevant knowledge about user interests that a system is able to capture and elaborate on, different techniques can be used to personalize search and retrieval results, which come into play at different points of a retrieval system. For instance, user preferences have been used to reformulate user queries, by expanding or refining the information in the query (e.g. adding, reweighing, or even disambiguating terms), using informa- tion from the user profile and history [Rocchio, 1971]. Preferences can also be applied on the search result after the query has been answered. For example, the results can be re-ranked by a complementary similarity mea- sure between documents and user preferences [Chirita et al., 2005, Micarelli and Sciarrone, 2004]; clustering techniques have also been proposed to group results in sets of categories, ranking higher the more relevant cat- egories by user preference [Liu et al., 2004]. A long series of variations of the popular PageRank algorithm for Web search have been researched as well, which use different initialization strategies based on user preferences, in such a way that the resulting PageRank value is biased towards individual user interests [Jeh and Widom, 2003]. In scenarios where information (preferences, history) of a large number of users is centrally available to the personalization system, it is also possible to apply collaborative filtering techniques in a way that thousands of users benefit from each other’s experience without even getting ever acquainted, as discussed earlier. We have explored the potential for improvement that can be gained in such different strategies by building on the semantic enhancements for the representation of user interests described in Section 7.3.1. To begin with, ontologies enable the automatic extension of user preference reach by means of inference steps. For instance, 82 Iván Cantador, Fernando López, Jesús Bescós, Pablo Castells, and José M. Martínez a preference for pets would automatically expand to cats and dogs. On a more complex path, a preference for award-winning films can be defined in an ontology language like OWL by a restriction class defined as the movies that have won some award. The user can have his preference fulfilled by the system inferring that, say, an Oscar is a subclass of award, and a movie that won some Oscar is an instance of the preferred class by the user. Our primary application of these principles has been brought to a basic content filtering approach, which re-ranks search results by comparing user preferences and content metadata in the ontology-based vector space as described in earlier sections. This comparison is achieved by a similarity measure that compares these two concept vectors (user preference and content semantics) based on the cosine function. This measure is then combined with the relevance score returned by the search engine (which is taken here as a black box) for each content item and the query being answered, thus introducing a personalized bias in the ranking [Cantador et al., 2008]. This approach is further elaborated by contextualizing the preference model before it is matched to content descriptions, based on the live user focus introduced in the previous section. In our approach, putting prefer- ences in context means to activate only those persistent preferences that are semantically close to what is going on in the session and what the user has in mind at that time. This matches the way human preferences work in real life where, for instance, bringing up a long-term (e.g. professional) user interest for pharmaceuticals when a user is searching for a good movie in the weekend could be generally out of place. This feature is addressed as follows. During a retrieval session, the system collects all the concepts involved in user queries and metadata of the content se-lected by the user. A vector of weighted concepts (where weights decay with time) is built in this way, which is taken to be representative of the short-term user attention, or at least to be related to it. Then, the contextualization mechanism starts by estimating the semantic distance between each concept in the preference set and the live focus set, which is achieved by scanning the number and length of paths connecting them through the semantic network defined in the ontology. This is achieved by computing the extended fuzzy semantic context of user preferences and focus, and their fuzzy intersection. The resulting degrees of membership of concepts to this fuzzy set is combined with the original preference weight in the long-term user profile to produce a contextualized preference vector, where the weight of preferences that are semantically unrelated or far from the current user focus will be close to zero, thus achieving the desired effect [Mylonas et al., 2008, Vallet et al., 2007]. The semantic context modelling and user profile extension techniques have also been applied in our re- search to elaborate on the collaborative recommendation approach, as introduced in Section 7.3.1. Besides the extension of similarity functions based on indirect semantic comparisons discussed earlier in this chapter, we have explored a step beyond this by further refining the model with the notion of semantic layers within user preferences, as we explain next. In typical collaborative filtering approaches, the comparison between users is performed globally, in such a way that partial, but strong and useful similarities might be missed. For instance, two people may have a highly coincident taste in cinema, but a very divergent one in sports. Their opinions on movies could be highly valuable for each other, but risk to be ignored by many recommender systems, because the global similarity between the users might be low. We thus contend for the distinction of different layers within the interests of users, as a useful refinement