On the Direction of Discrimination: An Information-Theoretic Analysis of Disparate Impact in Machine Learning