MAXCFU
Evaluation Methods AIMech:MaxCFU DRAFT
Aliases: Optimization-based counterfactual-fairness sensitivity
A multivariate sensitivity algorithm that uses gradient ascent through fitted additive-noise models to search for large counterfactual prediction discrepancies under bounded correlations of structural error terms.
Identity Notes & Concept Scope
MAXCFU is the named optimization-based tool for multivariate confounding, distinct from the bivariate grid procedure, its CFU discrepancy and the counterfactual-fairness criterion. The record describes the published search procedure, not a certified optimizer. 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
2 Claims
Mechanism
#method-scope
Fit the working model and its residual-based predictor once. Parameterize candidate error correlations through a transformed matrix product, fit the alternative model for each candidate, propagate counterfactual features, and measure squared prediction changes using the original predictor and working-model residual transform. Algorithm 1 uses minibatches and automatic differentiation to ascend this discrepancy with respect to the correlation parameters, then evaluates it on the full data.
Limitation
#optimization-limits
The supplement identifies nested local likelihood fitting and a nonconvex sensitivity objective as computational obstacles. Its matrix counterexample shows that setting all allowed correlations to the same bound can produce an invalid covariance structure; it does not supply a general global-optimization guarantee.
⚠️ Curated Limitations & Negative Results
- The published off-diagonal tanh transformation gives a symmetric unit-diagonal matrix with bounded entries but explicitly does not guarantee positive semidefiniteness. A resulting matrix is therefore not automatically a valid correlation matrix.
- Nested fitting and the outer search are nonconvex; local optimization, initialization and differentiation through fitting matter. No certified global worst-case value is asserted.
- Sensitivity is conditional on the assumed structural model, predictor, intervention, confounding budget and sample. It does not identify the true hidden confounding or establish overall fairness.
- The sparse clique parameterization is discussed but not exercised in the paper’s experiments, and its expressivity and numerical constraints are not certified here.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:CausalGraphicalModel
Causal Graphical Model
|
related_match | Uses counterfactuals from parameterized causal graphical models to assess prediction sensitivity. The evaluation algorithm and the model class are related, without an equivalence or subclass assertion. |
Evidence Sources & Primary Literature
2 Sources| 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…
|
supplement
|
Niki Kilbertus, Philip J. Ball, Matt J. Kusner, Adrian Weller, Ricardo Silva
|
primary paper
full_text
|
ID: https://proceedings.mlr.press/v115/kilbertus20a.html
sha256:238a9527530e…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-11T04:57:48Z
Updated: 2026-10-11T04:57:48Z
Source File:
data/curated/MaxCFU.yaml