Reproduce “Bridging Conversational and Collaborative Signals for Conversational Recommendation” on OpenResearch

About this paper

Ahmad Bin Rabiah, Nafis Sadeq, Julian McAuley

The BridgeCRS framework integrates collaborative filtering signals with conversational context using a newly constructed Reddit-ML32M dataset to enhance conversational recommender systems. This approach consistently improved recommendation accuracy and ranking quality across various metrics compared to existing methods.

Key insights

• Combining LLM-driven conversational understanding with collaborative filtering (CF) signals yields a synergistic effect, where the strengths of each modality compensate for the other's weaknesses, leading to enhanced recommendation accuracy.
• The creation of densely linked datasets from disparate sources (conversational and collaborative) is essential for learning robust item representations and mitigating data sparsity in complex recommender systems.
• A multi-stage framework, leveraging an LLM for initial context-aware generation and then refining those recommendations with pre-trained collaborative item embeddings, provides an effective architectural blueprint for hybrid conversational recommenders.

Summary by alphaXiv.