Bias Mitigation - Post-processing

Explanation

Post-processing bias mitigation checks how effectively a corrective adjustment applied after the model produces scores reduces residual disparity — for instance recalibrating auto-loan pricing outputs so groups are treated more equitably without distorting genuine risk. It matters because models often retain bias even after careful training, and a transparent post-hoc adjustment is sometimes the most practical lever, provided it actually shrinks the gap. It is measured by the Post-hoc Adjustment Score, computed as Adjustment Score = (Final_Disparity − Initial_Disparity) / Initial_Disparity for auto-loan pricing models: the fractional reduction in disparity achieved by the adjustment. To implement it, measure the disparity before adjustment, apply the post-processing step (e.g. group-aware threshold or score calibration), re-measure disparity, and log both values, the adjustment logic, and the resulting score as auditable evidence, alongside risk-alignment metrics to confirm pricing still reflects true risk. Run it whenever the mitigation is calibrated and re-validate periodically. The threshold requires greater than 0.7 reduction while maintaining risk alignment; if the reduction is 0.7 or less, or risk alignment degrades, it triggers a breach — revisit the mitigation strategy, consider in-processing fixes, and document the decision, in line with RBI Fairness and Equity and NIST MAP 2.1/4.1.