P$^2$CE Offers Plausible, Pareto-Optimal Counterfactual Explanations for ML Models.

The growing adoption of machine learning algorithms in sensitive social applications, such as loan approvals or job selections, has highlighted the critical need for fairness and transparency. Counterfactual explanations are a key tool in this regard, empowering individuals to comprehend and potentially alter adverse decisions by suggesting actionable changes to input features that would lead to a desired outcome. However, existing methods often struggle to simultaneously achieve feasibility, plausibility, and computational efficiency. To address these challenges, researchers introduce P$^2$CE, an algorithm designed to generate plausible Pareto-optimal counterfactual explanations. This algorithm provides users with a diverse array of optimal trade-offs across various notions of feasibility. P$^2$CE incorporates an auxiliary isolation forest outlier detector to ensure that the generated explanations align with the underlying data distribution, thereby enhancing their plausibility. Furthermore, the algorithm leverages SHAP values to achieve optimal results with significantly reduced computation times, making it efficient regardless of the specific machine learning model being explained. Empirical evaluations conducted on three different datasets demonstrate that P$^2$CE delivers superior performance in both the quality of its solutions and its computational efficiency when compared to related techniques.