Graph Contrastive Learning for Multi-Behavior Recommender System

Multi-behavior recommender system aims to model user preference representation based on multiple types of user-item interactions (e.g., viewing, adding to favorites, adding to the cart, and purchasing). However, existing works have two limitations in general: 1) Most of them only concern the sparse observed user-item interactions (explicit interaction) and ignore the huge amount of unobserved user-item interactions (implicit interaction), which are incapable of fully capturing user preference in recommender systems. 2) Previous works typically tend to only extract valuable information by distinguishing target and auxiliary behaviors to model user reference representation, and they fail to explore the fine-grained commonality between target and auxiliary behaviors. To tackle these limitations, we propose a new model named Graph Contrastive learning with Multi-Behavior (GCMB) for a multi-behavior recommender system. Specifically, we utilize Randomized Singular Value Decomposition (rSVD) to inject implicit interaction into the model, and then combine explicit interaction and implicit interaction to learn user preference by graph contrastive learning. Furthermore, we consider the multi-level commonality between target and auxiliary behaviors to capture the fine-grained commonality and then model high quality of user preference representation. Extensive experiments on two real-world datasets demonstrate that our method consistently outperforms various state-of-the-art recommender methods.