Dual Channel Hybrid Messaging-Passing Graph Convolutional Network for Recommendation

Abstract

Graph convolution network (GCN) is widely used in recommendation. It has become the mainstream technology of collaborative filtering as its powerful graph representation learning ability. Since historical interactive data can be naturally modeled as a user-item bipartite graph, many excellent methods have been proposed for recommendation based on bipartite graphs. These methods have achieved impressive successes. However, most of these methods only utilize the intuitive topology information of bipartite graph and ignore semantic information. In this paper, we propose a new Dual Channel Hybrid Messaging-Passing Graph Convolutional Network(DCH-GCN) recommendation model. In our model, first the user-item bipartite graph and item-item co-occurrence graph is constructed based on historical interaction data. Then the explicit messaging passing channel is constructed for topology information by modeling user-item explicit interactions. In addition, the implicit message-passing channel is constructed for semantic information by modeling item-item implicit interactions. Finally the loss of the predictive user preference is co-trained and co-optimized. It is more interpretable that we add semantic information into the model in bipartite graph. We perform a comprehensive experiment on two public benchmark datasets to verify the effectiveness of our model. The experimental results show that our method can better learn user and item embedding by using dual channels, and achieve significant and consistent improvements over competitive baselines.