I. Introduction
In online information systems, users interact with a system in a variety of forms. In traditional recommender systems, only user-item interaction data of one behavior type is considered for collaborative filtering [1]. Existing approaches for multi-behavior recommendation can be divided into two categories. The first category is based on collective matrix factorization (CMF) [2]–[4], which extends the matrix factorization (MF) method to jointly factorize multiple behavior matrices. The second category approaches the problem from the perspective of learning [5]–[7]. For example, [5], [7] extends the Bayesian Personalized Ranking (BPR) [8] framework to address multi-behavior recommendation by enriching the training data of relative preference from the multi-behavior data.