Agent4ExpRec: Towards Generating and Evaluating User-Centric Recommendation Explanation
Abstract
AbstractExplainable Recommendation (ER) aims to improve the transparency of Recommender Systems (RS) by generating natural language explanations. However, existing methods primarily optimize textual similarity to reference explanations, which often leads to item-centric explanations focusing on product attributes rather than users’ personalized preferences. Moreover, they usually do not care how the generated explanations influence the behavior of users. To address these limitations, we introduce Agent4ExpRec, a closed-loop user-centric pipeline designed for generating and evaluating recommendation explanation. Based on the original explanations, it enables the peer agent to analyze the user’s requirements from multiple aspects and provides feedback, and automatically plans the refinement process to generate explanations focusing on user preferences. During the evaluation stage, we analyze how different explanations influence user behavior, providing evidence that our generated explanations are more consistent with users’ personalized preferences. Since the actual user behavior on large-scale explanation text is difficult to obtain in practice, we use LLM-powered agents equipped with user profile to simulate user interactions and decision-making in ER scenarios. Extensive experiments demonstrate that the generated explanations consistently align with user preferences, highlighting the potential of user-centric refinement for explainable recommendation tasks. Our code is available at https://github.com/yy-Hou/Agent4ExpRec.