How can machine learning be both fair and accurate?