Collaborative filtering on the Internet

Collaborative filtering is a recent software technology that provides personalized recommendation. It is used for recommender systems and in particular for the personalization of Web sites. Users indicate their preferences (i.e. by rating objects) which the collaborative filtering system uses to match together users with similar interests and then predict unknown preferences. This thesis focuses on improving collaborative filtering algorithms and their application in Web sites. The lack of preference information (a.k.a. sparsity) may lead to failure of collaborative filtering algorithms. We propose a new collaborative filtering algorithm based on hierarchical clustering, which is designed to better address situations with few preference information. For a better understanding of Web based applications which apply collaborative filtering for personalization (user-adapted Web sites), we implemented a prototype user-adapted Web site: The Active WebMuseum is a public online Web museum for art paintings. In order to improve collaborative filtering with the help of weak content-based information which can automatically be indexed, such as color or texture of paintings, we propose two variations for the combination of collaborative filtering with the conceptually different content-based filtering. When collaborative filtering is applied in an on-line application, such as the Active WebMuseum it is important to define how the predictions are used to personalize the application and how the personalization is valued by the user. We propose the Multi-Corridor-Access-Paradigm as a model for formalizing the user interaction with a user-adapted Web site and therefore allowing the derivation of performance metrics related to the personalization performance. To improve the training phase when new users start using a collaborative filtering system, we propose several approaches for selecting objects for which preferences should be queried from new users.