Interactive Methods for Model-based Collaborative Filtering Recommender Systems

Abstract

Recommender systems have become very popular for reducing the information overload users are often confronted with in today's web. Collaborative filtering is the method of choice for generating personalized recommendations, supporting users in finding items that best match their preferences, from news articles and movies to all kinds of consumer goods and services. Model-based techniques have achieved great success in terms of recommendation accuracy and algorithmic performance. While there is a large body of research on these aspects, only little effort has been spent on improving user control and experience. As a consequence, users of contemporary systems usually have no other option than rating single items to indicate their preferences and thus to influence the recommendations. In this thesis, we propose a set of interactive methods for model-based collaborative filtering recommender systems. With these methods, we aim at providing users richer possibilities to specify their preferences and to control the outcome of the systems according to situational needs. In general, users should be enabled to take a more active role throughout the process of finding suitable items. Guided by a structured model of user interaction, we first present a choice-based preference elicitation method. For systems that rely on matrix factorization, one of the most commonly applied techniques in the area of model-based collaborative filtering, this method provides an alternative to rating items in cold-start situations. Furthermore, we describe an algorithmic enhancement, content-boosted matrix factorization. Based on the additional item-related information that is considered by this method, we give several examples of advanced interactive features that allow users to control the recommendations in an even more expressive manner, also later in the process. Finally, we propose a concept called blended recommending. This concept is designed to merge model-based collaborative filtering with other established methods in a way that users can be supported also in complex scenarios with the full range of options they need to reach their search goal. All these methodological contributions are complemented by empirical evaluations. Overall, we conducted four user experiments with n=35, 46, 54 and 33 participants, respectively. The results underline that our methods can effectively be implemented in existing recommender systems in order to turn them into fully interactive, user-controlled applications. This is finally confirmed with the help of an integrated recommendation platform that demonstrates that all our developments can be combined with each other in a single holistic system.