A Large Language Model-Based News Recommendation Method Integrating User Hierarchical Interest Representation

Large language models (LLMs) show great application potential in personalized news recommendation. Current LLM-based recommendation methods rarely explore hierarchical interests from user behaviors, leading to inadequate descriptions of user interests. This paper proposes a LLMs-based news recommendation method that integrates hierarchical user interest representation (LNR-HI). It leverages LLMs to finely depict users' browsing interests in news from aspects such as themes, key topics, and interest degrees. Meanwhile, it adopts hierarchical theme summarization to track changes in user interests, ultimately achieving end-to-end news recommendation. Empirical studies are conducted using real-world news recommendation data. The experimental results demonstrate that our method integrating multi-granularity user interest representation can accurately identify users' core interests and broad interests, making it easier to meet personalized information needs.