Eliminating Disparate Impact in Modeling
Daniel M. Tom · 2024
There is considerable regulatory oversight on credit lending. Besides having to perform well predicting credit defaults, a credit model also needs to be explainable and non-discriminatory on protected demographics. A regression model neutralization technique eliminates disparate impact discrimination. We use it in our Logistic Regression with classic AI Beam Search, so our model meets these criteria. In contrast, the adversarial debiasing methodology has shortcomings, which we discuss.