The paper "Equality of Opportunity in Supervised Learning" by Moritz Hardt, Eric Price, and Nathan Srebro proposes a criterion for measuring and removing discrimination in supervised learning based on protected attributes. The authors introduce the concept of "oblivious" measures, which only depend on the joint distribution of the predictor, the target, and the protected attribute, without evaluating individual features or the functional form of the predictor. They define two criteria: equalized odds and equal opportunity, which ensure that the predictor is independent of the protected attribute given the target. The paper provides a framework for constructing classifiers that satisfy these criteria from any learned predictor, minimizing the loss in utility. It also discusses the limitations of oblivious measures and presents a case study using FICO credit scores to illustrate the approach. The authors argue that their framework provides a meaningful measure of discrimination and achieves higher utility compared to other fairness notions like demographic parity.The paper "Equality of Opportunity in Supervised Learning" by Moritz Hardt, Eric Price, and Nathan Srebro proposes a criterion for measuring and removing discrimination in supervised learning based on protected attributes. The authors introduce the concept of "oblivious" measures, which only depend on the joint distribution of the predictor, the target, and the protected attribute, without evaluating individual features or the functional form of the predictor. They define two criteria: equalized odds and equal opportunity, which ensure that the predictor is independent of the protected attribute given the target. The paper provides a framework for constructing classifiers that satisfy these criteria from any learned predictor, minimizing the loss in utility. It also discusses the limitations of oblivious measures and presents a case study using FICO credit scores to illustrate the approach. The authors argue that their framework provides a meaningful measure of discrimination and achieves higher utility compared to other fairness notions like demographic parity.