Pairwise probabilistic matrix factorization for implicit feedback collaborative filtering

Abstract

Collaborative filtering (CF) has been widely applied to improve the performance of recommendation systems. With the motivation of the Netflix Prize, researchers have proposed a series of CF algorithms for rating datasets, such as the 1 to 5 rating on Netflix. In this paper, we investigate the problem about implicit user feedback, which is a more common scenario (e.g. purchase history, click-through log, and page visitation). In these problems, the training data are only binary, reflecting the user's action or inaction. Under these circumstances, generating a personalized ranking list for every user is a more challenging task since we have less prior information. We consider it as a ranking problem: collaborative ranking (CR) skips the intermediate rating prediction step, and creates the ranked list directly. In order to solve the ranking problem, we propose a new model named pairwise probabilistic matrix factorization (PPMF), which takes a pairwise ranking approach integrated with the popular probabilistic matrix factorization (PMF) model to learn the relative preference for items. Experiments on benchmark datasets show that our proposed PPMF model outperforms the state-of-the-art implicit feedback collaborative ranking models by using different evaluation metrics.