Rethinking Class Imbalance in Machine Learning