Knowledge MaxxingIntersectional evaluation
Intersectional evaluation measures a model's performance for people belonging to combinations of multiple groups at once, not just one group category at a time.
Prerequisite: Subgroup evaluation
What matters
A model can perform acceptably when each group category is checked separately, yet still fail badly for a specific combination of categories that was never evaluated together.
Example
A model evaluated separately by age group and separately by gender might still perform poorly for older women specifically, a combination neither single check would reveal.
Common misconception
Checking fairness for each group category separately does not guarantee fairness for every combination of categories; disparities can hide specifically at the intersection.
Recall question
Why might separate checks for each group category miss a disparity that intersectional evaluation would catch?