Evaluating ML classifiers using relative error instead of absolute accuracy (stephantul.github.io)