I’m testing ProxyML with an example notebook and found this great example of why you should really care about counterfactual examples. Here we see a synthesized sample that was flagged by our “black box” credit risk model as being a bad credit risk. But ProxyML’s surrogate found a nearby counterfactual example that was identical except for the purpose of the loan switching to the purchase of a used car instead of a new car. Our “black box” model was asked to evaluate the counterfactual, and this synthesized applicant is now a “good” credit risk. Imagine being the loan officer that can now share a path forward and not a rejection.
No other changes, and our sensitive data never left the server!
