Reinforced model-agnostic counterfactual explanations for recommender systems

Explanation is an important requirement for transparent and trustworthy recommender systems. When the recommendation model itself is not explainable, an explanation must be generated post-hoc. In contrast to traditional post-hoc explanation methods, counterfactual methods can provide scrutable and actionable explanations with high fidelity. Existing counterfactual explanation methods for recommender systems are either not generalizable or face a huge search space. In this work, we propose a reinforcement learning counterfactual explanation method MACER (Model-Agnostic Counterfactual Explanations for Recommender Systems) which generates item-based explanations for recommender systems. We embed the discrete action space into a continuous space, making it possible to use the process of finding counterfactual explanations as a task of reinforcement learning. This method treats the recommender system as a black box (model-agnostic) and has no requirement on the type of recommender system, and thus is applicable to all recommendation systems.