Knowledge Graph Attention Network with Attribute Significance for Personalized Recommendation

The recommendation system based on the knowledge graph usually introduces attribute information as supplements to improve the accuracy. However, most existing methods usually treat the influence of attribute information as consistent. To alleviate this problem, we propose a personalized recommendation model based on the attribute significance of the knowledge graph attention network (AS-KGAN). Firstly, to obtain user preferences on the knowledge graph and maintain the inherent structure and semantic information, we consider the entity types in the knowledge graph and conduct modeling in an end-to-end approach. Secondly, when introducing item attribute information, we consider the significance of different attributes on items and use the graph attention network to distinguish them. Thirdly, to obtain deeper representations of user preferences, the weights of node neighborhoods are learned during propagation. Finally, to better obtain the high-order relationships and ensure interpretability, we model the high-order connectivity explicitly in the knowledge graph. The model generates top-K recommendations for target users by predicting the probability that users will interact with the item. We applied this model to three public datasets, and the experimental results indicate that the recommendation quality of our model has improved compared with other models.