Equality Of OddsHow Can We Actually Fix the Bias?
If a model left to itself is unfair, we have to step in. There are two main places to do it.
During Training
Training with a fairness rule. While the model learns, we add a condition, just like the constraints of a linear programming problem: make the error as small as possible, subject to the TPR gap between the groups staying below a small limit, say 0.02. The model gives up a little accuracy in exchange for a smaller gap.
After Training
Post-processing thresholds. If the model cannot be trained again, we keep its scores and choose a separate threshold for each group so that the rates match.
The two sliders you used in the last section were
Where the Curves Cross
Engineers look for fair thresholds on ROC curves. Each school gets its own curve, and every threshold is one point (FPR, TPR) on it. Wherever the Hill School curve and the River School curve meet, both schools can have the same TPR and the same FPR. Find all three meeting points.
The two corners are the lazy solutions. The crossing in the middle is the only useful fair point — and each school reaches it with a different threshold, 60 and 45, because the model's scores mean different things in the two schools.
Fair to Whom?
Each way of choosing thresholds protects someone. Compare four rules: how many qualified students each school loses, how many false alarms there are, and how many decisions are right overall. Look at all four.
Under equalized odds, Arjun from Hill School scores 57 and is not shortlisted, while Kavya from River School scores 48 and is. Arjun's family says this is unfair to him. Which statement is the most accurate?
Tip
No single setting pleases everyone: when groups differ, the main fairness rules usually cannot all hold at once. That is why the choice should be made openly — by the people responsible, listening to the people affected — and written down, not left hidden inside a model.
Speed Round: Which Rule Holds?
Each slide shows an audit of a model for the two schools. Get five right to move on.