Recommender systems [41] constitute one of the fastest growing segments of the Internet economy today. They help reduce information overload and provide customized information access for targeted domains. Building and deploying recommender systems has matured into a fertile business activity, with benefits in retaining customers and enhancing revenues. Elements of the recommender landscape include customized search engines, handcrafted content indices, personalized shopping agents on e-commerce sites, and news-on-demand services. The scope of such personalization thus extends to many di erent forms of information content and delivery, not just web pages. The underlying algorithms and techniques, in turn, range from simple keyword matching of consumer profiles, collaborative filtering, to more sophisticated forms of data mining, such as clustering web server logs. Recommendation is often viewed as a system involving two modes (typically people and artifacts, such as movies and books) and has been studied in domains that focus on harnessing online information resources, information aggregation, social schemes for decision making, and user interfaces. A recurring theme among many of these applications is that recommendation is implicitly cast as a task of learning mappings (from people to recommended artifacts, for example) or of filling in entries to missing cells in a matrix (of consumer preferences, for example). Consequently, recommendation algorithms are evaluated by the accuracies of their predicted ratings. We approach recommendation from a di erent but complementary perspective of considering the connections that are made.
