Knowledge MaxxingProxy variable
A proxy variable is a feature that closely correlates with a sensitive attribute, like race or gender, letting a model indirectly rely on that attribute even if it's not used directly.
Prerequisite: Measurement bias
What matters
Removing a sensitive attribute from a dataset doesn't guarantee fairness if another feature strongly correlates with it, since the model can still learn the same pattern through the proxy.
Example
A zip code can act as a proxy variable for race in some regions, letting a model indirectly discriminate even if race itself was never included as a feature.
Common misconception
Removing a sensitive attribute from training data does not guarantee a model can't still learn a related pattern through a correlated proxy variable.
Recall question
Why can a model still behave unfairly on a sensitive attribute even after that attribute is removed from training data?