Explaining Black-Box Algorithms Using Probabilistic Contrastive Counterfactuals

Algorithmic systems are increasingly used to aid in decision-making, with potentially significant consequences for individuals, institutions and society. This has led to much interest in explainable artificial intelligence (XAI), which aims to reduce the opaqueness of AI-based decision-making systems, allowing humans to better understand and trust these systems. Much work in this context has focused on the attribution of responsibility for an algorithms decisions to its inputs; responsibility is typically approached as a purely associational concept. In this paper, we propose a principled causality-based approach for explaining black-box decision-making systems that unifies existing methods in XAI and addresses their limitations. At the core of our framework lies probabilistic contrastive counterfactuals, a concept that can be traced back to philosophical, cognitive, and social foundations of theories on how humans generate and select explanations. We show that such counterfactuals can serve as a theoretical foundation for both quantifying direct and indirect influences of a variable on decisions made by an algorithm as well as generate an actionable recourse for individuals negatively affected by the algorithm's decision. Unlike previous proposals, our system (1) can computeprovably correct explanations and recourse at local, global and contextual levels, (2) is designed to work with users with any level of knowledge of the underlying causal model, and (3) makes no assumptions about the internals of an algorithmic system except for the availability of its input-output data. We evaluate our proposal on real and synthetic data and demonstrate improvement over state-of-the-art approaches in XAI, including popular LIME, SHAP and actionable recourse methods.