Intersectional Fairness Analysis
What you'll learn: How to examine fairness across overlapping protected attributes to detect compound discrimination that single-attribute analysis misses.
Why Single-Attribute Analysis Isn't Enough
When you measure fairness for race separately and gender separately, you might conclude your model is fair. But what about Black women specifically? Or elderly Hispanic men? Intersectional fairness analysis examines combinations of protected attributes to uncover discrimination that affects people at the intersection of multiple identities.
The term comes from legal scholar Kimberlé Crenshaw, who showed how discrimination against Black women couldn't be understood by looking at race or gender alone—the experience was qualitatively different at the intersection.
How It Works in Practice
Instead of computing fairness metrics like equal opportunity or demographic parity for just "gender" or just "race," you compute them for intersectional groups:
- Women (all races) vs Men (all races)
- Black individuals (all genders) vs White individuals (all genders)
- Black Women vs White Women vs Black Men vs White Men
That third analysis often reveals hidden disparities. Your model might perform equally well for men and women overall, but Black women specifically could face significantly worse false negative rates than any other group.
The Challenge of Granularity
As you add more protected attributes (age, disability status, etc.), the number of intersectional subgroups explodes exponentially. You also face data sparsity—some combinations may have few examples. The key is identifying which intersections matter most for your context and ensuring adequate representation in your audit data.
Key Takeaway: Intersectional fairness analysis examines performance across overlapping protected attributes (like race AND gender together) to detect compound discrimination invisible in single-attribute audits—essential for truly equitable AI systems.