Research map / Learning theory and algorithms / Fairness

Standardized Interpretable Fairness Measures for Continuous Risk Scores

ICML 2024 · Learning theory and algorithms › Fairness · approach: Fairness-constrained classification · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes standardized, Wasserstein-based disparity measures for continuous risk scores covering independence, separation and calibration, shows they are invariant to monotone transforms, and relates them to ROC-based fairness measures.

Problem

Existing fairness measures are designed for binary classifiers or ROC curves; they are hard to interpret, not invariant to monotone score transformations, and can miss group disparities in continuous risk scores.

Approach

Generalize classifier rate differences (independence, equal opportunity, predictive equality, calibration) to continuous scores by averaging absolute bias over thresholds; uniform weighting gives a normalized Wasserstein-1 distance, while frequency-weighting gives a rank-based, transformation-invariant Wasserstein measure that also bounds ROC-based biases.

Key results

• On COMPAS, the standardized equal-opportunity bias is 0.161 and predictive-equality bias 0.154 (both p<0.01) while ROC bias is only 0.016, showing ROC measures miss separation bias.
• On UCI Adult, XGBoost achieves the lowest equal-opportunity bias (0.057) versus 0.107 for logistic regression.
• Removing the protected attribute from German Credit logistic regression lowers the EO bias from 0.083 to 0.048 but flips the direction of disadvantage to favor females on the positive component.

Measured effects reported in the paper

Surprising findings

• Removing the protected attribute can shift the bias direction, with the debiased German Credit model predominantly favoring the previously disadvantaged group on the positive component.
• ROC-based bias is low on COMPAS and in favor of females on Adult even while separation biases show clear disadvantage, so ROC measures can mask bias.

Limitations

• Measures are purely observational and do not establish causal discrimination; not all disparities are unfair.
• Only l1-norm based measures are developed; l_p variants lose the Wasserstein interpretation.
• Experiments are limited to three benchmark datasets and simple models, with no bias mitigation method evaluated.

Experimental setup

Datasets: COMPAS UCI Adult German Credit synthetic Gaussian scores

Benchmarks: COMPAS German Credit UCI Adult synthetic data

Metrics: standardized score bias (EO, PE, calibration) ROC bias cross-ROC bias permutation-test p-values positive/negative component shares

Baselines: ROC-based bias (ABROCA-style) cross-ROC bias AUROC-based intra-group and inter-group fairness calibration bias

Builds on: ABROCA equalized odds calibration within groups

Closest papers by meaning

• A Sequentially Fair Mechanism for Multiple Sensitive Attributes (AAAI 2024 · similarity 0.87)
• The Fragility of Fairness: Causal Sensitivity Analysis for Fair Machine Learning (NeurIPS 2024 · similarity 0.86)
• Understanding challenges to the interpretation of disaggregated evaluations of algorithmic fairness (NeurIPS 2025 · similarity 0.85)
• Post-hoc bias scoring is optimal for fair classification (ICLR 2024 · similarity 0.85)
• Reconciling Predictive and Statistical Parity: A Causal Approach (AAAI 2024 · similarity 0.84)
• Size-adaptive Hypothesis Testing for Fairness (NeurIPS 2025 · similarity 0.84)