Detecting Bias in Black-Box Models Using Transparent Model Distillation.

Black-box risk scoring models permeate our lives, yet are typically proprietary and opaque. We propose a transparent model distillation approach to detect bias in such models. Model distillation was originally designed to distill knowledge from a large, complex teacher model to a faster, simpler student model without significant loss in prediction accuracy. We add a third restriction - transparency. In this paper we use data sets that contain two labels to train on: the risk score predicted by a black-box model, as well as the actual outcome the risk score was intended to predict. This allows us to compare models that predict each label. For a particular class of student models - interpretable tree additive models with pairwise interactions (GA2Ms) - we provide confidence intervals for the difference between the risk score and actual outcome models. This presents a new method for detecting bias in black-box risk scores by assessing if contributions of protected features to the risk score are statistically different from their contributions to the actual outcome.