CATM: Candidate-Aware Temporal Multi-head Self-attention News Recommendation Model

User interests are diverse and change over time. Existing news recommendation models often do not consider the relationship and temporal changes between browsing news when modeling user characteristics. In addition, the wide range of user interests makes it difficult to match candidate news to users’ interests precisely. This paper proposes a news recommendation model based on the candidate-aware time series self-attention mechanism(CATM). The method incorporates candidate news into user modeling based on considering the temporal relationship of news sequences browsed by users, effectively improving news recommendation performance. In addition, to obtain more rich semantic news information, we design a granular network to obtain more fine-grained segment features of news. Finally, we also designed a candidate-aware attention network to build candidate-aware user interest representations further to better match candidate news with user interests. Extensive experiments on the MIND dataset demonstrate that our method can effectively improve news recommendation performance.