Push fairness up, accuracy often drops. The Pareto frontier shows the best trade.
From the chapter: Chapter 16: Ethics & Safety
Glossary: algorithmic fairness, demographic parity, equalised odds, pareto frontier
Transcript
A model that is highly accurate on the population as a whole may still treat subgroups unequally. Different fairness criteria capture this in different ways.
Demographic parity asks that each group receives the same fraction of positive predictions, regardless of base rate.
Equal opportunity asks that the true positive rate is the same across groups.
Calibration asks that a predicted probability of seventy per cent means seventy per cent in every group.
These criteria conflict. In general no model can satisfy all three at once unless the base rates are identical across groups, which they rarely are.
Plot accuracy on one axis and a fairness measure on the other. Each model is a point. The set of best-possible trade-offs traces out a frontier.
Pushing further along the frontier toward fairness usually costs accuracy. The size of the cost depends on how different the groups are.
The choice of where to sit on the frontier is not technical. It is a question about values: what kind of mistake are you willing to make, and against whom.