Collaborative Retrieval for Large Language Model-based Conversational Recommender Systems

Conversational recommender systems (CRS) aim to provide personalized recommendations via interactive dialogues with users.

While large language models (LLMs) enhance CRS with their superior understanding of context-aware user preferences, they typically struggle to leverage behavioral data, which have proven to be important for classical collaborative filtering (CF)-based approaches.

For this reason, we propose CRAG, Collaborative Retrieval Augmented Generation for LLM-based CRS. To the best of our knowledge, CRAG is the first approach that combines state-of-the-art LLMs with CF for conversational recommendations. In addition, we find that the two-step reflection proposed in CRAG is the key to incorporating context-aware collaborative filtering information into the LLM-based CRSs.