Exponentiated-gradient reduction for fair classification
Algorithms AIMech:ExponentiatedGradientFairClassification DRAFT
Aliases: Agarwal exponentiated-gradient fairness reduction
A cost-sensitive classification reduction that alternates classifier best responses with exponentiated-gradient updates of bounded Lagrange multipliers to approximate an error–fairness saddle point and return a randomized classifier mixture.
Identity Notes & Concept Scope
This is Algorithm 1 of Agarwal et al., not generic exponentiated-gradient optimization. It returns a distribution over classifiers and is distinct from the paper’s deterministic-candidate grid-search reduction and from postprocessing a fixed classifier. All 442 pinned AIO labels/scoped synonyms and 104 preceding native identities were compared, together with hidden/ignored-inclusive searches of both checkouts; source citations and pending frontier mentions were distinguished from existing concept identities.
Claims & Grounded Evidence
1 Claim
Mechanism
#method-scope
Express fairness as linear inequalities on conditional prediction moments whose conditioning events do not depend on the classifier. Given multiplier values, call a cost-sensitive classification oracle with modified per-example label costs. Update multipliers from constraint violations using exponentiated gradient, average the classifiers and multipliers, and stop when the computed saddle-point gap reaches the chosen tolerance.
⚠️ Curated Limitations & Negative Results
- The optimization analysis assumes a cost-sensitive best-response oracle; heuristic base learners need not realize the exact oracle guarantees.
- Bounds retain optimization tolerance, the multiplier-norm cap and statistical error, including dependence on the number of observations in each conditioning event.
- The constraint representation excludes conditioning events that depend on the learned predictor; the paper identifies predictive parity and calibration as outside this formulation.
- Protected attributes are required for training constraints. No separate protected-attribute argument is needed at prediction time only when the chosen feature vector and base classifier do not require it.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:BinaryClassification
Binary Classification
|
related_match | Trains binary classifiers under conditional-moment constraints. The reduction is a learning algorithm related to the binary classification task, not an equivalent task or ontology subclass assertion. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
original-paper
|
Alekh Agarwal, Alina Beygelzimer, Miroslav Dudík, John Langford, Hanna Wallach
|
primary paper
full_text
|
ID: https://proceedings.mlr.press/v80/agarwal18a.html
sha256:bcdbc1454b2a…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-11T05:03:15Z
Updated: 2026-10-11T05:03:15Z