Counterfactual Explanations for Recommendation Bias

Leonidas Zafeiriou, Panayiotis Tsaparas, Evaggelia Pitoura

Abstract: Today, we rely heavily on automated recommendation algorithms for assisting us in making several decisions. These algorithms are trained on large quantities of user interaction data, and as a result they incorporate various biases of the data in their recommendations. It is important to understand the origins of recommendation biases, however, this is becoming increasingly difficult, given the complexity of the recommenders. To address model complexity, researchers try to provide explanations for the behavior of the algorithms, such as counterfactual explanations. In this work, we consider explanations for recommendation bias, and we generalize counterfactual explanations to handle groups of users and items. We then consider a random-walk based recommender, and we propose efficient algorithms for computing the counterfactual explanations. We perform an experimental evaluation of our algorithms using both real and synthetic data.