Correcting Underrepresentation and Intersectionality Bias for Fair Classification