Bivariate counterfactual-fairness sensitivity analysis
Evaluation Methods AIMech:BivariateCounterfactualFairnessSensitivity DRAFT
Aliases: Kilbertus grid-based counterfactual-fairness sensitivity
A model-based sensitivity procedure that varies the correlation between two structural error terms and measures how a predictor built under an unconfounded additive-noise model changes on counterfactual data generated under the alternative confounded model.
Identity Notes & Concept Scope
This record denotes the paper’s two-error grid-based sensitivity procedure. The mean squared discrepancy is its output, not another fairness criterion. The multivariate MAXCFU optimizer is separately represented because it searches a matrix of correlations with a different algorithm. All 442 pinned AIO labels/scoped synonyms and 99 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
Fit an unconfounded working model and a predictor of its residuals. At a selected error correlation, fit an alternative additive-noise model, retain its inferred errors while changing the protected attribute, and propagate the intervention through the causal ordering. Recompute residuals of these counterfactual features using the original working-model coefficients, apply the fixed original predictor and average squared prediction differences. Repeat across correlations to obtain a sensitivity curve.
⚠️ Curated Limitations & Negative Results
- Alternative Model B is a hypothetical data-generating model, not an identified causal ground truth; the sensitivity parameter and acceptable discrepancy require domain judgment.
- The inspected construction uses specified additive-noise structural functions, embeddings and Gaussian error covariance, rather than arbitrary hidden confounding.
- Fitting, covariance invertibility and numerical grid resolution matter. The record does not certify a global maximum over all admissible models or extrapolate the source’s bivariate boundary statement to multiple error terms.
- Squared prediction differences depend on the prediction scale and assessed sample; they are not automatically calibrated probabilities of discrimination or a universal fairness score.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:CausalGraphicalModel
Causal Graphical Model
|
related_match | Evaluates predictor sensitivity using alternative causal graphical models and their counterfactuals. This evaluation procedure is not itself a causal model family or a type of bias. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
original-paper
|
Niki Kilbertus, Philip J. Ball, Matt J. Kusner, Adrian Weller, Ricardo Silva
|
primary paper
full_text
|
ID: https://proceedings.mlr.press/v115/kilbertus20a.html
sha256:f7083ab492a2…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-11T04:57:48Z
Updated: 2026-10-11T04:57:48Z