Multi-stage Knowledge Propagation Network for Recommendation

Abstract

As the knowledge graph can provide items with rich attributes, it has become an important way to alleviate cold start and sparsity in recommender systems. Recently, some knowledge graph based collaborative filtering methods use graph convolution networks to aggregate information from each item’s neighbors to capture the semantic relatedness, and significantly outperform the state-of-the-art methods. However, in the process of knowledge graph convolution, only the item nodes can make use of knowledge, while the user nodes only contain the original ID information. This gap in information modeling makes it difficult for prediction function to capture the user preference for high-order attribute nodes in knowledge graph, which leads to the introduction of noise data. In order to give full play to the ability of knowledge graph convolution in mining high-order knowledge, we propose Multi-Stage Knowledge Propagation Networks (MSKPN), an end-to-end recommender framework which combines the graph convolution on both knowledge graph and user-item graph. It uses the collaborative signal latent in user-item interactions to build an information propagation channel between the user nodes and item nodes, so as to complement user representations. We conduct extensive experiments on two public datasets, demonstrating that our MSKPN model significantly outperforms other state-of-the-art models. Further analyses are provided to verify the rationality of our model.