Clarifying user Preference with Maximum Entropy Based Recommendation

Conversational recommender systems (CRSs) aim to capture the preferences of users with minimal interaction and then recommend appropriate items. Previous work formalizes a CRS as a Markov decision process that explores strategies for query timing and recommendation. In this paper, we propose a CRS that incorporates negative feedback from recommendations based on maximum entropy to clarify user preferences (MECRS). We formalize how MECRS explores and exploits new preferences. We also implement MECRS based on a classical CRS. On a real-world collected dataset, we find that MECRS achieves better performance when the user simulator is unable to accurately answer preferences.