Equality Of OddsSummary
One page to revise from — the ideas first, then every formula you need.
- A model can be accurate overall and still be unfair: its mistakes can fall more often on one group, usually because of the data it learned from rather than anyone's intention.
- A fairness audit works out the error rates group by group. Real audits compare groups defined by protected attributes.
- Equal opportunity: the same true positive rate in every group, so every qualified person has the same chance.
- Equalized odds: the same true positive rate and the same false positive rate in every group.
- One threshold for everyone can treat groups very differently. Post-processing gives each group its own threshold; a fairness rule can also be built into training.
- On an ROC plot, equalized odds holds where the groups' curves cross. The corners, where everyone or no one is shortlisted, are lazy solutions.
- Fairness usually costs some accuracy. Which rule to use is a human decision, to be made openly.
Rates
Fairness rules
equal opportunity:
equalized odds: also