Interpretable Linear Dimensionality Reduction based on Bias-Variance Analysis