to produce better recommendations. This idea is achieved by analyzing the structure of the domain ontology, the weighted links between users and ontology concepts (as defined by preferences), the links between concepts and contents (annotations), and the links (explicit ratings) between content and users. Based on this rich interrelation within and across the three spaces (users, concepts, content), we develop strategies of coordinated clustering which produce focused recommendations based on partial but cohesive similarities. Our approach finds groups of interests shared by users, and communities of interest among users. Users who share interests of a specific concept cluster are connected in the corresponding community, where their preference weights determine the degree of membership to that cluster. This enables focused recommendations in the different communities [Cantador and Castells, 2006]. At the same time, this approach tackles the common sparsity problems of the collaborative filtering approach in large-scale retrieval spaces (i.e. user preference information being scarce when an item is new, a user is new, or user ratings are thinly spread over a huge number of items [Adomavicius and Tuzhilin, 2005]), by finding indirect similarities among users and content items. Note that links between users (contacts) 7 Enhanced Media Descriptions for Audiovisual Content Retrieval 83 and between documents (e.g. hyperlinks or other cross-references) could also be considered as part of the available information to further enhance the analysis, which we foresee as future research.
7.4 Content Adaptation
Once a specific content item is picked by the user from a list of choices retrieved by the system according to user queries, preference, or any of the strategies discussed in the previous section, the time comes to actually deliver the content itself in the most suitable form. At this point, the development of new access networks providing multimedia capabilities, and a wide and growing range of terminals, makes the adaptation of content a major issue for future multimedia services. Content adaptation is the main objective of a set of technologies that can be grouped under the umbrella of Universal Multimedia Access (UMA) [Vetro et al., 2003]. This means the capability of accessing to rich multimedia content through any client terminal and network. In this way content adaptation bridges content authors and content consumers in a world of increasing multimedia diversity. In order to perform content adaptation it is necessary to have the description of the content and the descrip- tion of the usage environment. To enhance the user’s experience [Pereira and Burnett, 2003], not only terminal and network characteristics and conditions should be taken into account when adapting, but also user prefer- ences and disabilities, as well as environmental conditions. All this information imposes some constraints to the content to be obtained after adaptation.
7.4.1 Content Adaptation Taxonomy
According to the level of understanding applied to the media, multimedia adaptation can be performed in two different ways:
Signal level adaptation, committed to transcoding media resources without knowledge of the meaning of • the content. Semantic level adaptation, which modifies the media assuming that there exists some knowledge about • the meaning of the content.
According to the information used to decide the adaptation to perform, multimedia adaptation can be based on three different sources:
Usage environment driven adaptation. The terminal, network and user preferences are taken into account • to decide the adaptation to perform to the multimedia content. Semantic driven adaptation evaluates the content of the media to select the more relevant parts. This kind • of adaptation has been mainly used to perform content summarization. Perception driven adaptation performs transformations according to some user preferred sensation, or • assists a user with a certain perception limitation or condition. The user perception of the content will be different from its original version, for example, to address the needs of users with visual handicaps (e.g. colour blind deficiencies, or specific preferences) in terms of visual temperature sensations. This class of customization operations considers all types of adaptations related to some human perceptual preferences, characteristics or handicaps.
According to the way to group media, adaptation can be performed at two different levels:
Media level adaptation, committed to adapt a unique resource, named media resource, and usually stored • as a file. Content to be adapted at this level can also be represented as a stream of bytes. Typically, a standardization body defines the whole format of this single modality media. System level adaptation (also named multimedia level), committed to adapt one or more resources grouped • as a compound resource, named system resource (e.g. a web page, or a SMIL file). This system resource can also convey metadata (descriptors). Within the system level compositions, we can identify three main kinds of adaptation: (i) structure level adaptation adapts a resource that is the union of (or has references to) other 84 Iván Cantador, Fernando López, Jesús Bescós, Pablo Castells, and José M. Martínez
media resources (e.g. DI, HTML); (ii) layout level adaptation changes the arrangements of the constituent elements in the scene (e.g. HTML, SMIL); (iii) synchronization level adaptation modifies the timeline of the constituent resources (e.g. MPEG files summarization, or SMIL image serialization). For example, an MPEG-21 Digital Item Declaration [Burnett et al., 2006] is capable of conveying media resources and descriptors grouped as a system resource named Digital Item (DI). During the adaptation, the number of resources and descriptor elements may change.
