Technical Summary: MESD – A Risk-Sensitive Metric for Explanation Fairness Across Intersectional Subgroups

Problem Statement

Current fairness evaluation in machine learning is predominantly outcome-oriented, relying on metrics like Demographic Parity and Equalized Odds to ensure statistical consistency in predictions across protected groups. However, these metrics fail to detect procedural unfairness: situations where a model produces statistically fair outcomes but utilizes systematically different, unstable, or biased reasoning processes for different demographic groups.

This issue is exacerbated by intersectionality. Models may appear fair when evaluated on single attributes (e.g., race or gender independently) while exhibiting significant disparities for intersectional subgroups (e.g., Black women), a phenomenon known as fairness gerrymandering. Furthermore, existing procedural fairness metrics often rely on counterfactual comparisons or single protected attributes, failing to address the stability of explanations under perturbation across complex, multi-dimensional subgroup intersections.

Methodology

The paper proposes Multi-category Explanation Stability Disparity (MESD), a procedural fairness metric designed to quantify disparities in explanation quality across intersectional subgroups. MESD is integrated into a multi-objective optimization framework called UEF (Utility, Explainability, and Fairness).

1. The MESD Metric

MESD operates through a three-stage robust procedure:

• Intersectional Subgroup Construction: The metric constructs subgroups via the Cartesian product of multiple protected attributes ($G = A_1 \times A_2 \times \dots \times A_m$). This captures the unique experiences of multiply-marginalized individuals.
• Stability Evaluation via Perturbation: To measure explanation consistency, the method generates perturbed versions of input instances using a hybrid framework of Gaussian noise and stochastic feature masking. It computes an ensemble attribution vector (combining multiple explainers like SHAP and LIME) for the original and perturbed inputs. Stability is quantified by the inverse of the Euclidean distance between these explanations; a stable explanation yields a score near 1, while inconsistent ones approach 0.
• Label-Aware Aggregation with Empirical-Bayes Shrinkage: To prevent misleading disparity estimates caused by small sample sizes in intersectional groups, MESD aggregates stability scores conditionally on the true label ($y$). It employs Empirical-Bayes (EB) shrinkage to pull unstable estimates from small groups toward the label-wise mean, reducing variance without masking true disparities.
• CVaR-Weighted Risk Measurement: To address fairness gerrymandering, MESD uses Conditional Value-at-Risk (CVaR). Instead of averaging disparities, it focuses on the "worst-case" tail of the disparity distribution. It calculates pairwise stability disparities ($D_{ij}$) and assigns higher weights to pairs involving groups with low procedural fairness (high risk), ensuring the metric is sensitive to extreme sub-group instability.

2. The UEF Optimization Framework

To generate models that are fair in both outcome and procedure, the authors introduce UEF, a multi-objective optimization framework using NSGA-II (Non-Dominated Sorting Genetic Algorithm II). UEF simultaneously optimizes three objectives:

1. Utility ($F_{perf}$): Predictive performance (maximized AUC).
2. Outcome Fairness ($F_{out}$): Minimized Demographic Parity gap across intersectional groups.
3. Procedural Fairness ($F_{proc}$): Minimized MESD score.

A Chebyshev scalarization is applied post-optimization to select a single model from the Pareto front that balances these competing objectives.

Key Contributions

1. MESD Metric: The introduction of a procedural fairness metric that explicitly combines intersectional subgroup analysis, explanation stability (robustness to perturbation), and risk-sensitive aggregation (CVaR) to detect fairness gerrymandering.
2. UEF Framework: A multi-objective optimization approach that jointly optimizes utility, outcome fairness, and procedural fairness, demonstrating that these dimensions can be balanced rather than treated as mutually exclusive.
3. Comprehensive Validation: Empirical evidence showing that outcome-oriented fairness does not guarantee procedural fairness, and that optimizing for MESD reveals disparities invisible to standard metrics.

Experimental Results

The authors evaluated MESD and UEF on three benchmark datasets (Adult Income, Recidivism, German Credit) against four baselines: Empirical Risk Minimization (ERM), Reweighing (RW), Reductions (RD), and Adversarial Debiasing (ADL).

• Balanced Performance: UEF consistently achieved the most balanced performance across all three objectives. On the Adult Income dataset (the most complex with three protected attributes), UEF achieved the highest AUC and F1 scores while simultaneously achieving the lowest Demographic Parity (DP) and MESD scores.
• Pareto Efficiency: UEF produced 12 out of 15 non-dominated solutions across all datasets and folds. Notably, it was the only method to achieve non-dominated solutions on all 5 folds of the Adult Income dataset, where intersectional fairness gerrymandering is most prevalent.
• Detection of Hidden Disparities: Baseline methods optimized for outcome fairness (like RD) often achieved low DP gaps but high MESD scores, confirming that they failed to ensure consistent reasoning across groups.
• Ablation Studies:
  • CVaR: Removing CVaR weighting (using variance or max-gap instead) masked worst-case disparities in small subgroups.
  • EB Shrinkage: Without shrinkage, small intersectional groups exhibited artificially low stability scores due to high variance; shrinkage regularized these estimates.
  • Hyperparameters: MESD was found to be robust to perturbation baselines and stabilizes with sufficient perturbation counts ($K \ge 20$).

Significance and Claims

The paper positions MESD as a necessary component for procedural justice in AI, grounded in the theoretical distinction between distributive justice (outcomes) and procedural justice (process). The authors claim that:

• Procedural Fairness is Distinct: A model can be statistically fair yet procedurally unfair if it relies on different, unstable reasoning for different groups. This divergence undermines the legitimacy of the decision system.
• Intersectionality is Critical: Single-axis fairness analysis obscures the experiences of multiply-marginalized individuals. MESD is among the first metrics to explicitly address explanation stability at the intersection of multiple protected attributes.
• Optimization Trade-offs: Jointly optimizing for utility, outcome, and procedural fairness yields models that are superior to those optimized for single objectives, particularly in complex, high-dimensional fairness landscapes.
• Regulatory Relevance: The work supports compliance with emerging regulatory frameworks (e.g., EU AI Act) by providing a mechanism to audit the process of decision-making, not just the result.

The authors acknowledge limitations, noting that MESD relies on post-hoc explainers (which are approximations) and is currently limited to tabular data with binary protected attributes. They suggest future work should explore in-processing explainability and extensions to continuous or non-tabular modalities.