Bayesian Recommender Systems: Models and Algorithms

This thesis is about how Bayesian methods can be applied to explicitly model and efficiently reason about uncertainty to make optimal recommendations. We are interested in three dimensions of recommender systems: (1) preference elicitation, (2) set-based recommendations, and (3) matchmaking. The first dimension concerns how one can minimize the elicitation efforts in learning a user’s utility function to propose the maximal utility recommendation. The second dimension concerns set-based recommendations, and the problem of how one can optimize the relevance of the recommended set, with respect to uncertainty over the relevance of each item in the set. The third dimension concerns the problem of matchmaking, which is the process of pairing competitors based on similar latent skill levels, given match outcomes, e.g., score, or win/lose/draw. All three dimensions face an inherent problem of handling uncertainty : user utility uncertainty for preference elicitation, query topic uncertainty for set-based retrieval in the context of document retrieval, and skill uncertainty in matchmaking. Bayesian approaches prove to be extremely flexible in modeling various problems, and are robust to risk. However, it is not until recently that efficient Bayesian inference techniques have been introduced for complex models. Thus, we utilize recent advances in Bayesian approaches for addressing these three problems. Our contributions in the thesis are twofold. First, we present compact Bayesian graphical models for dimensions (1)–(3). Second, for each dimension, we make use of advanced Bayesian inference techniques to learn and make optimal recommendations.