New Framework for Counterfactual Explanations Enhances ML Interpretability

Counterfactual explanations (CEs) are a key method for making machine learning models more interpretable by identifying minimal input changes that alter a model's output to a desired state. Traditionally, CEs are framed as a distance-minimization problem, but their underlying theoretical basis has been less explored. This research establishes a theoretical link, demonstrating that distance-minimization CEs can be understood as a maximum a posteriori (MAP) estimate within a Generalized Bayes framework, specifically when a distance-based prior is used. This new perspective, termed Distance-Prior Generalized Bayes CE (DP-GBCE), opens doors for more sophisticated decision-making. Building on this, the paper introduces additional decision rules beyond MAP, including a Bayes decision rule that minimizes expected loss and a risk-averse CVaR-CE. It also proposes an extension to account for model multiplicity by mixing posterior distributions from multiple models. Finally, new metrics are defined for evaluating both individual CEs and the overall posterior distribution, with experiments on simulated and real-world data illustrating the trade-offs of these new decision rules.