Comparing reward-based optimal behaviors in user-adapted Recommender Systems

Conversational Recommender Systems (CRSs) are intelligent E-commerce applications that interactively assist online users by following a default recommendation strategy. Typically, the strategy remains hard-coded during the interaction, thus making it impossible for CRSs to adapt to the dynamic user needs. In a previous paper, we have proposed and validated a novel technology that allows CRSs to autonomously improve a default strategy and learn the optimal (best) one. Our approach is based on the Reinforcement Learning (RL) paradigm, and uses a numerical reward model that is pre-selected arbitrarily by the system designers. In this context, it becomes important to determine the effect of using different reward models on the optimal system behavior. In this paper, we investigate the issue of strategy-learning under different reward model selections. We show that, even in a limited (simulated) setting, different and intelligent optimal behaviors are learnt under different rewards, and that our results largely correlate with our hypotheses about the expected optimal system behavior. We also deduce the importance of selecting the reward model carefully, based on the users´ contextual needs.