“Equality of Opportunity in Supervised Learning,” by Moritz Hardt, Eric Price, and Nathan Srebro, was published in 2016 and gave the field two of its most-cited fairness definitions: equalized odds and equal opportunity. The paper responds to a weakness in earlier notions like demographic parity, which can force a model to make obviously bad decisions just to equalize approval rates across groups.
Equalized odds requires that a classifier’s true-positive rate and false-positive rate be equal across protected groups, meaning the model is equally good at catching real positives and equally prone to false alarms regardless of group membership. Equal opportunity is the weaker version that only requires equal true-positive rates, which is appropriate when the harm of a missed positive is what matters most, for example denying a qualified applicant.
A useful feature of the framework is that it is oblivious: it depends only on the joint statistics of the prediction, the true outcome, and the protected attribute, so it can be checked and even enforced after the fact by adjusting a trained predictor. The authors demonstrate the idea on FICO credit-score data, a setting where these tradeoffs are concrete.
Why a business reader should care: equalized odds and equal opportunity are the criteria many fairness audits and tools actually measure. Understanding the difference, and that you usually cannot satisfy these alongside demographic parity at once, is essential when an organization commits to a fairness standard.