Beyond Adult and COMPAS: Fair Multi-Class Prediction via Information Projection

We formulate this problem in terms of "projecting" a pre-trained (and potentially unfair) classifier onto the set of models that satisfy target