Simple lending use case which is predicted the probability of late payment / default.

A synthetic feature was created to make a based line model unfair. Specifically, if a record was female, and that record was defaulted, the distribution on the LIMIT_BAL feature was pushed right.

Accounting for the protected class, and completings assessment of this model, we will see that is is unfair based on Equalized Odds Difference which is max( difference false postive rate, difference in false negative rate)

Mitigation is pursued via

• postprocessing
• reduction -> using fairness measures as constraints in model fitting