Incorporating Adjacent User Modeling into Session-based Recommendation with Graph Neural Networks

Recently, session-based recommendation system (SRS) based on anonymous sessions has attracted increasing attention for practical applications. Most existing methods are confined by limited behaviors within a single session while ignoring similar interests between different sessions. In this paper, we treat each session as a virtual user and propose an Adjacent and Target information United Graph Transition Network (ATU-GTN) to activate both user-level and session-level modeling for SRS. The central idea is that we construct adjacent user graph and session graph, and then explore two graph topologies to obtain the target session embedding. Specifically, the adjacent user graph is constructed via the co-occurrence of the last click following the phenomenon that the last click is highly correlated to current user interests. User embedding extracted from the adjacent user graph reveals users’ common interests hidden in homonymous or similar users. Current and global embeddings are extracted from the session graph based on within-session items, which represent current and long-term interests, respectively. By concatenating the three embeddings, the final embedding achieves an explicit and comprehensive representation of the target session. After that, the Neural Tensor Network (NTN) is firstly introduced to learn the sufficient interactions between the target session and candidate items for recommendation. The rationality and effectiveness of ATU-GTN are extensively evaluated on two real-world datasets over state-of-the-art baselines.