Do Not Read the Same News! Enhancing Diversity and Personalization of News Recommendation

Abstract

Personalized news recommendation by machine is one of the widely studied areas. As the production of news articles increases and topics are diversified, it is impractical to read all the articles available to users. Therefore, the purpose of the news recommendation system should be to provide relevant news based on the user's interest. Unlike other recommendation systems, explicit feedback from users on each item such as ratings is rarely provided in news recommendation systems. Most news recommendation systems use implicit feedback such as click histories to profile user interest, which leads to biased recommendation results towards generally popular articles. In this paper, we suggest a novel news recommendation model for more personalized recommendations. If a user reads news not widely clicked by others, the news reflects the user's personal interest rather than other popular news clicked. We implement two user encoders, one to encode the general interest of the set of users and another one to encode the user's individual interest. We also propose regularization methods that induce two encoders to encode different types of user interest. The experiment on real-world data shows that our proposed method improves the diversity and the quality of recommendations for different click histories without any significant performance drops.