k-CoFi: Modeling k-Granularity Preference Context in Collaborative Filtering

Abstract

Collaborative filtering (CF) is a highly applicable technology for predicting a user’s rating to a certain item. Recently, some works have gradually switched from modeling users’ rating behaviors alone to modeling both users’ behaviors and preference context beneath rating behaviors such as the set of other items rated by user u. In this paper, we go one step beyond and propose a novel perspective, i.e., k-granularity preference context, which is able to absorb existing preference context as special cases. Based on this new perspective, we further develop a novel and a generic recommendation method called k-CoFi that models k-granularity preference context in collaborative filtering in a principled way. Empirically, we study the effectiveness of factorization with coarse granularity, fine granularity and smooth granularity, and their complementarity, by applying k-CoFi to three real-world datasets. We also obtain some interesting and promising results and useful guidance for practitioners from the experiments.