Knowledge-Augmented News Recommendation via LLM Recall, Temporal GNN Encoding, and Multi-Task Ranking

Personalized news recommendation is difficult because news content does not last long, user interests change quickly, and the text is often short. To solve these problems, we present HKNR (Hybrid Knowledge-Augmented News Recommender), a single recommendation system that combines three parts: choosing candidates using LLaMA-2-7B embeddings, modeling users with graphs, and ranking with added knowledge. HKNR uses large language model embeddings to find articles with similar meanings, graph convolutional networks to learn how users behave over time, and a multi-layer ranking network that mixes article meanings with outside information like entities and topics. The system is trained using a loss that includes classification, ranking, and reconstruction. To make the system more general, we also use methods like dynamic negative sampling, embedding dropout, and gradient clipping. Tests show that HKNR works well on many evaluation measures and performs better than other common methods.