GNN-SSER: Graph neural networks with self-attention and session enhanced representation for session-based recommendation

Session-based recommendation (SBR) is a technically demanding task that aims to recommend items based on sequences of anonymous behaviors. Graph Neural Networks (GNNs) have demonstrated significant potential in SBR through constructing graphs from behavioral sequences. GNN-based recommendation models mainly involve message recursively passing along item-item interaction edges to refine encoded embeddings. While GNN-based methods have shown effectiveness, they face challenges such as constrained receptive fields and noisy, irrelevant connections. Transformer-based models are superior in adaptively and globally aggregating information. In addition, it is also critically important for SBR methods to leverage information beyond the current session. Navigating these challenges, we propose a novel model for SBR called Graph Neural Networks with Self-Attention and Session Enhanced Representation (GNN-SSER). GNN-SSER alternates between Transformer and GNN layers to enhance their capabilities mutually. In particular, GNN-SSER utilizes Transformer layers to expand the receptive field, allowing aggregation of information from more pertinent nodes, thus improving the message passing in GNNs. In addition, we propose a session enhanced channel (SEC) to utilize the inter-session information. Comprehensive experiments show that GNN-SSER delivers better performance than other state-of-the-art models.