Embed Size (px)
Citation preview
18 1541-1672/07/$25.00 © 2007 IEEE IEEE INTELLIGENT SYSTEMSPublished by the IEEE Computer Society
G u e s t E d i t o r s ’ I n t r o d u c t i o n
Recommender SystemsAlexander Felfernig and Gerhard Friedrich, University of Klagenfurt
Lars Schmidt-Thieme, University of Hildesheim
R ecommender systems support users by identifying interesting products and ser-
vices in situations where the number and complexity of offers outstrips the user’s
capability to survey them and reach a decision. Interest in recommender systems has
dramatically increased owing to the demand for personalization technologies from large,
successful e-commerce applications (such as Amazon.com). Nowadays, numerous online shops employrecommender applications, which many regard as akey enabling technology of e-commerce.
Corresponding applications recommend every-thing from news, Web sites, CDs, books, and moviesto more complex items such as financial services,digital cameras, or e-government services. Recom-mendations are determined either by explicitly con-ducting sales dialogues with online users or by ana-lyzing existing purchasing data from a single user ora community of users. Following the latter approach,the first recommender applications, developed in themid-1990s, aimed to aggregate existing rating infor-mation to derive new user recommendations.
Recommendation approachesOne of the most frequently used representations
of recommendation tasks is collaborative filtering,1,2
essentially a nearest-neighbor method applied to aratings matrix. For example, if two customers boughtsimilar CDs and rated them similarly, the systemwould recommend to one customer CDs that theother customer bought and rated positively. Collab-orative filtering perfectly implements the idea ofword-of-mouth promotion, in which the opinions offriends and benchmarking reports are the predomi-nant influences on a purchaser’s buying decision.
The first applications of recommender technolo-gies were the personalized recommendation of newsand Web sites; the former application is often basedon collaborative filtering, and the latter is based oncontent-based filtering.3 Content-based filtering uses
features of items the user liked in the past to infernew recommendations. Unlike collaborative ap-proaches, content-based filtering can’t provide ser-endipitous recommendations. It selects and recom-mends all products on the basis of purchasinginformation available from the current user. In bothcollaborative and content-based filtering, user pro-files are long-term models. These approaches don’texploit deep knowledge about the product domain,and both are excellent for supporting the recom-mendation of simple products such as books ormovies. A major strength of these approaches is thatno additional knowledge acquisition efforts are nec-essary if historical data is available.
In recent years, researchers have proposed morecomplex statistical models for recommender systemsthat improve recommendation quality—for exam-ple, Bayesian networks with a hidden class variable2
or compound classification models.4 Some have triedhybridization strategies that combine collaborativeand content-based information. Meanwhile, othershave worked on attribute-aware recommender mod-els that try to take co-user and product informationinto account in parallel.5
Customers purchasing complex products such asfinancial services, computers, or digital camerasneed both information and intelligent interactionmechanisms that support the selection of appropri-ate solutions. So, an explicit representation of prod-uct, marketing, and sales knowledge is necessary.Knowledge-based approaches use this type of rep-resentation.6–8 Such deep knowledge lets these rec-ommender systems
• calculate solutions that fulfill certain qual-ity requirements,
• explain solutions to a customer, and• support customers when the system can’t
find a solution.
In particular, explicit knowledge represen-tation enables the validation of a recommendersystem’s quality regarding calculated solu-tions. For example, in many applications, thegenerated solutions must suit customer re-quirements, adhere to legal regulations, and bein line with the company’s sales strategy.Knowledge-based recommender technologiesprovide the formalisms necessary in this con-text. Unlike the word-of-mouth promotion thatcollaborative filtering provides, knowledge-based recommendation implements explicitsales dialogues that help users select items.
In this issueThe articles in this special issue cover a wide
range of research topics. Francesco Ricci andQuang Nhat Nguyen present knowledge-based recommendation techniques that helpon-the-move travelers select pretravel ser-vices. The recommendation environment re-lies on a critique-based approach that supportsthe acquisition and revision of user prefer-ences. The recommendation system is con-nected to a recommender user interfacespecifically designed for mobile environ-ments. In addition to the detailed discussionof the user interface and its support for revis-ing critiques, the article also sketches theintegration of a mobile recommendationenvironment (MobyRek) with a pretravelplanning system (NutKing).
