Variational Fair Autoencoder
Architectures AIMech:VariationalFairAutoencoder DRAFT
Aliases: VFAE
A variational autoencoder architecture for learning representations with reduced sensitive or nuisance information, combining sensitive-variable-conditioned inference and generation with an MMD penalty on group-conditional aggregated latent distributions and a semi-supervised predictive component.
Identity Notes & Concept Scope
VFAE denotes the MMD-regularized semi-supervised variational architecture introduced by Louizos et al. It is distinct from the paper’s unregularized VAE baseline, prototype-based LFR, ordinary autoencoding and the model-dependent FAIRLEARNING algorithm. Prior independence and aggregated-posterior matching do not establish causal counterfactual fairness. 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
Use neural inference and generative models conditioned on the observed sensitive variable, with latent structure separating that variable from the representation in the generative model. In the semi-supervised construction, introduce a second latent layer and label variable and jointly train the inference, reconstruction and predictive components. Penalize differences between aggregated latent distributions across sensitive groups with maximum mean discrepancy, approximated using random Fourier features in the reported implementation.
⚠️ Curated Limitations & Negative Results
- Independence in the generative prior does not alone ensure independence of inferred representations; the MMD regularizer addresses remaining group dependence.
- Finite data, optimization, kernel choice and random-feature approximation limit how strongly matching can be claimed; no exact population independence or universal fairness guarantee is asserted.
- The encoder conditions on the sensitive variable. Semi-supervised here allows missing task labels, not arbitrary absence of sensitive-attribute information.
- The paper reports sampling the inferred latent representation for prediction rather than replacing it by its mean. Numerical objective-sign and bandwidth conventions are not certified by this conceptual record.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:VariationalAutoEncoder
Variational Auto Encoder
|
related_match | Uses variational autoencoding with additional sensitive-variable conditioning, a supervised component and MMD regularization. Because the pinned target definition specifies unsupervised learning, this mapping records the architectural relationship without asserting an unqualified subclass relation. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
original-paper
|
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/VariationalFairAutoencoder.yaml