An Empirical Study on Learning Fairness Metrics for COMPAS Data with Human Supervision