A Hierarchical Knowledge and Interest Propagation Network for Recommender Systems

To address the data sparsity and cold start issues of collaborative filtering, side information, such as social network, knowledge graph, is introduced to recommender systems. Knowledge graph, as a sort of auxiliary and structural data, is full of semantic and logical connections among entities in the world. In this paper, we propose a Hierarchical Knowledge and Interest Propagation Network(HKIPN) for recommendation, where a new heterogeneous propagation method is presented. Specifically, HKIPN propagates knowledge and user interest simultaneously in a unified graph combined by user-item bipartite interaction graph and knowledge graph. During the propagation, a hierarchical method is devised to aggregate a node's high-order neighbors explicitly and concurrently. Besides, an attention mechanism is employed to discriminate the importance of neighbors. Furthermore, due to information decay in the process of propagation, the decay factor, as the weight of each hierarchical representation to compose the final user-and-item representations, is taken into account. We apply the proposed model to three benchmark datasets about movie, book, and music recommendation and compare it with state-of-the-art baselines. The experiment results and further studies demonstrate that our approach outperforms compelling recommender baselines.