GESD: A New Procedural Fairness Metric for Machine Learning Explanations

Researchers have introduced a new fairness measure called Group-level Explanation Stability Disparity (GESD), which shifts the focus from just looking at outcomes to ensuring procedural fairness in machine learning. Unlike typical metrics like statistical parity or equal opportunity, GESD assesses how stable and robust model explanations are across different subgroups within a protected category. This measure works independently of specific models or explainers, enabling a more comprehensive analysis of fairness at the explanation level. GESD is part of an optimization framework named FEU, designed to balance utility, outcome fairness, and explanation fairness all at once. This work, found in arXiv:2605.15295v1, aims to deepen our understanding of bias in decision-making processes.