Home / Curated / Evaluation Criteria / Equalized odds

Equalized odds

Evaluation Criteria AIMech:EqualizedOdds DRAFT

Aliases: Equalised odds

A fairness criterion requiring a predictor's output to be conditionally independent of a specified protected attribute given the target outcome; for binary predictions and outcomes, it requires equal true-positive and false-positive rates across the compared groups.

Identity Notes & Concept Scope This is the distributional criterion in Hardt, Price and Srebro's Definition 2.1, not their postprocessing algorithm, a bias type, demographic parity or a claim of overall social fairness. Equal opportunity is a weaker positive-outcome-only constraint. All 442 pinned AIO labels/scoped synonyms and 94 preceding native identities were compared, with ignored-inclusive text searches in both checkouts.

Claims & Grounded Evidence

1 Claim

Scope #criterion-scope

Assess whether the conditional distribution of the prediction given the outcome is the same across protected groups. In the binary case, test both P(prediction=1 | A=a,Y=1) and P(prediction=1 | A=a,Y=0) across a. The general independence definition also applies to scores and nonbinary variables; thresholding an equalized-odds score by a common rule preserves the property.

supports Source: Equality of Opportunity in Supervised Learning · Locator: Section 2, Definition 2.1, PDF p. 3; Sections 2.2–2.3, PDF p. 4; Section 6, PDF pp. 12–14; discussion of reliable targets and nonconclusive fairness tests, PDF pp. 20–21

⚠️ Curated Limitations & Negative Results

The conditioning outcome and protected groups must be specified, with estimable nonzero-probability conditioning strata for rate comparisons. Biased or unreliable outcome labels can undermine interpretation.
Equal true-positive and false-positive rates do not require identical overall accuracy when group outcome prevalences differ. The record uses the formal conditional-independence definition, not the source's informal equal-accuracy wording.
This is an oblivious distributional test: identical prediction/outcome/group statistics can arise from different underlying dependency structures. Satisfaction or violation alone is not conclusive evidence of overall fairness or unfairness.
Finite-sample uncertainty, a tolerance and domain-specific interpretation are needed in an application; no empirical result, universal fairness guarantee or legal-compliance claim is asserted.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:Classification` Classification	related_match	The criterion constrains classifier predictions and error rates in its central binary setting. Classification denotes the prediction task; equalized odds denotes a condition used to assess its outputs, so this is a direct task/criterion relationship rather than equivalence or subclassing.

Evidence Sources & Primary Literature

1 Source

Source ID	Publication / Document	Type / Access	Reference & Checksum
`original-paper`	Equality of Opportunity in Supervised Learning Moritz Hardt, Eric Price, Nathan Srebro	primary paper full_text	ID: arXiv:1610.02413 sha256:e92b4f579180…

Curation Provenance

Curator: Codex

Method: ai_assisted

Model: GPT-6

Created: 2026-10-11T04:48:50Z

Updated: 2026-10-11T04:48:50Z

Source File: data/curated/EqualizedOdds.yaml