Equalized odds is a group-fairness criterion introduced by Hardt, Price, and Srebro in their 2016 paper “Equality of Opportunity in Supervised Learning.” It requires that a classifier perform equally well, in two specific senses, for every protected group. Concretely, the true-positive rate (how often the model correctly flags real positives) and the false-positive rate (how often it raises a false alarm) must be the same across groups such as gender or race.
The intuition is that the model should be no more likely to wrongly accuse, and no less likely to correctly identify, one group than another. A weaker variant, equal opportunity, requires only equal true-positive rates, which is the right choice when the most important harm is missing a genuine positive, for instance failing to extend credit to someone who would have repaid.
Equalized odds is central to the impossibility results that shaped algorithmic fairness. When two groups have different base rates of the outcome being predicted, it is mathematically impossible for a classifier to satisfy equalized odds and certain other fairness definitions at the same time, the same tension that drove the public dispute over the COMPAS recidivism tool.
Why a business reader should care: equalized odds is the criterion many fairness toolkits actually compute, and unlike demographic parity it accounts for whether decisions are correct. Knowing that it generally cannot be reconciled with other definitions forces an explicit, defensible choice about which kind of fairness an organization is committing to.