GRE: A GAT-Based Relation Embedding Model of Knowledge Graph for Recommendation

Compared with collaborative filtering, knowledge graph embedding based recommender systems greatly boost the information retrieval accuracy and solve the limitations of data sparsity and cold start of traditional collaborative filtering. In order to fully explore the relationship and structure information hidden in knowledge graphs, we propose the GAT-based RelationEmbedding (GRE) model. In our model, we propose a Triple Set to denote a set of knowledge graph triples whose head entities are linked by items in interaction records, and a Triple Group to denote a group of knowledge graph triples extracted from Triple Set according to different relations. The proposed GRE is a neural model that aims at enriching user preference representation in recommender systems by utilizing Graph Attention Network (GAT) to aggregate the embeddings of adjacent tail entities to head entity over Triple Group and embedding the representation of relation in the process of polymerization of Triple Groups in Triple Set. By embedding relation information into each Triple Group representation and concatenating Triple Group representations in Triple Set, this proposed novel relation embedding method addresses the problem that GAT-based models only consider aggregating the neighboring entities and ignore the effect of relations in triples. Through extensive experimental comparisons with the baselines, we show that GRE has gained state-of-the-art performance in the majority of the cases on two open-source datasets.