An Axiomatic Theory of Provably-Fair Welfare-Centric Machine Learning

Neural Information Processing Systems 

We avoid this issue by working directly with malfare and loss.