A Fairness Metric Jumps 700% Under Biased Label Noise That Barely Touches Accuracy, a 22-Model Benchmark Finds

A University of Amsterdam benchmark, published in the Proceedings of the VLDB Endowment for VLDB 2026, corrupted 15 tabular datasets with six kinds of label noise and ran 22 classification models against each. The headline asymmetry: models tolerate random label errors up to 40% with only a modest hit to performance, but as little as 7% biased noise, errors concentrated in one subgroup, cuts ResNet's AUC by 4.4% on average, and on the ACS Unemployment dataset, 30% subgroup-targeted noise raises a fairness gap for the minority group by up to 700% while accuracy barely moves.

What the benchmark actually tested

The paper is "Fault Lines: Benchmarking the Impact of Label Data Quality on ML Robustness and Fairness," by David Jackson, Paul Groth and Hazar Harmouch, all at the University of Amsterdam. It appears in the Proceedings of the VLDB Endowment, Volume 19, Issue 4, pages 670 to 683 (DOI 10.14778/3785297.3785308), tied to the VLDB 2026 conference. Code and artifacts are released at github.com/fault-lines/fault_lines_benchmark.

The 15 binary classification datasets are reused from TableShift (Gardner, Popovic and Schmidt, arXiv:2312.07577, NeurIPS 2023), ranging from FICO HELOC (2,220 rows, predicting credit delinquency) up to Public Coverage (4,006,249 rows, by race), and spanning healthcare, finance and social outcome domains. Each dataset carries a designated sensitive subgroup, such as race, age, poverty status or veteran status, that the paper uses to inject noise and to measure fairness.

Instead of only the random label flips most prior noise research uses, the authors define six noise strategies: Random Noise, Class-Conditional Noise, Conditional Feature Noise, Correlated Feature Noise, Concatenated Feature Noise, and Temporal or Contextual Noise. For most of the paper's headline results, they combine Conditional Feature Noise with Class-Conditional Noise into what they call biased noise: labels flip only inside a chosen subgroup, and at a different rate depending on the true class.

The 22 models span linear baselines (SVM, logistic regression), gradient boosting (XGBoost, LightGBM, CatBoost), feedforward and attention-based neural networks (MLP, ResNet, FT Transformer, TabTransformer, SAINT), optimization methods built specifically to resist noise (a noise-resistant GBDT variant, DRO, Group DRO, IRM), domain adaptation methods (DANN, MixUp, VREX) and three explicitly fairness-aware reductions (FairEG, FairGS, FairRS). Four fairness metrics are tracked throughout: Equality of Opportunity, Demographic Parity, Equalized Odds and Predictive Parity.

Random noise: models shrug it off until they suddenly don't

Under purely random label flips, ranging from 0% to 50% across all 15 datasets, the strongest models barely move. The noise-resistant GBDT variant goes from 85.35% accuracy and 82.96 AUC clean to 84.26% accuracy and 78.97 AUC at 30 to 40% noise. CatBoost and FT Transformer show a similar pattern. The authors ran ANOVA and Tukey's HSD across the noise levels and found no statistically significant degradation up to 40% noise (p = 0.670, Tukey mean difference -0.0591, adjusted p > 0.698). Performance only collapses at 50% noise, the point where the signal-to-noise ratio hits 1:1 (p < 0.01).

Weaker baselines fare worse: SVM's best-case accuracy drops 5.32 percentage points from clean to high-noise conditions, and MixUp, a domain-adaptation method, starts so far below the pack on clean data (77.11% accuracy) that its noise tolerance barely matters in practice. Even average-trial performance across hyperparameter settings, not just the best trial, showed FT Transformer degrading 2.22% in accuracy from 0% to 30% noise, four times the 0.8% drop seen in its best-trial numbers, a sign that transformer robustness depends heavily on tuning rather than being a fixed property of the architecture.

Random noise is not harmless to fairness either, even though it is blind to group membership. In the College Scorecard dataset, an SVM's Equalized Odds disparity between veterans and non-veterans nearly doubled, from 0.065 at 0% noise to 0.124 at 30% noise, despite the flips being applied with no regard to veteran status at all. The mechanism: when a subgroup already has a lower base rate for the positive class, symmetric random noise corrupts a larger share of that subgroup's few positive labels, and the effect compounds.

