Enhancing conversational recommendations through contrastive user preference modeling

Conversational recommender systems (CRSs) aim to deliver items that best align with user preferences through interactive, multi-turn dialogues. While recent studies emphasize the value of incorporating sentiment signals into recommendation, most existing methods rely on a single user representation, which often fails to capture conflicting user intentions such as likes and dislikes, leading to suboptimal outcomes. To address this limitation, we present CUPID (Contrastive User Preference In Dialogue), a novel framework for enhancing conversational recommendations through contrastive user preference modeling. CUPID uncovers both positive and negative signals by expanding user preferences with contrastive reasoning, explicitly disentangling conflicting intentions. These contrastive preferences are then integrated into the recommendation process via preference-aware learning, enabling finer-grained personalization. Extensive experiments on benchmark datasets demonstrate that CUPID consistently outperforms state-of-the-art baselines, achieving substantial improvements in recommendation accuracy and robustness across multiple evaluation metrics. The code is publicly available at https://github.com/AnandMayank/CUPID.