A Tag-Based Post-Hoc Framework for Explainable Conversational Recommendation

Explanations are important for conversational recommendation. They help users to understand how the recommender system works, and elicit user's responses by allowing users to provide informative feedback on them. During the interaction with users, explainable conversational recommender system provides explanations and collects user feedback to further refine the recommendations. Current conversational recommender systems, however, usually make recommendations with black-box prediction models, bringing difficulty in model explainability. Moreover, existing methods for explainable recommendation commonly provide one-shot explanations and fail to leverage user feedback. In this paper, we propose a tag-based post-hoc framework for explainable conversational recommendation (TPECR), which enables black-box recommendation models to provide explanations and refine recommendations based on user preference on tags (e.g., item attributes). Specifically, given the recommendation model being explained, TPECR trains a generation model to construct user embeddings based on their tag preferences. The explanations are provided by utilizing the generation model to estimate the contributions of different tags with respect to each item prediction. Given user feedback, the recommendation at the next turn is refined by tuning the tag preference and generating modified user embedding with the generation model. We instantiate the generation model with conditional variational auto-encoder (CVAE), which reconstructs user embedding conditioned on his tag preference. We conducted experiments by applying TPECR to different models and the results demonstrated the effectiveness of our TPECR on both synthetic and real datasets.