Biased noise: small, targeted errors do outsized damage

For the biased-noise experiments, the authors apply a 30% flip rate inside a chosen sensitive subgroup only, leaving the rest of each dataset untouched. Because only a slice of each dataset is affected, the resulting dataset-level noise rate averages about 7%, ranging from 5.7% on the Sepsis dataset to 9.99% on the Voting dataset. Across 12 representative classifiers, that noise produced significant increases in Equality of Opportunity (mean difference 0.216, p < 0.05), Demographic Parity (mean difference up to 0.143, p < 0.05) and Equalized Odds ratio (mean difference up to 0.348, p < 0.05). Predictive Parity moved unevenly: some models shifted by a mean of 0.102 (p < 0.05), others by only 0.0002, a split the paper ties directly to architecture.

Model family mattered more than it did under random noise. ResNet showed an 0.254 swing in Equalized Odds ratio between clean and noisy conditions, flagging it as highly sensitive to biased noise. TabTransformer shifted only 0.037 on the same metric, but that apparent robustness is partly an artifact of comparison: models like MixUp that also show small shifts start from a much lower clean-data AUC, so "more stable" is not the same as "more useful."

Fairness-aware models were not a reliable fix. FairEG, the best-performing of the three, held its accuracy drop to 0.3 to 0.8% and its AUC drop to 0.8 to 3.6% under biased noise, but its clean-data AUC already ran 23.97% below CatBoost's, a cost the stability numbers alone do not show. FairGS, constrained to Equalized Odds, saw its own Equality of Opportunity rise 169% and its Equalized Odds ratio rise 173% under the same noise. FairRS fared worst: Equalized Odds ratio up more than 300%, AUC down 4.6%, often landing below plain boosting models it was meant to improve on.

One genuine surprise: FT Transformer, the strongest transformer baseline, beat XGBoost on 10 of 13 datasets under biased noise, with 33% smaller average AUC degradation. The tradeoff is stability: FT Transformer's accuracy varied 6.2% across hyperparameter trials versus 2.03% for XGBoost, so the advantage is real but harder to land without careful tuning.

The ACS Unemployment case: a fairness gap that grows 700% while accuracy holds

The paper's closest look is at the ACS Unemployment dataset, using race as the primary sensitive attribute and migration status as a second. On clean data, TabTransformer and FT Transformer reach near-perfect fairness (Equality of Opportunity near 0.0), MixUp and Label Group DRO sit close behind (0.00 to 0.013), MLP and ResNet are moderate (Equalized Odds ratio 0.084 to 0.164), and logistic regression starts with the highest baseline disparity of the group (0.160 to 0.165).

Apply 30% biased noise to the targeted subgroup, and that ordering breaks apart unevenly. Logistic regression and ResNet deteriorate most, with Equality of Opportunity climbing into the 0.353 to 0.693 range across their best-performing trials, and logistic regression's worst trial reaching 0.790. TabTransformer and FT Transformer lose their near-perfect fairness too, though less severely, settling into a Demographic Parity difference of 0.0 to 0.144. Label Group DRO is the sharpest illustration of the paper's central claim: its accuracy stayed within roughly 97% of its clean-data level across conditions, while its fairness metrics still rose by 0.897 to 0.938, a large, hidden failure that an accuracy-only evaluation would miss entirely.

Switch from subgroup-targeted noise to correlated noise (flipping labels only where two features intersect, covering just 4.39% of the dataset here), and the disconnect gets starker still: CatBoost's accuracy barely moves, about a 0.12% change, while Predictive Parity and Equality of Opportunity for the minority group rise by up to 700%. Methods built specifically to resist exactly this kind of corruption made it worse in this setting instead: VREX's Equalized Odds reached 47.9%. Concatenated noise, which covers a broader 35.50% of the dataset, produces a more conventional roughly 7% AUC drop alongside subgroup harm. And fairness constraints only worked when they targeted the actual corrupted attribute: when a constraint protected one group but noise hit an unconstrained one, bias simply moved there, in one case a 40% rise in Equalized Odds for an unprotected minority race group.

Where this sits next to Cleanlab's own label-error research

The paper is explicit about what it adds on top of earlier label-quality work. It contrasts itself directly with Northcutt, Athalye and Mueller's "Pervasive Label Errors in Test Sets Destabilize Machine Learning Benchmarks" (arXiv:2103.14749, NeurIPS 2021), the confident-learning research that underlies Cleanlab's label-error detection tooling and estimated at least 3.3% error rates across 10 widely used benchmark test sets. Fault Lines notes that earlier line of work "focused primarily on accuracy," while its own contribution is to show that subgroup-specific, feature-correlated noise exacerbates fairness disparities well beyond what random label errors produce, a gap that a confident-learning-style detector tuned mainly to catch mislabeled individual examples does not automatically close.

The paper also cites earlier work (Guha et al., and Ni et al.) finding that automated data cleaning pipelines can themselves introduce new errors or quietly damage fairness while fixing accuracy, and its own results echo that: in a handful of cases, models trained on data with under 10% random noise showed small, not statistically significant, accuracy gains over models trained on clean data, a reminder that "more cleaning" and "better data" are not automatically the same instruction.