Jinhyung Cho, Kwiseok Kwon, and Yong-tae Park propose a collaborative-filteringmethod based on group behavior theory inconsumer psychology. The approach exploitsthe idea of defining dual information sources(groups of similar users and expert users).Depending on a user’s receptivity regardingthose groups, the system automatically adaptsits recommendation strategy. This approachcan improve classification accuracy com-pared to conventional collaborative filtering.
Product reviews and shared experiencesabout products are powerful informationsources that can improve recommendationquality. Silvana Aciar, Debbie Zhang, SimeonSimoff, and John Debenham present a struc-tured approach to including such informationsources in calculating product recommenda-tions. Their system exploits text-mining tech-niques to make textual information units
applicable for calculating recommendations,which can improve the quality of recom-mendations based on collaborative filtering.
Gediminas Adomavicius and YoungOkKwon present two new recommendationapproaches that use multicriteria recommen-dation. Multicriteria ratings provide addi-tional information about user preferencesregarding several aspects of an item, such asstory and acting quality in a movie. Experi-mental results from tests with a commercialdata set show that multicriteria approachesimprove recommendation accuracy comparedto traditional single-rating strategies.
Publicly available recommender systemspresent a security problem. Attackers can intro-duce biased profiles to try to adapt the recom-mendation behavior in a way that’s advanta-geous to the attacker. Bamshad Mobasher,Robin Burke, Runa Bhaumik, and J.J. Sand-vig present an approach to understanding,identifying, and defeating such profile injec-tion attacks. As a response to major attack pat-terns, they present algorithmic approaches forthe design of more robust recommenders.
Neil J. Hurley, Michael P. O’Mahony, andGuenole C.M. Silvestre discuss security inrecommender systems, focusing on a cost-benefit analysis of the financial gains thatresult from attacks on recommender systems.They also present a corresponding analysisframework.
Markus Zanker, Markus Jessenitschnig,Dietmar Jannach, and Sergiu Gordea evalu-ate different recommendation techniquesusing commercial data from an online cigarretailer. The evaluation compares three fun-damentally different recommendation ap-proaches (knowledge-based approaches,collaborative filtering, and content-based fil-tering). Primarily, they point out that collab-orative-filtering algorithms provide goodresults for big data sets but are less effectivefor small data sets, where hybrid and knowl-edge-based approaches perform better.
Finally, Dan Melamed, Bracha Shapira,and Yuval Elovici present an approach tosolving a problem that accompanies the appli-cation of collaborative-filtering systems: thefree-ride problem. This occurs when usersconsume other people’s evaluations withoutproviding their own evaluations. The articlepresents an incentive mechanism based on amarket-based model for pricing evaluationsthat motivates users to provide evaluations.
Future directionsIn addition to the topics this special issue
covers, we deem the following challenges asimportant future research directions.
Automated productdata extraction
In a knowledge-based system, the qualityof recommendations is directly correlatedwith the data quality. In many cases, productdata are available only in an unstructuredform. Furthermore, data in recommenderknowledge bases become outdated. The ma-jor research focuses in this context are theautomated extraction of product data fromdifferent information sources and the auto-mated detection and adaptation of outdatedproduct data. This includes identifying rele-vant information sources (for instance, inmost cases, Web pages), extracting productdata, and resolving contradictions in thosedata. A related recent challenge is extractingproduct information directly from digitalmultimedia products such as books, CDs,DVDs, and TV programs.
Community-basedrecommendation
State-of-the-art knowledge-based recom-mender environments focus on a single pointof knowledge acquisition. Such centralizedapproaches can’t guarantee the quality ofknowledge bases because those knowledge
MAY/JUNE 2007 www.computer.org/intelligent 19
bases integrate the views of only a smallgroup of experts. Mechanisms that effec-tively integrate the knowledge of a commu-nity of interest can potentially improveknowledge bases’ quality and make relatedknowledge-acquisition processes more effec-tive. Such participative architectures willenable the development of integrated knowl-edge bases and user applications.
