News Recommendation Based On Multi-Feature Sequence Transformer

Personalized news recommendation system aims to screen out the news that users are interested in from explosion amount of news for display. In recent years, deep learning methods have been widely used in news recommendation system. However, whether in traditional news recommendation methods or advanced deep learning models, most of them are only modeled after feature extraction of news titles or modeled after adding user preferences. There are two problems: insufficient expression of news and insufficient exploration of the implicit meaning of users, continuous behavior. Therefore, in this paper, we propose a news recommendation model based on a multi-feature sequence transformer (MFST). It first extracts multiple attributes of news and merges them together for learning unified news representation. Secondly, a powerful Transformer component is applied to process the user's historical reading behavior seq uence information to express the news in more details by strengthening the learning ability of news representation and capturing the meaning behind the user's continuous historical reading behaviors. In addition, we also attached an attention network to calculate the closeness of the clicked news to the candidate news. Experimental results based on the real-world news dataset confirmed that our proposed MFST model is effective for personalized news recommendation compared the state-of-the-art deep learning models.