Equal Opportunity
A machine learning fairness criterion requiring that, among genuinely qualified individuals, a model's true positive rate be consistent across groups. It was systematically formalized by Moritz Hardt and colleagues in 2016 and corresponds more closely than statistical parity to the practical fairness demands of settings such as employment and credit.
Timeline (last 90 days)
Commonly used fairness metrics include demographic parity, equal opportunity, and predictive parity; in practice there is often a trade-off between predictive parity and equal opportunity, and it is recommended to choose a primary metric according to the business scenario and document the reasons for the trade-off for audit purposes.
Algorithmic fairness refers to the property that machine learning models do not produce systematic discrimination or bias across different groups (such as gender, race, age); it is a set of conflicting metrics, for example demographic parity and equal opportunity often cannot be satisfied simultaneously.
Pre-deployment fairness assessment should include three dimensions: statistical parity analysis across demographic groups, equal opportunity assessment for qualified candidates, and intersectional bias assessment.
Equal opportunity was systematically formalized by Hardt and other scholars in 2016, requiring that among people whose true label is qualified or positive, the model's true positive rate be consistent across groups.
Statistical parity and equal opportunity often cannot be satisfied simultaneously in mathematical terms, and actual deployment requires choosing a primary metric based on the business scenario and documenting the reasons for the trade-off.
All Facts (5)
Commonly used fairness metrics include demographic parity, equal opportunity, and predictive parity; in practice there is often a trade-off between predictive parity and equal opportunity, and it is recommended to choose a primary metric according to the business scenario and document the reasons for the trade-off for audit purposes.
50%UnverifiedAlgorithmic fairness refers to the property that machine learning models do not produce systematic discrimination or bias across different groups (such as gender, race, age); it is a set of conflicting metrics, for example demographic parity and equal opportunity often cannot be satisfied simultaneously.
50%UnverifiedPre-deployment fairness assessment should include three dimensions: statistical parity analysis across demographic groups, equal opportunity assessment for qualified candidates, and intersectional bias assessment.
50%UnverifiedEqual opportunity was systematically formalized by Hardt and other scholars in 2016, requiring that among people whose true label is qualified or positive, the model's true positive rate be consistent across groups.
50%UnverifiedStatistical parity and equal opportunity often cannot be satisfied simultaneously in mathematical terms, and actual deployment requires choosing a primary metric based on the business scenario and documenting the reasons for the trade-off.
50%