Preference and assortment dynamics
With a few remarkable exceptions, mostactual recommender models work on profilesthat have been aggregated over time by sim-ply collecting preference indicators. But inmany domains, users’ preferences developover time, as do their assessments of certainproducts or product characteristics. Forinstance, a user who rated a digital cameraas technically innovative two years ago mightconsider that camera outdated today. Takingthis dynamic into account during modelbuilding will introduce a new level of com-plexity to recommender system models.
Intelligent testingSuccessfully developing and maintaining
recommender knowledge bases requires in-telligent testing environments that can guar-antee recommendations’correctness. So, fu-ture research must focus on developingmechanisms to automatically configure opti-mal test suites that both maximize the prob-ability of identifying faulty elements in therecommender knowledge base and minimizethe number of test cases needed to achievethis goal. Minimizing the number of testcases is important because domain expertsmust validate them manually.
Recommendation of configurableproducts and services
With the production of the Model T about100 years ago, Henry Ford revolutionizedmanufacturing by employing mass produc-tion (the efficient production of many iden-tical products). Nowadays, mass productionis an outmoded business model, and compa-nies must provide goods and services that fitcustomers’ individual needs. In this context,mass customization—the production ofhighly variant products and services undermass production pricing conditions—hasbecome the new paradigm.9 A phenomenonaccompanying mass customization is massconfusion,10 which occurs when items aretoo numerous and complex for users to sur-
vey. Developing recommender technologiesthat apply to configurable products and ser-vices can help tackle mass confusion.
Context awarenessIntegrating contextual elements into rec-
ommender algorithms is extremely impor-tant for improving recommendation results’quality. Such contextualization can play animportant role in mobile environments—forinstance, in the selection of travel services ordestinations for on-the-move tourists. In suchsituations, recommendations depend not onlyon customer preferences but also on context,which can include attributes such as time ofday, season, weather conditions, and ticketavailability.
Theories of consumer buying behavior
Another major research issue is the activeintegration of theories from cognitive psy-chology and decision theory11 into recom-mender applications. This integration willprovide new insights into the major factorsinfluencing consumer behavior and, conse-quently, could result in more efficient rec-ommender systems in interactive sellingenvironments.
Intelligibility and explanationTo be convincing, recommendations must
be explained to customers. Only when theycan challenge a recommendation and seewhy a system recommended a specific prod-uct will customers start to trust that system.Conceptual, knowledge-based, logical mod-els that can provide such explanations shouldwork in tandem with recommender systemsbased on machine-learning models. Com-bining statistical models’ predictive powerwith the intelligibility of conceptual, human-
understandable models might turn out to bea major success factor for recommender sys-tems in the next decade.
Recommender technologies’high indus-trial demand has triggered the devel-
opment of techniques to improve user inter-action with recommenders (in terms of moreuser-friendly and effective dialogue struc-tures), the quality of result sets presented tousers (in terms of precision/recall), and usertrust in presented recommendations. The arti-cles presented in this special issue addressthese techniques, and the field presents a widearray of open research questions.
AcknowledgmentsWe thank all the authors for their contributions
to this special issue. Furthermore, we thank thereviewers for their valuable comments, which sig-nificantly contributed to the accepted articles’highquality. Finally, we express our gratitude to Edi-tor in Chief James Hendler and the IEEE Intelli-gent Systems staff, who supported this special issueon recommender systems and served as mentorsand guides for this project.
References
1. J. Konstan et al., “GroupLens: Applying Col-laborative Filtering to Usenet News,” Comm.ACM, vol. 40, no. 3, 1997, pp. 77–87.
2. T. Hofmann, “Latent Semantic Models forCollaborative Filtering,” ACM Trans. Infor-mation Systems, vol. 22, no. 1, 2004, pp.89–115.
3. M. Pazzani and D. Billsus, “Learning andRevising User Profiles: The Identification ofInteresting Web Sites,” Machine Learning,vol. 27, no. 3, 1997, pp. 313–331.
4. L. Schmidt-Thieme, “Compound Classifica-tion Models for Recommender Systems,”Proc. IEEE Int’l Conf. Data Mining (ICDM05), IEEE Press, 2005, pp. 387–385.
5. K. Tso and L. Schmidt-Thieme, “Evaluationof Attribute-Aware Recommender SystemAlgorithms on Data with Varying Character-istics,” Proc. 10th Pacific-Asia Conf. Knowl-edge Discovery and Data Mining (PAKDD 06),Springer, 2006, pp. 831–840.
