This GitHub repo has:
- code to clean publicly-available data sets with both risk scores and actual outcomes (COMPAS, Chicago Police, NYPD Stop-and-Frisk, Lending Club). This code is run-able.
- pseudocode for distillation setup proposed in the paper "Distill-and-Compare: Auditing Black-Box Models Using Transparent Model Distillation" presented at AAAI/ACM AIES 2018. https://arxiv.org/abs/1710.06169. You will need to pick a transparent model class and implement it / call an implementation of it.
Please email ht395@cornell.edu if you have questions.