Graph Convolutional Matrix Completion via Relation Reconstruction

Abstract

To alleviate sparsity and improve recommender systems performance, it is necessary to go beyond modeling user-item interactions and take auxiliary information into account. Besides user-item interactions, auxiliary information can be used to build relation graphs. Recently, Graph Convolution Networks (GCNs), which can integrate content information and structural information of nodes, have been demonstrated to be powerful in learning on graph data and applied in recommendation systems. However, existing approaches do not consider multiple types of relations between nodes and high-order structural information. In this paper, we propose a new model called Graph Convolutional Matrix Completion via relation reconstruction (RE-GCMC) to capture structural information and relations between nodes in the graph. We construct user-user, item-item, and user-item relation graphs by evaluating the feature similarity of the nodes. Then, we introduce the Graph Convolutional Networks (GCNs) and self-attention mechanism to be applied in the graphs to refine feature embeddings. We apply the proposed model to four datasets and experimental results demonstrate that our approach outperforms state-of-the-art recommender baselines.