6. R. Burke, “Knowledge-Based RecommenderSystems,” Encyclopedia of Library and Infor-
G u e s t E d i t o r s ’ I n t r o d u c t i o n
20 www.computer.org/intelligent IEEE INTELLIGENT SYSTEMS
mation Systems, vol. 69, supplement 32, A.Kent, ed., Marcel Dekker, 2000, pp. 180–200.
7. A. Felfernig et al., “An Environment for theDevelopment of Knowledge-Based Recom-mender Applications,” Int’l J. ElectronicCommerce, vol. 11, no. 2, 2006, pp. 11–34.
8. A. Felfernig and B. Gula, “An EmpiricalStudy on Consumer Behavior in the Interac-tion with Knowledge-Based RecommenderApplications,” Proc. IEEE Joint Conf. E-Commerce Technology and Enterprise Com-puting, E-Commerce and E-Services (CEC06/EEE 06), IEEE Press, 2006, pp. 288–296.
9. B. Pine and S. Davis, Mass Customization:The New Frontier in Business Competition,Harvard Business School Press, 1999.
10. C. Huffman and B. Kahn, “Variety for Sale:Mass Customization or Mass Confusion,” J.Retailing, vol. 74, no. 4, 1998, pp. 491–513.
11. J. Payne, J. Bettman, and E. Johnson, TheAdaptive Decision Maker, Cambridge Univ.Press, 1993.
For more information on this or any other com-puting topic, please visit our Digital Library atwww.computer.org/publications/dlib.
MAY/JUNE 2007 www.computer.org/intelligent 21
T h e A u t h o r sAlexander Felfernig is an associate professor at the University Klagen-furt’s Institute of Applied Informatics. He’s also a member of the univer-sity’s Intelligent Systems and Business Informatics research group andcofounder and CEO of ConfigWorks. He received his PhD in computer sci-ence from the University of Klagenfurt. His research interests include knowl-edge-based configuration, knowledge-based recommenders in e-commerce,psychological aspects of consumer buying behavior, model-based diagnosiswith a special focus on knowledge acquisition, and approaches to test automa-tion in knowledge-based-systems development. He’s a member of the IEEEComputer Society. Contact him at Alpen-Adria Universität Klagenfurt, Uni-
versitätsstrasse 65, 9020 Klagenfurt, Austria; [email protected].
Gerhard Friedrich is a professor at the University of Klagenfurt, where hedirects the research group on Intelligent Systems and Business Informatics.His research interests include knowledge-based systems for configuringproducts and services, knowledge-based advisory systems, distributed knowl-edge-based systems, knowledge acquisition, integration of knowledge-basedsystems, diagnosis and repair of processes, and self-healing Web services.He received his PhD in computer science from the Technical UniversityVienna. He was previously head of the technology center for configuratorsand diagnosis systems of Siemens Austria, where under his responsibility
numerous knowledge-based systems were successfully implemented. He’s an editor of AI Commu-nications and is associate editor of the International Journal of Mass Customisation. Contact him atAlpen-Adria Universität Klagenfurt, Universitätsstrasse 65, 9020 Klagenfurt, Austria; [email protected].
Lars Schmidt-Thieme is a professor at the Institute for Economics andInformation Systems and the Institute of Computer Science at the Universityof Hildesheim, where he directs the Information Systems and Machine Learn-ing Lab. His research interests include machine learning and data mining,especially classification and pattern mining for complex data. He receivedhis PhD in economics from the University of Karlsruhe. He’s involved inseveral research projects, such as the EC FP6 integrated project X-Media.Contact him at the Univ. of Hildesheim, Marienburger Platz 22, D-31141Hildesheim, Germany; [email protected].
Tomorrow's PCs, handhelds, and Internet will use technology
that exploits current research in artificial intelligence. Breakthroughs
in areas such as intelligent agents, the Semantic Web, data mining,
and natural language processing will revolutionize your work and
leisure activities. Read about this research as it happens in
IEEE Intelligent Systems.
See the Future ofComputingNow
in IEEEIntelligent Systems
IEE
E
SUBSCRIBE NOW! http://computer.org/intelligent
