MGSF: Towards Multi-Graphs Semantic Fusion for Multi-behavior Recommendation.

In real life, users’ behavior preferences reflect their interests. The purpose of the multi-behavior recommendation is to model users’ preferences based on their multiple behavior characteristics and make accurate recommendations. However, existing approaches exploit weights to reflect the significance of different behaviors, which are unable to fully extract the semantic information carried by various behaviors of users. In addition, these models employ functions independent of node type on the user and item side, which ignores the inherent differences between nodes. To alleviate these problems, we propose a novel multi-behavior recommendation framework based on GNN, namely MGSF, which is capable of making Multi-Graphs Semantic Fusion under a specific interactive environment. In particular, we devise a heterogeneous message encoding layer that incorporates different behavior semantic information into the node representation. Then, we design the implicit message passing mechanism to deliver high-order supervisory information of the same type. In addition, we utilize separate transformation and aggregation operations to respectively handle two different types of nodes for users and items. Experiments on two real datasets show that our method outperforms the baseline models.