Mission accomplished
This chapter’s mission was to see whether there was unfair bias in predicting whether a particular defendant would recidivate. We demonstrated that the FPR for African American defendants is 1.87 times higher than for Caucasian defendants. This disparity was confirmed with WIT, indicating that the model in question is much more likely to misclassify the positive class based on race. However, this is a global interpretation method, so it doesn’t answer our question regarding a specific defendant. Incidentally, in Chapter 11, Bias Mitigation and Causal Inference Methods, we will cover other global interpretation methods for unfairness.
To ascertain whether the model was racially biased against the defendant in question, we leveraged anchor and counterfactual explanations – they both output race as a primary feature in their explanations. The anchor did it with relatively high precision and coverage, and Counterfactuals Guided by Prototypes...
