Cascading graph convolution contrastive learning networks for multi-behavior recommendation

In most real-world recommendation scenarios, there are multiple types of behaviors that tend to follow a specific chain of behaviors (e.g., view->-cart->collect->purchase). To forecast users’ possible preferences for the target behavior (e.g., purchase), existing multi-behavior recommendations use auxili-ary behaviors (e.g., view, collect, and cart), and they often emphasize the difference between behaviors. However, different sorts of behavior partially reflect the same user preferences, and similarities between them are largely ignored. At the same time, the subsequent behavior in the chain usually reflects stronger user preferences than the previous behavior, and most multi-behavior models cannot capture the interdependence in the chain of behaviors. To tackle these issues, we propose a Cascading Graph Convolution Contrastive Learning Networks (CGCCN) for Multi-Behavior Recommendation. Specifically, We use LightGCN to learn user and item embeddings, and then we combine multi-task learning with contrastive learning to explicitly exploit behavioral dependence in embeddings learning and capture differences between embeddings. We conduct comprehensive experiments on two real-world datasets to validate the efficiency of our model. The results further demonstrate the rationality and effectiveness of the designed CGCCN, where the maximum improvement can reach to 42.98%.