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

Abstract

The notion of individual fairness requires that similar people receive\nsimilar treatment. However, this is hard to achieve in practice since it is\ndifficult to specify the appropriate similarity metric. In this work, we\nattempt to learn such similarity metric from human annotated data. We gather a\nnew dataset of human judgments on a criminal recidivism prediction (COMPAS)\ntask. By assuming the human supervision obeys the principle of individual\nfairness, we leverage prior work on metric learning, evaluate the performance\nof several metric learning methods on our dataset, and show that the learned\nmetrics outperform the Euclidean and Precision metric under various criteria.\nWe do not provide a way to directly learn a similarity metric satisfying the\nindividual fairness, but to provide an empirical study on how to derive the\nsimilarity metric from human supervisors, then future work can use this as a\ntool to understand human supervision.\n

Ähnliche Arbeiten