These two adaptation levels (media and system) can be performed in a signal or semantic way. Conversely, signal and semantic level adaptations can be performed in a media or system level. Another dimension to consider in an adaptation framework are the different adaptation approaches, that can be also applied with content understanding (semantic level adaptation) or without it (blind or signal level adaptation); all these approaches are expected to coexist [Vetro, 2003]: Transcoding. This is the most frequent adaptation, where the media parameters are changed in order to fit • the constraints imposed by the usage environment. Scaling, colour reduction, coding standard, etc. Transmoding. This adaptation implies a change of modality of the original media in order to enable it to • be presented in the selected scenario. Text to audio, video to slideshow or image to video are among this new generation of content adaptation. Scalable content truncation. Scalable formats are based on the co-existence in the same file or stream of • different versions of the same content at different temporal, spatial and quality resolutions. Efficient tools for management of such content are provided by bitstream description tools within MPEG-21 DIA [Vetro and Timmerer, 2005]. Summarization. Video and audio skimming (creating excerpts of the temporal content, as for video trail- • ers) are the most common approaches to summarization, although there are other proposals based on ad- ditional transmoding (e.g., a video poster consisting of the composition into one poster image of selected keyframes of the summarised video). The presentation of information provides different options to sum- marization, whilst the common ground are the content analysis technologies for the identification of the relevant parts of the content: keyframes or relevant video segments of a video, chorus of a song, main theme of a symphony, etc.
7.4.2 Content Adaptation Engine
Content adaptation is performed by a module of the complete content retrieval system, that is usually named Content Adaptation Engine. It aims to provide the user the best experience when accessing the requested content within the current usage environment. A Content Adaptation Engine includes two main functionalities: deciding the adaptation to perform, and managing the decided adaptation process. In the simplest cases an Adaptation Engine is only able to perform one adaptation (e.g., reducing the spatial resolution of an image to a half in each dimension) and therefore no decision phase is required. The next complexity step comes up when the Content Adaptation Engine can choose between different values for certain parameters of the adaptation (e.g. reducing the spatial resolution in a predefined range). Additional complexity appears when a Content Adaptation Engine is able to perform two (or more) adaptation processes, each with its own parameters (e.g. for achieving a target file size, spatial resolution or bits per pixel can be reduced). In this case, the decision is not only parameter-based; instead, a combination of adaptation processes and their parameters can achieve the target adaptation. According to the identified functionality, a generic Content Adaptation Engine can be modelled to include two main submodules: a Decision Module and a set of Content Adaptation Tools (CATs). Each CAT is able to perform a particular adaptation. The different available CATs may diverge in the adaptation approach (e.g. transcoding, transmoding, etc), the range of parameters values, the supported input and output formats, the per- formance (in terms of processing requirements, quality, etc), and so forth. Hence, there is a need for describing the CAT adaptation capabilities [Valdés and Martínez, 2006a], so that the Decision Module is able to incor- porate also this description when making a decision primarily based on the aforementioned content and usage environment descriptions. 7 Enhanced Media Descriptions for Audiovisual Content Retrieval 85
Several approaches have been proposed to perform content adaptation [Jannach et al., 2006, Magalhaes and Pereira, 2004, Martínez et al., 2005, Tseng et al., 2004, Vetro and Timmerer, 2005]. We have developed a Content Adaptation Engine, named CAIN (for Content Adaptation Integrator) [Martínez et al., 2005], aimed to the integration of different content adaptation approaches, that is, different CATs [Valdés and Martínez, 2006b]. In CAIN the Decision Module uses Constraints Programming [López and Martínez, 2007] for selecting the CAT that is best suited to the optional and mandatory constraints imposed, respectively, by terminal and network characteristics and user preferences [López et al., 2006]. It should be noted that [Jannach et al., 2006] proposes the use of a scheduling algorithm to find a chain of elementary adaptation operations which transform the media accordingly, whilst the CAIN framework considers CATs that perform several elementary adaptations. The Decision Module selects only one CAT from the available ones (we are evaluating to extend our solution in order to allow the concatenation of CATs in the future). The adaptation methods are constructed upon the foundations provided by MPEG description standards: content descriptions are based on MPEG-7 MDS [Manjunath et al., 2003] and MPEG-21 DIA BSD [Burnett et al., 2006], whilst the context descriptions are based on a subset of MPEG-21 DIA Usage Environment Descriptions tools [Burnett et al., 2006].
