Intersectional Fairness

The foundational methodology of intersectional fairness that disaggregates model performance metrics across compound demographic subgroups rather than isolated protected attributes.

• Evaluates metrics like false positive rate for specific intersections (e.g., Black women over 60) rather than race or gender alone
• Reveals masking effects where acceptable average group performance conceals severe discrimination against a nested subgroup
• Requires sufficient sample sizes per intersectional cell to achieve statistical significance
• Extends standard fairness metrics—demographic parity, equalized odds—to operate on cross-product subgroups

Without this decomposition, a lending model might appear fair to both women overall and Black applicants overall while systematically denying loans to Black women at a disproportionately high rate.

An approach that uses structural causal models to distinguish between direct discrimination, indirect discrimination via proxy variables, and legitimate explanatory factors that correlate with protected intersections.

• Maps causal pathways through which sensitive attributes influence decisions, identifying path-specific effects that constitute unfair discrimination
• Addresses the redlining proxy problem: ZIP codes, browsing history, and device types often encode intersectional identity information
• Enables counterfactual fairness testing at the intersectional level—asking whether a decision would change if an individual's race and gender were simultaneously altered
• Distinguishes between legitimate business necessity and discriminatory causal chains that perpetuate historical disadvantage

This framework is essential for compliance with regulations like the EU AI Act, which requires explainability for high-risk automated decisions affecting vulnerable populations.

Specialized quantitative measures designed to capture compound unfairness that single-axis metrics cannot detect.

• Differential Fairness: Extends differential privacy concepts to measure whether outcomes for intersectional subgroups diverge from population baselines beyond acceptable thresholds
• Minimax Fairness: Optimizes for the worst-performing subgroup, ensuring no intersectional group bears a disproportionate accuracy or error burden
• Subgroup Calibration: Verifies that predicted probabilities maintain their real-world meaning consistently across all intersectional cells
• Multicalibration: A stronger condition ensuring calibration holds simultaneously across every computationally identifiable subgroup, preventing hidden discrimination against emergent clusters

These metrics directly address the failure mode documented in the landmark Gender Shades study, where commercial facial recognition systems achieved high aggregate accuracy while catastrophically failing on darker-skinned women.

Operational practices for building training datasets that adequately represent intersectional identities and enable meaningful subgroup evaluation.

• Disaggregated demographic labeling: Collecting multiple attribute labels per data point with explicit consent, enabling post-hoc intersectional analysis
• Small group oversampling: Deliberately increasing representation of rare intersectional subgroups to achieve statistically meaningful sample sizes for evaluation
• Community-driven taxonomy design: Engaging affected communities to define relevant identity categories rather than imposing top-down demographic schemas that may erase critical distinctions
• Synthetic augmentation with constraints: Generating synthetic data for underrepresented intersections while preserving authentic distributional characteristics and avoiding stereotypical amplification

Without these practices, intersectional fairness evaluation becomes mathematically impossible—you cannot measure what you have not labeled and cannot mitigate what you cannot measure.

Technical interventions applied at different stages of the ML pipeline specifically targeting compound discrimination.

Pre-processing: Reweighting or resampling training data to balance representation across intersectional subgroups while preserving feature-label relationships.

In-processing: Adding adversarial debiasing constraints that prevent a model from encoding intersectional identity information in its latent representations, or optimizing directly for minimax fairness objectives during training.

Post-processing: Applying separate calibration thresholds per intersectional subgroup to equalize error rates, or using fair representation learning to transform model outputs into a space where intersectional identity is not predictive of outcomes.

Each approach involves trade-offs between fairness, accuracy, and computational complexity—and the choice depends on whether the protected intersectional attributes are available at inference time.