Knowledge Graph Attention Network Enhanced Sequential Recommendation

Abstract

Knowledge graph (KG) has recently been proved effective and attracted a lot of attentions in sequential recommender systems. However, the relations between the attributes of different entities in KG, which could be utilized to improve the performance, remain largely unexploited. In this paper, we propose an end-to-end Knowledge Graph attention network enhanced Sequential Recommendation (KGSR) framework to capture the context-dependency of sequence items and the semantic information of items in KG by explicitly exploiting high-order relations between entities. Specifically, our method first combines the user-item bipartite graph and the KG into a unified graph and encodes all nodes of the unified graph into vector representations with TransR. Then, a graph attention network recursively propagates the information of neighbor nodes to refine the embedding of nodes and distinguishes the importance of neighbors with an attention mechanism. Finally, we apply recurrent neural network to capture the user’s dynamic preferences by encoding user-interactive sequence items that contain rich auxiliary semantic information. Experimental results on two datasets demonstrate that KGSR outperforms the state-of-the-art sequential recommendation methods.