Home / Curated / Evaluation Criteria / Lipschitz individual fairness

Lipschitz individual fairness

Evaluation Criteria AIMech:LipschitzIndividualFairness DRAFT

Aliases: Individual fairness (Dwork et al.), Fairness through awareness (Lipschitz criterion)

An individual-fairness criterion requiring the distance between the output distributions assigned by a randomized classifier to any two individuals to be no greater than their task-specific similarity distance.

Identity Notes & Concept Scope The record is the explicit (D,d)-Lipschitz condition in Dwork and colleagues' framework. It is narrower than every possible notion of individual fairness and separate from their loss-minimizing linear-program construction, metric-learning algorithms, differential privacy and group parity. Identity assessment covered 442 AIO labels/scoped synonyms, 94 preceding native identities and ignored-inclusive searches of both checkouts.

Claims & Grounded Evidence

1 Claim

Scope #criterion-scope

Given a mapping M from individuals to distributions over outcomes, an externally specified task distance d and an output-distribution distance D, require D(M(x),M(y)) <= d(x,y) for every compared pair. The criterion constrains permissible mappings; choosing among them requires a separate utility objective. The paper studies total-variation and relative-l-infinity output distances as particular choices.

supports Source: Fairness Through Awareness · Locator: Sections 1.1–1.2, printed pp. 1–3 / PDF pp. 2–4; Section 2, Definition 2.1 and equation (1), printed p. 4 / PDF p. 5; Section 2.1 probability metrics and Remark 2.1, printed p. 5 / PDF p. 6

⚠️ Curated Limitations & Negative Results

The task-specific distance and output-distance choice determine the constraint. The framework does not certify that a chosen similarity metric represents an appropriate social or domain judgment.
A constant output distribution satisfies the condition; useful predictions require a separate utility objective. Criterion satisfaction alone does not guarantee predictive quality.
The formulation over an inspected finite population does not itself guarantee generalization to unobserved individuals. Group-parity implications require additional conditions on the metric and group distributions.
This record does not equate the criterion with all individual-fairness notions, certify the paper's full optimization/theorem results, or assert legal compliance.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:Classification` Classification	related_match	The condition directly constrains randomized classifiers' outcome distributions. Classification denotes the task while this local concept denotes a constraint on its mapping, so related_match preserves the abstraction difference. MetricLearning is not a substitute: the defining criterion assumes a task metric rather than specifying a neural procedure that learns one.

Evidence Sources & Primary Literature

1 Source

Source ID	Publication / Document	Type / Access	Reference & Checksum
`original-paper`	Fairness Through Awareness Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Reingold, Richard Zemel	primary paper full_text	ID: arXiv:1104.3913 sha256:0a5afde6e5a2…

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/LipschitzIndividualFairness.yaml