Learning Fair Representations
Algorithms AIMech:LearningFairRepresentations DRAFT
Aliases: LFR, Learned Fair Representations
A prototype-based representation-learning algorithm that jointly balances reconstruction, supervised prediction and differences between group-average prototype assignment probabilities.
Identity Notes & Concept Scope
The named LFR algorithm learns a prototype representation and a prediction rule. It is distinct from the general aim of learning fair representations, the finite-population FairnessLP formulation, and the later Variational Fair Autoencoder. The learned metric variant does not make it identical to AIO MetricLearning, which denotes a deep neural network in this pin. 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
2 Claims
Mechanism
#method-scope
Map an individual to softmax probabilities over learned prototypes using negative distances. Optimize prototype locations and bounded prototype label probabilities with a weighted sum of squared reconstruction loss, predictive cross-entropy and absolute differences of mean assignment probabilities between two protected groups. The paper also learns feature-distance weights separately by group and uses L-BFGS for its optimization.
Limitation
#representation-distribution-limit
Louizos and colleagues identify a higher-order-distribution limitation of matching average LFR assignment probabilities: groups can have the same average assignment while the distribution of those probability values differs. This concerns the exposed probability-vector representation, not a refutation of parity of an exactly matched sampled prototype index.
⚠️ Curated Limitations & Negative Results
- Fairness, reconstruction and predictive utility are weighted objectives; finite training and a nonzero penalty weight do not guarantee exact population parity.
- Equal group means of assignment probabilities do not imply equality of the distributions of the full assignment-probability vectors. Arbitrary nonlinear downstream processing or thresholding requires a separate fairness assessment.
- The method encourages local similarity; it does not certify the full supplied-metric pairwise Lipschitz criterion for every pair of individuals.
- The group-specific distance variant uses group membership for its mapping. Out-of-sample representation availability is not a proof of out-of-sample fairness.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:Classification
Classification
|
related_match | Learns a representation and prediction rule for classification. The algorithm is related to the classification task; the pin’s RepresentationLearning and MetricLearning classes specifically describe deep neural networks and are not asserted as exact or broad identities for this prototype method. |
Evidence Sources & Primary Literature
2 Sources| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
original-paper
|
Richard Zemel, Yu (Ledell) Wu, Kevin Swersky, Toniann Pitassi, Cynthia Dwork
|
primary paper
full_text
|
ID: https://proceedings.mlr.press/v28/zemel13.html
sha256:230af0b2ff49…
|
representation-critique
|
Christos Louizos, Kevin Swersky, Yujia Li, Max Welling, Richard Zemel
|
primary paper
full_text
|
ID: arXiv:1511.00830
sha256:3095f69e500a…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-11T05:03:15Z
Updated: 2026-10-11T05:03:15Z
Source File:
data/curated/LearningFairRepresentations.yaml