Distribution-guided Graph Learning for Sequential Recommendation

Sequential recommendations predict the next item by capturing behavioral patterns from a user’s sequence. Graph neural networks (GNN) have recently gained popularity in sequential recommendation for effectively capturing high-order information, which notably improves recommendation performance. However, some existing GNN-based methods represent the node embeddings in the graph as fixed vectors, which fails to capture the uncertainty generated by the transition of relations between nodes. To cope with the above challenge, we propose Distribution-guided Graph Learning for Sequential Recommendation (DGLSR). Specifically, it utilizes the Gaussian distribution (i.e., mean and covariance embeddings) to represent nodes in the user–item bipartite graph, modeling the node uncertainty while preserving the graph structure. Subsequently, we use a graph convolutional network to update user and item node distribution embeddings, and then introduce a personalization distribution embedding fusion operation to integrate them, which generates the final sequence representation. Furthermore, we design a temporal Wasserstein self-attention mechanism. This mechanism utilizes the Wasserstein distance to measure the distributional differences between any two items in the sequence while enhancing the model’s sensitivity to temporal dynamics, thereby improving the accuracy of the next item prediction. Experiments involving four real-world datasets reveal that the DGLSR we developed exceeds the performance of SOTA methods on benchmark metrics.