This paper introduces a framework for measuring and removing discrimination in supervised learning based on protected attributes. The authors propose two criteria: equalized odds and equal opportunity. Equalized odds requires that the true positive and false positive rates are equal across protected groups, while equal opportunity focuses only on the true positive rate for the "advantaged" outcome. The framework is "oblivious," meaning it depends only on the joint statistics of the target, predictions, and protected attribute, not on individual features or the predictor's functional form. The authors show that their framework allows for more accurate predictors and aligns fairness with the central goal of supervised learning. They also demonstrate that their approach can be implemented through post-processing of existing predictors, without altering the training process. The paper also discusses the limitations of oblivious measures and shows that different dependency structures can be indistinguishable based on such measures. The authors provide a case study using FICO credit scores to illustrate their framework.This paper introduces a framework for measuring and removing discrimination in supervised learning based on protected attributes. The authors propose two criteria: equalized odds and equal opportunity. Equalized odds requires that the true positive and false positive rates are equal across protected groups, while equal opportunity focuses only on the true positive rate for the "advantaged" outcome. The framework is "oblivious," meaning it depends only on the joint statistics of the target, predictions, and protected attribute, not on individual features or the predictor's functional form. The authors show that their framework allows for more accurate predictors and aligns fairness with the central goal of supervised learning. They also demonstrate that their approach can be implemented through post-processing of existing predictors, without altering the training process. The paper also discusses the limitations of oblivious measures and shows that different dependency structures can be indistinguishable based on such measures. The authors provide a case study using FICO credit scores to illustrate their framework.