An empirical evaluation of active learning strategies for profile elicitation in a conversational recommender system

Conversational Recommender Systems have received widespread attention in both research and practice. They assist people in finding relevant and interesting items through a multi-turn conversation. The use of natural language interaction also allows users to express their preferences with more flexibility. However, these systems often have to work in a cold-start situation, and most of the conversation is dedicated to the profile elicitation step. In order to ensure good recommendations, this profile should be as rich as possible, which requires great user effort. In this paper, we investigate the application of Active Learning techniques for improving the profile elicitation step in a Conversational Recommender System. We compared five different state-of-the-art techniques, and carried out a user study with 219 users in order to assess their effectiveness both in terms of recommendation accuracy and user effort. Results show that assisting users by providing personalized suggestions during the profile elicitation step improves the quality of the recommendations in terms of Hit Rate and nDCG, compared to a strategy that requires users to come up with preferences on their own.