Comment by tbrownaw

There are multiple different ways to measure performance. If different groups have different rates of whatever you're predicting, it is not possible to have all of the different ways of measuring performance agree on whether your model is fair or not.