Multi-level News Recommendation via Modeling Candidate Interactions

Due to the information explosion on the Internet, news recommendation, which helps users quickly find the news they are interested in, has become an essential issue for online news services. Previous research work usually adopts collaborative filtering or content-based methods which extract features and measure the similarities between users and each candidate news independently. However, candidate news often competes with each other for user attention, and modeling the interactions of multiple candidate news helps distinguish them better for news recommendation. In this paper, we propose a multi-level news recommendation method via modeling the interactions of multiple candidate news explicitly. Specifically, we design a Candidate Interaction Module (CIM) to generate interaction-enhanced candidate news representations. For each candidate news, the interaction-enhanced news representation contains information from other candidate news displayed to the user at the same time. Furthermore, in order to identify the connections between candidate news and user preferences at different semantic levels, we add a Multi-level Prediction Module (MPM) to exploit the category and subcategory information of news. Experimental results demonstrate that our proposed model achieves the state-of-the-art performance on two real-world benchmark datasets.