Asymmetric Pairwise Preference Learning for Heterogeneous One-Class Collaborative Filtering

Yongxin Ni, Zhuoxin Zhan, Weike Pan, Zhong Ming

Abstract: Heterogeneous one-class collaborative filtering (HOCCF) is a recent and important recommendation problem which involves two different types of one-class feedback such as purchases and examinations. In this paper, we propose a generic asymmetric pairwise preference assumption and a novel like-minded user-group construction strategy for the HOCCF problem. Specifically, our generic assumption contains six different pairwise preference relations derived from the heterogeneous feedback, where we introduce a series of weighting strategies to make our assumption more reasonable. Our group construction strategy introduces richer interactions within user-groups, which is expected to learn the users’ preference more accurately. We then design a novel recommendation model called asymmetric pairwise preference learning (APPLE). Extensive empirical studies show that our APPLE can recommend items significantly more accurately than the closely related state-of-the-art methods on three real-world datasets.