Collaborative filtering with implicit feedback via learning pairwise preferences over user-groups and item-sets

In this paper, we focus on an important recommendation problem known as one-class collaborative filtering (OCCF) and propose a novel preference assumption to model users' implicit one-class feedback such as "examinations" or "likes" in the studied problem. Specifically, we address the limitations of previous pairwise preference learning works by defining the pairwise relations on user-groups and item-sets in the vertical dimension and in the horizontal dimension, respectively. On the basis of the proposed generic dual pairwise preference assumption, we develop a novel recommendation algorithm, i.e., collaborative filtering with implicit feedback via learning pairwise preferences over user-groups and item-sets (CoFi(+)). The main merit of our CoFi(+) is its capacity for modeling both the horizontal and vertical ranking-oriented preference relations more sufficiently, as well as its generality of absorbing several existing pairwise preference learning algorithms as special cases. We conduct extensive empirical studies on three public datasets and find that our CoFi(+) performs significantly better than the state-of-the-art methods.