Personalized News Recommendation with CNN and Multi-Head Self-Attention

Abstract

With the globalization of information dissemination and information reception, it is especially important to obtain accurate user and news representations to recommend limited news that matches users' real interests in the infinite richness of news. Existing user representations are usually single and ignore the interrelationship between different components (e.g., titles, categories, and bodies) in the news. In this paper, we propose a personalized news recommendation approach using a convolutional neural network (CNN) and multi-head self-attention. The core of our approach is a user encoder and a news encoder, In the user encoder, we use multi-head self-attention to learn user representations from the news that users have browsed. In the news en-coder, we use convolutional neural networks to obtain local contextual information of words from news components and then use multi-head self-attention to model the interrelationship of words between different components of the news to learn news representations from multi-view news components. In addition, we use attention to select more important words and news components to enrich the user and news representations, and the experimental results show the advancement of our approach on real-world dataset.