Relation-Enhanced Multi-Graph Attention Network for Recommendation

Abstract

Knowledge graph captures structured information and relations between a set of entities. Researchers always introduce knowledge graph (KG) into recommender systems for more accurate and explainable recommendation. Recently, many researchers deploy Graph Neural Network (GNN) with knowledge graph in recommender systems. However, they do not consider proper aggregation and ignore the layer limitation of the GNN. To tackle these issues, we propose a novel recommendation framework, named Relation-Enhanced Multiple Graph Attention Network (REMAN for short), which models the heterogeneous and high-order relationships among entities in recommendation. Firstly, we encode user behaviors and item knowledge as a unified relational graph. Then we utilize a relation-specific attention aggregator to aggregate the embeddings of the heterogeneous neighbors. Thirdly, we propose a relation-enhanced user graph in order to make up for the limitations of the GNN layer in recommendation. Finally, we make prediction based on the embeddings we learned in graphs. Extensive experiments on three benchmark datasets demonstrate that our framework significantly outperforms strong recommender methods.