Using Large Language Models as user Interests Interpretation for Solving Cold-Start Item Recommendation

Recommender systems are essential for enhancing user experiences on digital platforms, but they often struggle with suggesting items lacking interaction data, known as the cold-start problem. This study introduces “LLM-search,” a novel method combining language models (PaLM -2) with data augmentation techniques to address this challenge. LLM -search analyzes users' interactions to identify an item most likely to be the next purchase. It then employs the BM25 search engine to find relevant cold items based on the items identified, which effectively creates synthetic cold item interactions for training the recommender model. The effectiveness of this method is demonstrated through experiments on the 2018 Amazon review dataset in the sports and outdoor category. The method improved recommendations for cold items at various recall values. However, there was a slight decrease in performance for warm items in some cases. This study demonstrates that the LLM -search method can enhance cold item recommendation recall, offering a novel solution to the persistent cold-start problem in recommender systems.