7.5 Conclusions
Automatic adaptation is a major issue in modern multimedia content delivery environments and infrastructures. In this chapter we have discussed a set of adaptation techniques that address this need, spanning across the whole content retrieval and delivery cycle, from the selection and choice of content units, based on high-level descriptions of the meanings within, to the actual delivery of content streams, adapted to the available conditions at the consumer end-point. The main innovations in our proposed approaches for personalized retrieval focus on the potential for im- provements enabled by working on (i) the representation of semantics and (ii) the consideration of the retrieval user context, where in our model so far the latter includes semantic contexts, the social environment, and the user focus. The ontology-based approach to semantics representation has its own trade-offs, the main ones being the limited availability of ontologies, and the development cost and formalization problems involved in defin- ing very detailed ones. However, the proposed personalization techniques are tolerant to incomplete knowledge, which means that they make the most of high-quality semantic information whenever it is available, and they degrade gracefully to the performance of any standard technique based on simpler representations (e.g. key- words, documents) with which our techniques can be combined, as the completeness of ontological knowledge decreases for the domain area at hand. The proposed techniques for content adaptation are characterized by the capability to both decide on the most appropriate adaptation strategy for the dynamic situation in process, and actually carry out and manage the selected adaptation approach. The adaptation of content is built upon the expressive capabilities of the MPEG standards. Content and access adaptation together address the challenges raised by the new order of magnitude in the scale of current content environments, and the heterogeneous and dynamic nature of delivery platforms and user audiences, in order to deliver the subjective content service quality that an enhanced user experience requires. In order to reach this point, automatic content understanding is currently the most challenging research issue in content analysis in order to be able to perform semantic based adaptation based on what is present in the content.
7.6 Acknowledgements
This research was partially supported by the European Commission under contracts FP6-001765 aceMedia and FP6-027685 MESH. The expressed content is the view of the authors but not necessarily the view of the aceMedia or MESH projects as a whole. Thanks are due to the members of the NETS and GTI research groups at EPS-UAM who have collaborated in these and other research projects, where the experiences on multimedia content customization reported here have been acquired: Miriam Fernández, Víctor Fernández- Carbajales, Javier Molina, Víctor Valdés, and David Vallet. 86 Iván Cantador, Fernando López, Jesús Bescós, Pablo Castells, and José M. Martínez References
G. Adomavicius and A. Tuzhilin. Toward the next generation of recommender systems: A survey of the state- of-the-art and possible extensions. IEEE Transactions on Knowledge and Data Engineering, 17(6):734–749, 2005. F. Baader, D. Calvanese, D. L. McGuinness, D. Nardi, and P. F. Patel-Schneider. The Description Logic Hand- book: Theory, Implementation, Applications. Cambridge University Press, 2003. D. Brickley and R. V. Guha. RDF Vocabulary Description Language 1.0: RDF Schema. W3C Recommenda- tion, February 2004. I. S. Burnett, F. Pereira, R. Van de Walle, and R. Koenen. The MPEG-21 Book. Wiley & Sons, 2006. I. Cantador and P. Castells. Multilayered semantic social network modelling by ontology-based user profiles clustering: Application to collaborative filtering. In 15th International Conference on Knowledge Engineer- ing and Knowledge Management, Podebrady, Czech Republic, 2006. I. Cantador, M. Fernández, D. Vallet, P. Castells, J. Picault, and M. Ribière. Advances in Semantic Media Adaptation and Personalization, chapter A Multi-Purpose Ontology-Based Approach for Personalised Con- tent Filtering and Retrieval. Springer Verlag, 2008. P. Castells, M. Fernández, and D. Vallet. An adaptation of the vector-space model for ontology-based informa- tion retrieval. IEEE