Course contentsShow
← Machine Learning and Deep Learning
Lesson 3302 of 3,53872. Bias and FairnessFree lesson

Intersectionality in Bias Measurement

Evaluating fairness across intersecting protected attributes (e.g., race and gender combined).

Intersectionality in Bias Measurement

What you'll learn: How to evaluate fairness when protected attributes overlap, revealing biases that single-attribute analysis misses.

The Core Problem

Up until now, we've measured bias by looking at one protected attribute at a time—checking if your model treats men and women fairly, or if it treats different racial groups fairly. But real people exist at the intersection of multiple identities. A Black woman experiences the world differently than either Black men or white women, and bias measurement that only looks at "race" or "gender" separately can completely miss discrimination against this intersectional group.

Why Single-Attribute Analysis Falls Short

Imagine a hiring model that achieves:

  • Equal opportunity for men vs. women overall
  • Equal opportunity for Black vs. white candidates overall

Sounds fair, right? But when you break it down further, you might discover the model systematically rejects Black women specifically, while being generous to Black men and white women. The intersectional subgroup was hidden in the aggregated statistics.

Measuring Intersectional Bias

The approach is conceptually straightforward: compute your fairness metrics (demographic parity, equalized odds, etc.) not just for single protected attributes, but for combinations of them:

  • Race × Gender (e.g., Asian women, Latino men)
  • Age × Disability status
  • Race × Gender × Socioeconomic background

This explodes the number of subgroups you need to evaluate, but it reveals real-world disparities that matter.

Key Takeaway: Intersectionality in bias measurement requires evaluating fairness metrics across combinations of protected attributes, not just single attributes in isolation, to uncover discrimination against multiply-marginalized groups that aggregate statistics would hide.