MnRec: A News Recommendation Fusion Model Combining Multi-granularity Information

Abstract

Personalized news recommendations can alleviate information overload. Most current representation matching-based news recommendation methods learn user interest representations from users’ behavior to match candidate news to perform recommendations. These methods do not consider candidate news during user modeling. The learned user interests are matched with candidate news in the last step, weakening the fine-grained matching signals (word-level relationships) between users and candidate news. Recent research has attempted to address this issue by modeling fine-grained interaction matching between candidate news and each news article viewed by the user. Although interaction-based news recommendation methods can better grasp the semantic focus in the news and focus on word-level behavioral interactions, they may not be able to abstract high-level user interest representations from the news users browse. Therefore, it is a worthwhile problem to make full use of the above two architectures effectively so that the model can discover richly detailed cues of user interests from fine-grained behavioral interactions and the abstraction of high-level user interests representations from the news that users browse. To address this issue, we propose MnRec, a framework for fusing multigranularity information for news recommendation. The model integrates the two matching methods via interactive attention and representation attention. In addition, we design a granularity network module to extract news multigranularity information. We also design an RTCN module to implement multilevel interest modeling of users. Extensive experiments on the real news dataset MIND verify the method’s validity.