Transactions on Knowledge and Data Engineering, 19(2):261–272, 2007. P. A. Chirita, W. Nejdl, R. Paiu, and C. Kohlschütter. Using ODP metadata to personalize search. In 28th International ACM SIGIR Conference on Research and Development in information Retrieval, Salvador, Brazil, 2005. S. Dasiopoulou, V. Mezaris, I. Kompatsiaris, V. K. Papastathis, and M. G. Strintzis. Knowledge-assisted se- mantic video object detection. IEEE Transactions on Circuits and Systems for Video, 15(10):1210–1224, 2005. S. Gauch, J. Chaffee, and A. Pretschner. Ontology-based personalized search and browsing. Web Intelligence and Agent Systems, 1(3-4):219–234, 2004. D. Jannach, K. Leopold, C. Timmerer, and H. Hellwagner. A knowledge-based framework for multimedia adaptation. International Journal on Applied Intelligence, 24(2):109–125, 2006. G. Jeh and J. Widom. Scaling personalized web search. In 12th International World-Wide Web Conference, Budapest, Hungary, 2003. D. Lenat and R. V. Guha. Building Large Knowledge-Based Systems: Representation and Inference in the Cyc Project. Addison-Wesley, 1990. F. Liu, C. Yu, and W. Meng. Personalized web search for improving retrieval effectiveness. IEEE Transactions on Knowledge and Data Engineering, 16(1):28–40, 2004. F. López and J. M. Martínez. Multimedia content adaptation modelled as a constraints matching problem with optimisation. In Workshop on Image Analysis for Multimedia Interactive Services, Santorini, Greece, 2007. F. López, J. M. Martínez, and V. Valdés. Multimedia content adaptation within the CAIN framework via constraints satisfaction and optimization. In 4th International Workshop on Adaptative Multimedia Retrieval, Geneva, Switzerland, 2006. J. Magalhaes and F. Pereira. Using MPEG standards for multimedia customization. Signal Processing: Image Communications, 19(5):437–456, 2004. B. S. Manjunath, P. Salembier, and T. Sikora. Introduction to MPEG-7 Multimedia Content Description Inter- face. Wiley & Sons, 2003. J. M. Martínez, V. Valdés, J. Bescós, and L. Herranz. Introducing CAIN: A meta-data-driven content adaptation manager integrating heterogeneous content adaptation tools. In Workshop on Image Analysis for Multimedia Interactive Services, Montreux, Switzerland, 2005. D. L. McGuinness and F. van Harmelen. OWL Web Ontology Language Overview. W3C Recommendation, February 2004. A. Micarelli and F. Sciarrone. Anatomy and empirical evaluation of an adaptive web-based information filtering system. User Modeling and User-Adapted Interaction, 14(2-3):159–200, 2004. P. Mylonas, D. Vallet, P. Castells, M. Fernández, and Y. Avrithis. Personalized information retrieval based on context and ontological knowledge. Knowledge Engineering Review, 23(1), 2008. 7 Enhanced Media Descriptions for Audiovisual Content Retrieval 87
F. Pereira and I. Burnett. Universal multimedia experiences for tomorrow. IEEE Signal Processing Magazine, 20(2):63–73, 2003. J. Rocchio. The SMART Retrieval System: Experiments in Automatic Document Processing, chapter Relevance feedback information retrieval. Prentice-Hall, 1971. S. Staab and R. Studer, editors. Handbook on Ontologies. Springer Verlag, 2004. B. L. Tseng, C. Y. Lin, and J. R. Smith. Using MPEG-7 and MPEG-21 for personalizing video. IEEE Multi- media, 11(1):42–53, 2004. V. Valdés and J. M. Martínez. Content adaptation capabilities description tool for supporting extensibility in the CAIN framework. In International Workshop on Multimedia Content Representation, Classification and Security, Istambul, Turkey, 2006a. V. Valdés and J. M. Martínez. Content adaptation tools in the CAIN framework. In 9th International Workshop Visual Content Processing and Representation, Sardinia, Italy, 2006b. D. Vallet, P. Castells, M. Fernández, P. Mylonas, and Y. Avrithis. Personalized content retrieval in context using ontological knowledge. IEEE Transactions on Circuits and Systems for Video Technology, 17(3):336–346, 2007. A. Vetro. Transcoding, scalable coding and standardized metadata. In 8th International Workshop on Visual Content Processing and Representation, Madrid, Spain, 2003. A. Vetro and C. Timmerer. Digital item adaptation: Overview of standardization and research activities. IEEE Transactions on Multimedia, 7(3):418–426, 2005. A. Vetro, C. Christopoulos, and T. Ebrahimi. Special issue on universal multimedia access. IEEE Signal Processing Magazine, 20(2), 2003.
