SEDGN: Sequence enhanced denoising graph neural network for session-based recommendation

Abstract

Session-based recommendation, which aims at predicting subsequent user actions based on anonymous sessions, plays a significant role in many online services. Based on the graph neural network (GNN), existing methods construct each session as a graph and capture the rich transition relationship between items to generate item representations. However, they neglect natural noise in a session, which is not conducive to making accurate recommendations. In addition, GNN-based methods cannot fully express sequential information of the session, such as the repeated nodes in a session and the starting node of the directed graph. Therefore, we propose a novel method, i.e., sequence enhanced denoising graph neural network (SEDGN). Specifically, we use a gated recurrent unit (GRU) to enhance the graph neural network, thereby addressing the GNN’s deficiency in modeling the sequential information of sessions. In addition, two denoise modules are designed to alleviate the impacts of natural noise by more accurate item representations from the perspective of sequence-structure and graph-structured data. Extensive experiments on five real-world datasets revealed that SEDGN outperforms the most the existed state of art methods and demonstrates that the components in the model are effective.