基于双增强图神经网络的会话推荐

In session-based recommendation, graph-based models primarily rely on item transition patterns to capture item similarity and dependency relationships. However, existing methods still have limitations. First, current graph structures update item embeddings solely based on item co-occurrence relationships, lacking user-centered modeling and failing to explicitly incorporate user-oriented information. Second, traditional graph structures typically model only the transitions between adjacent items, making it difficult to capture long-range dependencies within sessions. To address these issues, this paper proposes a session-based recommendation algorithm based on Dual-Augmented Graph Neural Networks (DAGNN). DAGNN consists of a user-item interaction graph and a hyper-session connection graph. The user-item interaction graph is constructed based on item transitions, with users as central nodes directly connected to multiple items. This allows GNN propagation to incorporate user-specific information, influencing all relevant items. Meanwhile, the hyper-session connection graph shortens the relative paths between different nodes, enabling efficient propagation of long-range dependencies and enhancing the overall modeling capability. Experiments on three public benchmark datasets demonstrate the effectiveness of the proposed model.