8 A Semantics-Based Approach Beyond Traditional Personalization Paradigms
Yolanda Blanco-Fernández, Alberto Gil-Solla and Manuel Ramos-Cabrer
Department of Telematics Engineering, University of Vigo, 36310, Spain {yolanda,agil,mramos}@det.uvigo.es
Summary. The so-called recommender systems have become assistance tools indispensable to the users in domains where the information overload hampers manual searching processes. In literature, diverse person- alization paradigms have been proposed to match the personal preferences of each user against the available items. All these paradigms are laid down on a common substratum that uses syntactic matching techniques, which greatly limit the quality of the offered recommendations due to their inflexible nature. To fight these limitations, this paper explores a novel approach based on reasoning about the semantics of both the users’ preferences and considered items, by resorting to less rigid inference mechanisms borrowed from the Semantic Web. Our reasoning-based approach shape enhanced personalization strategies that migrate from the traditional syntactic proposals to the inference of semantic relationships between the users’ interests and the available items. Thanks to the exploitation of semantics, the discovered relationships provide the recommender system with additional knowledge about the user’s preferences, thus leading to more accurate suggestion processes. The flexibility and strength of our approach enable its application in multiple domains and personalization tools. Specifically, for testing and validation purposes, our semantics-based strategies have been incorporated and evaluated in several tools bound to the field of Digital TV, where viewers are exposed to overwhelming amounts of audiovisual contents and challenged by the plethora of interactive functionality provided by the current digital receivers. The achieved results have revealed both the scalable nature and the computational feasibility of our domain-independent personalization approach based on semantic reasoning.
8.1 Introduction
Since the middle of nineties, recommender systems have been gaining an increasing interest as tools that help the users by selecting automatically items of interest in domains where the overload of information is notice- able [Adomavicius and Tuzhilin, 2005]. For that purpose, such personalization tools employ matching tech- niques that compare the available items with the users’ preferences, which are previously modeled in personal profiles. As Qing Li et al. describe elsewhere in this book, the most used strategies can be grouped into three main categories, named content-based filtering, collaborative filtering and hybrid approaches. Broadly speak- ing, a content-based filtering approach seeks to recommend similar items to those the user liked in the past, whereas the collaborative approaches select items for a given user that similar users also appreciated, as we detail next.
