Adjust the outputs
Post processing leaves the trained model alone and instead transforms its outputs. You take the scores the model already produces and adjust thresholds or scores per group to meet a fairness or calibration target.
Two related goals
- Group thresholding: choose a different decision threshold for each group so that a metric like equal opportunity holds.
- Calibration: ensure that a predicted probability means the same thing for every group, so among people scored at seventy percent, about seventy percent are truly positive in each group.
Why teams like it
- It needs only the model outputs and the protected attribute, not retraining or data access.
- It is fast and works with any black box scorer, including vendor models.
Limits
- It cannot fix a model that lacks the signal to be fair; it only redistributes existing errors.
- Using group specific thresholds means treating people differently by group, which may raise legal or ethical concerns.
Key idea
Post processing adjusts a trained model thresholds or scores per group to reach fairness or calibration without retraining, which is fast and model agnostic but only redistributes the errors the model already makes.