First-Order Probabilistic Model for Hybrid Recommendations

In this paper, we address the task of inferring user preference relationships about various objects in order to generate relevant recommendations. The majority of the traditional approaches to the problem assume a flat representation of the data, and focus on a single dyadic relationship between the objects. We present a richer theoretical model for making recommendations that allows us to reason about many different relations at the same time. The model is based on Markov logic, which is a simple and powerful language that combines first-order logic and probabilistic graphical models. We apply a hybrid, content-collaborative merging scheme through feature combination. We experimentally verify the efficacy of our theoretical model, and show that our method outperforms state-of-the-art recommendation approaches.