Content-based filtering relies on an ability to accurately represent recommendable items in terms of a • suitable set of content attributes, and to represent user profile information as an analogous feature set. In this strategy, the user’s profile effectively delimits a region of the item-space from which all future suggestions will be drawn, in such a way that the relevance of a given content item to a specific target user is proportional to the similarity of this item to her profile (see [Hammond et al., 1995, Smyth and Cotter, 1999]). To measure these similarity values, traditional approaches resort to syntactic metrics that 90 Yolanda Blanco-Fernández, Alberto Gil-Solla and Manuel Ramos-Cabrer
only permit to detect resemblance among items sharing identical attributes, thus leading to overspecialized recommendations that suffer from a limited diversity. This is particularly problematic for new users since their recommendations will be based on a reduced set of items represented in their immature profiles. So, items relevant to a user, but bear little or no resemblance to the items in her profile, will never be suggested in the future. Collaborative filtering. As explained in [Goldberg et al., 1992, Maltz and Ehrlich, 1995], the basic idea • in the collaborative paradigm is to move beyond the experience of an individual user’s profile, and instead draw on the experiences of a community of users. Typically, each target user is associated to a set of nearest- neighbor users by comparing their respective profiles. To delimit the user’s neighborhood, collaborative techniques look for correlation between users in terms of the ratings they assigned to the items contained in their profiles. From this it follows that two users are neighbors when they have rated common items in their profiles by assigning them similar ratings. Once the user’s neighborhood has been created, collaborative approaches predict her interest in the items rated by those neighbors. As a result, the items with highest ratings in the neighbors’ profiles are finally suggested to the target user. Collaborative approaches also suffer from severe drawbacks, such as the so-called sparsity problem and la- tency concerns. As for the first limitation, note that as the number of available items increases, it is unlikely that two users rate the same items in their profiles, thus hampering the selection of the user’s neighbors. Besides, in this situation, the creation of the user’s neighborhood (based on computing correlations be- tween vectors containing the users’ ratings) becomes too demanding in computational terms, resulting in substantial scalability problems. The second-mentioned limitation is related to the fact that collaborative filtering is not suitable for suggesting new items because only those already rated by other users can be recommended. So, if a new item is added to the content database in the recommender system, there can be a significant delay before this item will be considered for recommendation (see [Adomavicius and Tuzhilin, 2005] for details). Hybrid approaches. Bearing in mind the internals of content-based and collaborative filtering techniques, • researchers tended to define hybrid approaches where both techniques complement each other perfectly, al- lowing to exploit their advantages and mitigate their respective weaknesses. As Burke described in [Burke, 2002], the most of the hybrid recommender systems adopt the so-called “collaboration via content” paradigm by Pazzani [Pazzani, 1999], based on computing the similarity between two users’s preferences by considering both the content descriptions (just like in content-based approaches), and their respective ratings (considered in collaborative filtering). This way, Pazzani fights the sparsity problem by detecting that two users have similar preferences even when there does not exist overlap between the items contained in their profiles. However, in order to measure similarity in this case, it is necessary that these items have common attributes, hence the fact that Pazzani’s approach is still limited by the syntactic metrics used in the traditional content-based techniques. In this paper, we describe how to overcome this kind of limitations —derived from the purely syntactic techniques adopted by existing personalization paradigms— by exploiting the semantics of the application domain of each recommender system. To this aim, our approach resorts to reasoning mechanisms bound to the so-called Semantic Web, in which information is given a well-defined meaning, better enabling computers and people to work in cooperation. To achieve this goal, three main elements are required in the Semantic Web: firstly, metadata to annotate the available Web resources; secondly, a domain ontology where these semantic annotations are formalized; and finally, some reasoning techniques to infer semantic relationships among the annotated resources. Specifically, the personalization approach we describe in this paper proposes new matching techniques based on inferring semantic relationships between the user’s preferences and the available items, which would go unnoticed in traditional syntactic mechanisms. These relationships allow the recommender system to dis- cover a huge amount of knowledge about the user’s interests, thus enabling to select more accurately items of interest for her. To illustrate synergies existing between semantic reasoning and recommender systems, our domain-independent approach defines two strategies —using content-based and collaborative filtering, respectively– that can be either separately applied or mixed in an advanced hybrid technique. On the one hand, our content-based approach mitigates the overspecialization of the traditional sugges- • tions by using a new similarity metrics that reasons about the semantics of the compared items. Thanks 8 A Semantics-Based Approach Beyond Traditional Personalization Paradigms 91
to the flexibility enabled by semantics, our metrics measures resemblance between items that do not share identical attributes, thus endowing the offered recommendation with a rich diversity. Regarding the collaborative strategy, our reasoning capabilites are exploited during both the process of • neighborhood formation for each user, and the selection of her personalized recommendations. This way, it is possible to: (i) detect similarity between users’ profiles that do not contain common items, and (ii) suggest items that have not been rated by any user in the system, thus alleviating the adverse effects of both the sparsity problem and the latency concerns present in current collaborative approaches. This paper is organized as follows: Section 8.2 focuses on the two key components of our reasoning frame- work: the ontology where the semantic annotations of the items available in the personalization tool are for- malized, and the techniques employed to model the users’ preferences. Next, Sections 8.3 and 8.4 describe in detail how our content-based and collaborative filtering strategies mitigate the intrinsic drawbacks of the tradi- tional paradigms. Then, Section 8.5 highlights aspects related to the experimental evaluation of our strategies in different personalization system for Digital TV, and it discusses concerns about scalability and computa- tional viability of our approach. Finally, Section 8.6 concludes the paper and points out possible lines of further research.
