Fusing sequential recommender and ad-hoc planner for multi-faceted preference understanding in conversational recommendation

Recent advances in Large Language Models (LLMs) have accelerated the development of Conversational Recommender Systems (CRS). However, existing CRS approaches face two critical challenges: limited incorporation of historical user interactions and unrealistic experimental settings that fail to reflect real-world scenarios. To address these challenges, we propose FuseRec, a novel framework that enables any Sequential Recommender System (SRS) to function as a conversational recommender through integration with an ad-hoc Planner. The integration leverages the inherent strength of SRS in processing long-term historical interactions while the Planner learns conversation strategies. Additionally, a CRS module refines recommendations by incorporating conversational context, effectively fusing sequential patterns with real-time dialogue insights. To create realistic evaluation environments, we implement an advanced GenAI-based user simulator with stratified personas reflecting varying degrees of preference awareness, from users with clear preferences to those with abstract, uncertain preferences. To handle multi-faceted user behaviors, the Planner employs four sophisticated actions: Chitchat, semantic questioning (Semantic Q), attribute questioning (Attribute Q), and Recommend, dynamically adjusting based on user response patterns. We train the Planner through reinforcement learning with curriculum strategy based on user difficulty levels. Through extensive experiments, we demonstrate that FuseRec significantly outperforms existing approaches in recommendation accuracy while showing remarkable adaptability across different user types and recommendation scenarios.