Fed-ReMasker: Federated Tabular Imputation under Feature-Level Missingness
Ioannis Papathanail, Rooholla Poursoleymani, Lubnaa Abdur Rahman, Stavroula Georgia Mougiakakou
cs.LG · cs.AI
Quick verdict
Applies ReMasker via FedAvg to federated feature-level imputation; rigorous MCAR benchmark shows clear wins, but novelty is modest and privacy untested.
Summary
This paper adapts the ReMasker masked autoencoder to federated learning (Fed-ReMasker) for tabular imputation when entire features are unobserved at some clients, a setting the authors argue prior federated imputation work largely ignores. The method itself is a straightforward composition: clients train ReMasker locally and parameters are aggregated with FedAvg. The contribution is primarily an extensive benchmark: 324 homogeneous scenarios (nine synthetic and real datasets, missing-feature ratios 0–0.3, missing-value ratios 0.1–0.5, 3–10 clients) plus 36 heterogeneous scenarios (quantity, missingness, and feature-distribution skew) and a centralized reference. Fed-ReMasker achieves the lowest NRMSE in 93.2% of value-level and 96.7% of feature-level scenarios, with paired margins over Cafe (closest baseline) of 0.018/0.026; Fed-MIWAE falls below mean imputation on entirely missing features. FedProx and FedAdam add no benefit (FedAvg best or tied in 29/30 cells), and federated training trails the pooled-data model by about 3.0% on average, rising to 7.2% at 10 clients. Statistical reporting is careful (dataset-level paired t-intervals, seed variability), but the homogeneous grid rests on a single partition seed, evaluation is MCAR-only, datasets are fully-observed subsets of at least 9,200 samples, and no privacy mechanisms or downstream-task effects are assessed. Code and configurations are publicly released.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Fed-ReMasker achieves the lowest NRMSE in 93.2% of 324 value-level and 96.7% of 243 feature-level homogeneous scenarios across nine datasets, with paired margins of 0.018 (value-level) and 0.026 (feature-level) over the closest baseline Cafe.
Strengths
First systematic benchmark of feature-level federated imputation
Section IV-B defines 324 homogeneous scenarios (9 datasets × 4 MFR × 3 MVR × 3 K), 36 heterogeneous scenarios across four conditions, and 126 centralized-reference configurations, with value- and feature-level NRMSE reported separately — a gap the paper correctly identifies in FedHF-Impute's evaluation (Sec. II-B), which excludes unobserved-feature positions.
Unusually careful statistical reporting
Paired differences use 95% t-intervals computed across datasets (n=9) to respect within-dataset dependence (Sec. IV-D); all 30 exceptions to the win-rate claim are enumerated and bounded (max deficit 0.0058 value-level, 0.0044 feature-level); seed SDs up to ±0.028 (PhysioNet, feature-distribution skew) contextualize small differences in Table II.
Honest accounting of the cost of federation
Useful negative results on federated optimizers
Strong leakage control and reproducibility
Weaknesses
Minimal methodological novelty
Headline win rates rest on a single partition seed
Baseline tuning asymmetry
Realism gaps between motivation and evaluation
No privacy or downstream-task validation
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
Fed-ReMasker (arXiv 2609.28105, cs.LG) targets federated tabular imputation under feature-level missingness: entire variables unobserved at some clients. It claims (1) a federated adaptation of ReMasker, (2) the first systematic benchmark of this setting spanning value- and feature-level missingness, and (3) that plain FedAvg suffices, with error within 3.0% of a centralized pooled-data model on average. The scope is deliberately narrow: horizontal FL with a shared schema, MCAR masks (except one data-dependent condition), numeric features only, no privacy mechanisms. Within that scope, the claims are precisely quantified and mostly supported.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 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 feature-level missingness in federated learning?
It is when a participating site has no observations at all for a given variable — for example a hospital that never records a particular lab measurement — as opposed to value-level missingness, where only individual entries of a variable are missing. Fed-ReMasker lets such a site impute that entire feature using a shared model whose parameters were trained partly by sites that do observe it.
How much accuracy does Fed-ReMasker lose compared to centralizing all data?
On average about 3.0% higher NRMSE than a centralized ReMasker trained on the pooled data across 126 configurations. The gap is 1.2% with 3 clients, 3.2% with 5, and 7.2% with 10 clients, and under the feature-distribution skew condition the paired difference versus centralized training was not statistically distinguishable.
Which methods does Fed-ReMasker compare against and on what data?
Mean imputation, Cafe, Fed-MIWAE, and FedHF-Impute, evaluated on six synthetic datasets with known causal structure (20–100 features) plus three real datasets — Codon, PhysioNet, and NHANES — under varying missing-feature ratios (0–0.3), missing-value ratios (0.1–0.5), and client counts (3, 5, 10), with a separate client-heterogeneity study on the real datasets.
Is code available for Fed-ReMasker?
Yes. The authors state that code, configurations, and scripts to reproduce every table, figure, and reported statistic are publicly available at github.com/AIHNlab/Fed-ReMasker.
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.
- Figures 1–3 are referenced in the text but only describable from captions; per-scenario NRMSE curves and the PCA partition visualizations could not be checked against the claims.
- Table bold/underline markers were flattened in extraction; best/second-best attributions per cell were inferred from the numeric values, not the formatting.
- The abstract in the metadata is truncated mid-sentence ('comes within 3.0% on avera…'), but the full-text conclusion supplies the complete claim.
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.