8.2 The Reasoning Framework 8.2.1 The Domain Ontology In the field of the Semantic Web, an ontology is a formal specification of a conceptualization, that is, an abstract and simplified view of the world that we wish to represent, described in a language that is equipped with a formal semantics [Berners-Lee et al., 2001]. An ontology characterizes that semantics in terms of concepts and their relationships, which are represented by classes and properties, respectively. Both entities are hierarchically organized in the conceptualization, which is populated by including specific instances of both classes and properties. As an example, note that in the context of a recommender system, instances of classes represent the available items and their attributes, whereas instances of properties link the items and attributes each other. As a sample of this formalization process, we depict in Fig. 8.1 a brief excerpt from an ontology for the domain of television. Here it is possible to identify several instances that identify specific TV programs, which belong to a hierarchy of classes referred to different genres (e.g. Fiction, Sports, Music, Leisure). As we commented before, the attributes of these TV contents (e.g. cast, intented audience, topics) are also identified by hierarchically organized classes, and are related to each program by means of labeled properties (e.g. hasActor, hasIntendedAudience, isAbout). To define these semantic annotations it is possible to reuse existing collection of metadata that describe in detail the specific domain of application. For instance, in the TV field, the TV- Anytime metadata specification [TV-Anytime Forum, 2001] provides very detailed description about generic audiovisual contents, thus greatly simplifying the formalization process. Ontologies have become the cornerstone in the Semantic Web due to two reasons. On the one hand, as these conceptualization represent formally a specific domain, they permit to employ inference processes to discover new knowledge from the formalized information. On the other hand, ontologies facilitate automated knowledge sharing, by allowing an easy reusing between users and software agents. To this aim, several standard imple- mentation languages have been developed in the Semantic Web. The former proposals were RDF [Beckett, 2004] and RDFS [Brickley and Guha, 2004], which added a formal semantics to the purely syntactic specifi- cations provided in XML. Next, DAML [DAML, 2000] and OIL [Fensel et al., 2001] arose, which have been finally fused and standardized by W3C as OWL [McGuinness and van Harmelen, 2004]. Nowadays, OWL is the most expressive language in which three sub-levels have been defined (Lite, DL and Full). In this regard, note that the language used to implement the ontology required in our reasoning approach depends on the knowledge and expressive necessities of each application domain and each recommender system.
8.2.2 The User Modeling Technique To reason about the user’s preferences, our approach needs a formal representation including semantic descrip- tions of the items which are interesting and unappealing to him/her (named positive and negative preferences, 92 Yolanda Blanco-Fernández, Alberto Gil-Solla and Manuel Ramos-Cabrer
[4] [3] About your M. Schumacher Idols F1 race Formula1 Interviews
NON-FICTION Documentaries CONTENTS [2] Motor Sports SPORTS Tom FICTION Comedy The Money Pit Hanks
CONTENTS Kart racing
¡ ¡ ¡ ¡ ¡ ¡ ¡ ¢ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¢ ¡ ¡ ¡ ¡
[2]
Karting
’
You ve Got
Mail Jacky
[3]
Blues &