Shap-enhanced counterfactual explanations for recommendations

Abstract

Explanations in recommender systems help users better understand why a recommendation (or a list of recommendations) is generated. Explaining recommendations has become an important requirement for enhancing users' trust and satisfaction. However, explanation methods vary across different recommender models, increasing engineering costs. As recommender systems become ever more inscrutable, directly explaining recommender systems sometimes becomes impossible. Post-hoc explanation methods that do not elucidate internal mechanisms of recommender systems are popular approaches. State-of-art post-hoc explanation methods such as SHAP can generate explanations by building simpler surrogate models to approximate the original models. However, directly applying such methods has several concerns. First of all, post-hoc explanations may not be faithful to the original recommender systems since the internal mechanisms of recommender systems are not elucidated. Another concern is that the outputs returned by methods such as SHAP are not trivial for plain users to understand since background mathematical knowledge is required. In this work, we present an explanation method enhanced by SHAP that can generate easily understandable explanations with high fidelity.

Topics & Concepts