Self-supervised Dual Graph Learning for Recommendation

Collaborative filtering (CF) is one of the common approaches for recommendation. Recently, graph-based methods, which use the holistic bipartite graph structure to learn user preferences for items, have emerged as effective techniques for CF. Nevertheless, existing graph-based recommenders primarily consider entity connections (i.e., user-user, item-item, and user-item relations) but ignore semantic associations among interaction behaviors (e.g., relationships between different interactions of the same user or item). Because the modelings of entities and interaction behaviors are crucial components in CF, we believe that jointly learning discriminative relationships of these two parts achieves performance gain. In this work, we develop dual graph learning modules, which model both entity-level and behavior-level relations for better learning of the entities and interactions. One entity-centric module adopts graph convolutions on the user-item bipartite graph to capture connections between entities. The other behavior-centric module introduces the user-item line graph (where nodes represent interaction behaviors) and then creates aline graph neural network on this graph to distill semantic associations of interaction behaviors. In addition, self-supervised learning is utilized to model cooperative signals between the two graph learning modules, complementing the representation learning capabilities of each module. We name our method self-supervised dual graph learning (SDGL). Experiment results on the seven real-world datasets indicate the superiority of our SDGL over the state-of-theart baselines. Specifically, SDGL achieves a performance improvement of 1.30% similar to 7.99% in the Precision@10 metric compared to the best baseline models.