Sequence-Enhanced Graph Neural Networks for Session-Based Recommendation

Session-Based Recommendation (SBR), aiming at predicting user next action in view of anonymous behavior sequences, is a critical task for online services (e.g. e-commerce, streaming media). Recently, Graph Neural Networks (GNN) have achieved significant success in session-based recommendation modeling tasks. However, the graph structure data cannot adequately reflect the sequential information in original sessions, and it is also hard for GNN to capture high-order item relationships. Focusing on the challenges mentioned, we proposed a method called Sequence-enhanced Graph neural networks for Session-Based Recommendation (SG-SBR) to provide an effective solution. First, we introduced positional embedding mechanism to merge sequential information and session representations generated by GNN. Second, we proposed self-attention network to capture both the sequential information and high-order item relationships, which improves the ability to capture sequential information for GNN. The above method not only improves the performance of recommendation, but also can be effortlessly integrated into other session-based recommendation combined with GNN. We conducted extensive experiments on two real datasets, and the results show that SG-SBR outperforms baseline methods consistently.