News Recommendation via Jointly Modeling Event Matching and Style Matching

Abstract

News recommendation is a valuable technology that helps users effectively and efficiently find news articles that interest them. However, most of existing approaches for news recommendation often model users’ preferences by simply mixing all different information from news content together without in-depth analysis on news content. Such a practice often leads to significant information loss and thus impedes the recommendation performance. In practice, two factors which may significantly determine users’ preferences towards news are news event and news style since users tend to read news articles that report events they are interested in, and they also prefer articles that are written in their preferred style. Such two factors are often overlooked by existing approaches. To address this issue, we propose a novel Event and Style Matching (ESM) model for improving the performance of news recommendation. The ESM model first uses an event-style disentangler to extract event and style information from news articles respectively. Then, a novel event matching module and a novel style matching module are designed to match the candidate news with users’ preference from the event perspective and style perspective respectively. Finally, a unified score is calculated by aggregating the event matching score and style matching score for next news recommendation. Extensive experiments on real-world datasets demonstrate the superiority of ESM model and the rationality of our design (The source code and the splitted datasets are publicly available at https://github.com/ZQpengyu/ESM ).