Errors, balanced by group
Equalized odds requires that the model have equal true positive rates and equal false positive rates across groups. In other words, among people who truly qualify, every group is approved at the same rate, and among people who truly do not, every group is wrongly approved at the same rate.
Why condition on the truth
Unlike demographic parity, equalized odds conditions on the actual outcome. It is a separation criterion: the prediction is independent of the protected attribute given the true label. This respects real differences in base rates while still demanding balanced errors.
What it protects against
- A model that is accurate for one group but makes far more false positives or false negatives for another.
- Hidden double standards where qualified members of one group are missed more often.
The cost
Meeting equalized odds may force the model to discard useful signal or randomize some decisions, which can reduce overall accuracy.
Key idea
Equalized odds demands equal true positive and false positive rates across groups conditioned on the true label, balancing both error types while respecting genuine base rate differences.