Knowledge-aware user multi-interest modeling method for news recommendation

News recommendation can filter massive news and provide personalized information services, and the key is to better portray news and users. Current news recommendation methods try to integrate the knowledge graph to introduce background knowledge for news features modeling and have achieved good performance. However, when modeling users’ interest features, these methods only consider the personalized information of the interacted news and ignore the fact that some typical entities can help depict multiple topics that users are interested in. In this paper, we propose a method called KMRec (knowledge-aware user multi-interest modeling method for news recommendation) to realize the fine-grained portrayal of users’ multi-interest based on knowledge graph. Specifically, we represent user multi-interest features through four modules: historical text encoder, individual entity encoder, common entity encoder and user encoder. The historical text encoder is for the text modeling of each news and information aggregation. The individual entity encoder achieves the entity feature modeling of each news and information aggregation based on knowledge graph. The common entity encoder realizes the common entity information extraction of users’ historical interacted news based on knowledge graph. The user encoder integrates the text feature, individual entity feature and common entity feature to comprehensively depict the multi-interest of each user. Offline experimental results on both MIND-small and MIND-large datasets show that leveraging KMRec for users’ interest modeling can effectively improve the performance of news recommendation. The idea of introducing the user multi-interest feature is also verified to be effective by comparative experiments with existing news feature representations.