MDCD: Meta-Learning Driven Conditional Diffusion for User Cold-Start in Conversational Recommendation

Abstract

The conversational recommendation system (CRS) breaks the limitations of traditional static methods, presenting a novel framework for personalized recommendations with real-time adaptability and dynamic interactions. Existing methods predominantly focus on the balance between ”exploration and exploitation (E&E)”, but they struggle to explore sufficiently in large-scale candidate sets and rely heavily on extensive historical interaction data to support accurate decision-making. Consequently, these methods often face issues such as instability in recommendations and poor adaptation to new users in cold-start scenarios. To address these challenges, we propose an innovative meta-learning-driven conditional diffusion model to replace E&E. Specifically, we use a generative model to directly model the data distribution and learn personalized user patterns, enhancing recommendation stability. Meanwhile, the meta-learning framework implicitly constrains the optimization process of the diffusion model and other components, enhancing the model's ability to rapidly adapt to new users and dynamic preferences. Extensive experimental results demonstrate that the proposed method outperforms representative baseline methods on several real-world datasets, validating its effectiveness and robustness.