FCx: An algorithm for finding Feasible Counterfactual Explanations
Kleopatra Markou, Vana Kalogeraki, Dimitrios Gunopulos
cs.LG
Quick verdict
Feasibility-aware conditional VAE adds causal, sparsity and LOF losses; wins feasibility in 7/8 settings vs 9 baselines, but circular metric, no stats or code.
Summary
The paper introduces FCx, a counterfactual-explanation generator built on a conditional variational autoencoder whose training objective couples seven terms: KL regularization, proximity, a hinge validity loss, hard domain-specified causal constraints (e.g., age must not decrease), soft covariance-based penalties for automatically inferred feature relations, an L0/L1 sparsity term, and a Local Outlier Factor (LOF) regularizer computed in the latent space to push counterfactuals into dense regions. Immutable attributes are held fixed. On four public datasets (Adult, KDD-Census, Law School, Folktables) against nine baselines — Mahajan et al., REVISE, C-CHVAE, CEM, DiCE, FACE, VCNet, FCEGAN, CEILS — FCx reports 100% validity everywhere and the highest feasibility in seven of eight unary/binary settings (e.g., 82.92% Adult binary, 94.59% Law School binary), losing only Census-unary to C-CHVAE (98.87% vs 93.96%, though C-CHVAE's validity is 48.44%). An ablation (Table 4) attributes gains to sparsity and LOF, with LOF raising Adult epoch time from 3.2s to ~43s. Key limitations: the feasibility metric mirrors the training constraints (circularity), soft constraints use covariance rather than validated causal structure, results are single-run without significance tests, and no code release is stated.
Review complete
Assessment at a glance
6
risk flags identified
● high 1
● medium 4
● low 1
4 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- FCx is the first method to efficiently generate counterfactual explanations that are simultaneously realistic, low-cost, and feasible.
- Feasibility can be enforced via hard user-specified constraints (age, LSAT) and soft constraints automatically inferred from data through covariance structure.
- Embedding hard/soft causal penalties, L0/L1 sparsity, and latent-space LOF density into a conditional VAE loss yields the highest feasibility in 7 of 8 unary/binary settings across four datasets, with 100% validity everywhere.
- The LOF term eliminates outlier counterfactuals by pushing generation into dense regions; the ablation (Table 4) attributes added feasibility to sparsity and LOF at acceptable training cost.
Strengths
Broad, current benchmark
Four public tabular datasets × 9 baselines including 2024-era SOTA (VCNet, FCEGAN, CEILS, Mahajan et al.) plus CARLA reproductions and DiCE — broader than most CF-feasibility papers (§5.5–5.6, Tables 1–2).
Component-level ablation
Table 4 isolates sparsity and LOF contributions across all four datasets (e.g., Folktables 65.20→71.20→84.32), and reports the compute cost of each component (Adult 3.21→43.44 s/epoch).
Concrete dual treatment of feasibility
Validity never sacrificed
Weaknesses
Circularity between objective and evaluation
Covariance framed as causality
No statistical rigor
Reproducibility gaps
Asymmetric treatment of the one lost setting
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
FCx (arXiv 2609.19383, cs.LG) targets feasibility in counterfactual (CF) explanations: the changes suggested to flip a model's decision should be constructively achievable. The paper claims (1) FCx is 'the first to efficiently generate CFs that are realistic, low-cost and feasible'; (2) feasibility can be enforced through hard domain-specified constraints (age, LSAT) and soft constraints automatically inferred from data; (3) a conditional VAE trained with a seven-term loss (KL, proximity, hinge validity, hard + soft causal penalties, L0/L1 sparsity, latent-space LOF) 'matches state-of-the-art performance ... while guaranteeing feasibility.' Evidence: benchmarks on four public tabular datasets (Adult, KDD-Census, Law School, Folktables) against nine baselines plus an ablation (Table 4).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is FCx and what problem does it solve?
FCx generates counterfactual explanations — minimal input changes that flip a machine-learning model's prediction — while ensuring the suggested changes are actually feasible in the real world (it will not, for example, suggest lowering a person's age). It uses a conditional VAE trained with hard constraints from domain experts (e.g., immutable attributes like race) plus soft constraints automatically inferred from feature covariance in the data.
How well does FCx perform compared to other counterfactual methods?
On four public datasets (Adult, Census-Income KDD, Law School, Folktables) against nine baselines, the paper reports 100% validity everywhere and the highest feasibility scores in 7 of 8 constraint settings (74.91%–94.59%), losing only the Census unary setting to C-CHVAE (98.87% vs 93.96%, though C-CHVAE's validity is only 48.44%).
Is there code available for FCx?
The reviewed arXiv text does not mention a code release. Reproduction requires re-implementing the conditional VAE with its seven-term loss on public tabular datasets, with baselines available through the CARLA and DiCE Python libraries.
What are the main criticisms of the FCx paper?
The main criticisms are: the feasibility evaluation uses the same age/LSAT constraints the model is trained to satisfy (circularity risk); the 'soft causal' constraints are covariance penalties rather than verified causal structure; results are single-run with no statistical significance testing; and the claim of 'guaranteeing' feasibility overstates what a soft loss penalty can achieve.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Large portions of Sections 2, 3.2, 4.1 and 5.1–5.4 were elided in text extraction; the causal-graph inference procedure and dataset/classifier setup could not be fully verified.
- Figures 1–8 are not analyzable from text; training-loss curves (Fig. 8) are taken on trust.
- No code/data availability statement was visible in the extracted text; one may exist in omitted sections.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.