Measuring and correcting bias
Fairness can be measured, but there are several definitions, sometimes incompatible with one another.
ObjectiveDistinguish several definitions of fairness and choose the one that suits a given decision.
Saying that a system is fair requires specifying fair towards whom and by which criterion. Should each group have the same acceptance rate, the same error rate, or should a score be equally reliable whatever the group? The reference studies show that these criteria generally cannot all be satisfied at once.
- Demographic parity: each group receives the same proportion of positive outcomes.
- Equal error rates: false positives and false negatives are equally frequent in each group.
- Calibration: the same score means the same actual risk, whatever the group.
The choice is therefore a human decision that depends on context. In medical screening, missing a case is more serious than a false alarm. For a sanction, it is the reverse.
References
- Barocas, Hardt, Narayanan (2023). Fairness and Machine Learning: Limitations and Opportunities. MIT Press. fairmlbook.org/