Generating Conversational Recommendation Datasets from Real User Interactions

Modern recommendation systems typically follow two complementary paradigms: collaborative filtering, which models long-term user preferences from historical interactions, and conversational recommendation systems (CRS), which interact with users in natural language to uncover immediate needs. While CRS models often do not have access to collaborative signals, resulting in generic suggestions, traditional recommenders lack mechanisms to elicit users’ immediate needs interactively. Unifying these paradigms promises richer personalization but remains challenging due to the lack of conversational datasets grounded in real user behavior. We present ConvRecStudio, a novel framework that leverages large language models (LLMs) to automatically simulate realistic, multi-turn dialogs grounded in real timestamped user-item interactions and user reviews. ConvRecStudio follows a three-stage pipeline: (1) Temporal Profiling, which constructs user profiles and community-level item sentiment trajectories across fine-grained aspects; (2) Semantic Dialog Planning, which generates a structured plan over dialog acts; and (3) Multi-Turn Simulation, which instantiates the plan using paired LLM agents for the user and system, constrained by executional and behavioral fidelity checks. We apply ConvRecStudio to three domains, MobileRec, Yelp, and Amazon Electronics, producing over 12k multi-turn dialogs for each dataset. Human and automatic evaluations confirm the naturalness, coherence, and behavioral grounding of the generated conversations. To demonstrate utility, we build a cross-attention-based transformer model that jointly encodes user history and dialog context, achieving significant gains in Hit@K and NDCG@K metrics over baselines relying on either signal alone or naive fusion. These results establish ConvRecStudio as a scalable and effective foundation for unified conversational recommendations.