Critiquing with Confidence

The ability of a CBR system to evaluate its own confidence in a proposed solution is likely to have an important impact on its problem solving and reasoning ability; if nothing else it allows a system to respond with “I don’t know” instead of suggesting poor solutions. This ability is especially important in interactive CBR recommender systems because to be successful these systems must build trust with their users. This often means helping users to understand the reasons behind a particular recommendation, and presenting them with explanations, and confidence information is an important way to achieve this. In this paper we propose an explicit model of confidence for conversational recommendation systems. We explain how confidence can be evaluated at the feature-level, during each cycle of a recommendation session, and how this can be effectively communicated to the user. In turn, we also show how case-level confidence can be usefully incorporated into the recommendation logic to guide the recommender in the direction of more confident suggestions.