Retrieval-Augmented Generation for News Recommendation

Large language models (LLMs) have advanced rapidly in natural language processing, paving the way for their application in news recommendation systems. In real-world scenarios, the models usually need training on historical data to provide accurate recommendations. However, the log data for training contains noise data and exposure bias, thereby limiting the understanding of user preferences. In this paper, we propose Retrieval-Augmented Generation for News Recommendation (RAG4NR), a prompting procedure involving RAG to promote understanding for user interests and news contents. Our proposed method uses augmented user demographic generation and item selection to construct the user profile, then applies the RAG prompting templates to obtain recommended item lists in a zero-shot manner, eliminating the need for task-specific fine-tuning. Experiments on the MIND dataset demonstrate that RAG4NR outperforms existing zero-shot LLM-based rankers in news recommendation, showcasing its effectiveness.