Making Bias Amplification in Balanced Datasets Directional and Interpretable