What this means for a team shipping on noisy labels

Stress-testing a pipeline against random label flips is not stress-testing it against the kind of noise most likely to occur in practice: humans and collection processes tend to make correlated mistakes, not uniformly random ones. A validation suite that only checks accuracy under randomly corrupted labels can report green while a biased-noise failure mode sits undetected underneath it.

Accuracy and AUC are not sufficient fairness proxies, full stop. The paper's own Label Group DRO and ACS Unemployment correlated-noise results both show a model holding performance steady while a fairness metric moves by 70 to over 90 percentage points in the other direction. Any evaluation pipeline that reports a single performance number without a parallel fairness metric, on the actual sensitive subgroups in the data, cannot see this failure mode at all.

Fairness-aware training is not a substitute for checking which attribute the constraint actually protects. FairGS and FairRS both degraded sharply under biased noise despite being designed for fairness, and constraints that protected one group left an unconstrained group's bias to grow freely elsewhere. The paper's practitioner table is blunt on this point: sample and review labels from sensitive subgroups directly, prioritize data quality because model-side fixes cannot fully compensate for poor labels, and report multiple fairness metrics alongside predictive performance, not one or the other.

Check it yourself

The paper, its dataset table and the evaluation code are public:

# Read the paper's noise taxonomy (Section 3.1), dataset table
# (Table 1), model list (Table 2) and the ACS Unemployment deep
# dive (Section 4.5, Figure 7)
open https://www.vldb.org/pvldb/vol19/p670-jackson.pdf
open https://doi.org/10.14778/3785297.3785308

# Clone the released benchmark and evaluation framework
git clone https://github.com/fault-lines/fault_lines_benchmark
cd fault_lines_benchmark

# The underlying 15 TableShift datasets and their domain shifts:
open https://github.com/mlfoundations/tableshift
open https://arxiv.org/abs/2312.07577

# For the confident-learning label-error research this paper
# contrasts itself against (and that underlies Cleanlab's tooling):
open https://arxiv.org/abs/2103.14749
open https://labelerrors.com

What would prove this wrong

The core asymmetry, that a given overall noise rate does far more damage when concentrated in a subgroup than when spread randomly, is already demonstrated on 15 datasets and 22 models with reported statistical significance (p < 0.05 on the main fairness-metric shifts). The open question is generalization: all 15 datasets are tabular, binary-classification, and drawn from one benchmark family (TableShift). If an independent replication on a different set of tabular datasets, or on multi-class or regression tasks, which the authors themselves flag as untested, finds that biased noise at 5 to 10% does not produce materially larger fairness or performance damage than random noise at the same overall rate, the claim that this asymmetry is a general property of biased label noise, rather than an artifact of this specific benchmark's datasets and subgroup choices, would not hold up. A second, narrower test: if a data-cleaning or label-error-detection pipeline (confident-learning-based or otherwise) is shown to catch the specific correlated and concatenated noise patterns in Section 4.4 at a detection rate comparable to its rate on random noise, without being told in advance which feature or subgroup was targeted, that would undercut the paper's implicit claim that today's cleaning tools are poorly matched to this failure mode.

Sources

1. David Jackson, Paul Groth, Hazar Harmouch, "Fault Lines: Benchmarking the Impact of Label Data Quality on ML Robustness and Fairness," Proceedings of the VLDB Endowment, Vol. 19, No. 4, pp. 670-683, VLDB 2026. DOI: 10.14778/3785297.3785308. Full text: vldb.org/pvldb/vol19/p670-jackson.pdf.
2. Fault Lines benchmark code and artifacts, fault-lines/fault_lines_benchmark, accessed October 2026. github.com/fault-lines/fault_lines_benchmark.
3. Josh Gardner, Zoran Popovic, Ludwig Schmidt, "Benchmarking Distribution Shift in Tabular Data with TableShift," arXiv:2312.07577, submitted December 12, 2023, NeurIPS 2023 Datasets and Benchmarks Track (source of the 15 datasets reused by Fault Lines). arxiv.org/abs/2312.07577.
4. Curtis G. Northcutt, Anish Athalye, Jonas Mueller, "Pervasive Label Errors in Test Sets Destabilize Machine Learning Benchmarks," arXiv:2103.14749, submitted March 26, 2021, NeurIPS 2021 Datasets and Benchmarks Track (cited by Fault Lines as the accuracy-focused prior work it extends to fairness; underlies Cleanlab's confident-learning tooling). arxiv.org/abs/2103.14749.