Explaining Recommendations: Fidelity versus Interpretability

A hungry academic who is attending the WebMedia conference in Gramado uses a mobile phone app to obtain a recommendation for a place-to-eat. The restaurant that the app recommends is not within walking distance and serves a fusion-style cuisine with which the academic is unfamiliar. Should she accept the recommendation? Her confidence in the recommendation might be improved by an explanation of the recommender system's decision-making. But the explanations that recommender systems provide at present are often post hoc: fidelity to the system's decision-making is sometimes sacrificed for interpretability. Are fidelity and interpretability always in conflict? Or can they can be reconciled without damaging the quality of the recommendations themselves? This talk will review the kinds of explanations given by recommender systems. It will describe a new generation of recommender systems in which recommendation and explanation are more intimately connected and which seeks to maintain the quality of the recommendations while providing explanations that are both intelligible and reasonably faithful to the system's operation.