Title of Thesis: Hierarchical Probabilistic Relational Models for Recommender Systems Hierarchical Probabilistic Relational Models for Recommender Systems

In this thesis we describe an approach to the recommender system problem based on the Probabilistic Relational Model framework. Traditionally, recommender systems have fallen into two broad categories: content-based and collaborative-filteringbased recommender systems, each of which has a distinct set of strengths and weaknesses. We present a sound statistical framework for integrating both of the above approaches, which allows the strengths of one system to help mitigate the weaknesses of the other. To accomplish this goal we apply the framework of Probabilistic Relational Models (PRMs) to the recommender system problem domain. We begin by applying standard PRMs (sPRMs) to the EachMovie recommender system dataset, which uncovers several severe limitations of the sPRM framework. We then apply an extension of PRMs called Hierarchical PRMs (hPRMs) to the recommender problem, which from a theoretical perspective should address several of the limitations of sPRMs. We show through empirical results that hPRMs do, in fact, achieve superior results on the EachMovie dataset.