Decoupled classifiers for group-fair and efficient machine learning

When it is ethical and legal to use a sensitive attribute (such as gender or race) in machine learning systems, the question remains how to do so. We show that the naive application of machine learning algorithms using sensitive attributes leads to an inherent tradeoff in accuracy between groups. We provide a simple and efficient decoupling technique, that can be added on top of any black-box machine learning algorithm, to learn different classifiers for different groups. Transfer learning is used to mitigate the problem of having too little data on any one group.