GFLC: Graph-based Fairness-aware Label Correction for Fair Classification

Kallol Roy; Modar Sulaiman

REVIEW 4 major objections 7 minor 6 references

GFLC: Graph-based Fairness-aware Label Correction for Fair Classification

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper proposes Graph-based Fairness-aware Label Correction (GFLC), which scores training instances by prediction confidence, Ricci-flow-optimized graph Laplacian disagreement, and demographic-parity gain, then flips the…

desk verdict A plausible new label-correction pipeline, but the headline performance advantage is not established because a single tuned GFLC run is compared against 50-run averages of the baseline. read the letter →

arxiv 2506.15620 v1 pith:DJIAIUTB submitted 2025-06-18 cs.LG cs.AI

Modar Sulaiman , Kallol Roy This is my paper

classification cs.LGcs.AI

keywords fairnessinmachinelearningwithnoisylabelslabelcorrectiondemographicparitygraphLaplacianForman-RiccicurvatureRicciflowk-NN

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper proposes Graph-based Fairness-aware Label Correction (GFLC), a method for fixing label noise in binary classification while preserving demographic parity. GFLC scores each training instance by a combination of classifier uncertainty, disagreement with neighbors on a k-nearest-neighbor graph whose edge weights have been evolved by discrete Ricci flow, and the estimated gain in demographic parity from flipping the instance's label. It then flips the highest-scoring positive and negative candidates to bring each sensitive group's positive rate within a tolerance of the overall prevalence. Experiments on the bank account fraud dataset show GFLC outperforming the Fair-OBNC baseline: AUC rises from 0.813 to 0.874 at 5% noise, and demographic parity stays near 1.0 across decision thresholds while the baseline degrades sharply. The claim is that fairness-aware label correction can improve both accuracy and fairness with respect to the closest baseline.

What carries the argument

The central mechanism is a three-term per-instance score: score_i = α(1−M_i) + β L_i + γ ΔDP_i. Here (1−M_i) is largest for low-confidence predictions; L_i = Σ_{j∈N(i)} w_{ij}(y_i−y_j)^2 is the graph Laplacian disagreement on the k-NN graph, with edge weights updated by discrete Ricci flow (simplified Forman–Ricci curvature, $w^{{(t+1)}}$_{ij}=max($w^{{(t)}}$_{ij}+η F(e_{uv})^{(t)}, ε)); ΔDP_i is the change in the demographic parity ratio if label i were flipped. The score ranks candidates, and group-wise flip counts K+ and K− are chosen so each group's positive rate falls within prevalence bounds [P(y)(1−D), P(y)(1+D)]. The Laplacian term with Ricci-flow weights is meant to suppress spurious cross-group edges while the fairness term pushes toward parity.

What would settle it

Run GFLC on a dataset with clean labels whose group positive rates are known to differ (e.g., 10% positive in group A and 40% in group B), inject group-dependent noise, then compare the corrected labels to ground truth. If the corrected labels are less accurate than the noisy labels, or the final classifier's accuracy is lower than a model trained on the noisy labels without correction, the parity-preserving target is the wrong one and the central claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that group-dependent label noise can be corrected without sacrificing fairness by scoring instances with a combined measure: prediction confidence (margin term), graph Laplacian disagreement with neighbors on a k-NN graph with Ricci-flow-evolved edge weights, and the demographic-parity improvement that flipping each label would produce. The authors report that on the bank account fraud dataset (Variant II), GFLC achieves AUC 0.874 vs 0.813 for Fair-OBNC at 5% noise, 0.843 vs 0.783 at 10%, and 0.799 vs 0.752 at 20%, while the demographic parity ratio stays near 1.0 across thresholds where Fair-OBNC falls toward 0.05 or 0. They interpret this as an improved performance-fairness trade-off, particularly at high noise rates.

Load-bearing premise

The method assumes clean labels satisfy demographic parity, so forcing equal group positive rates is the correct target; on real data where group base rates legitimately differ, the fairness incentive would distort labels and reduce accuracy.

Editorial extensions

If this is right

At 5%, 10%, and 20% group-dependent label noise, GFLC reports AUC of 0.874, 0.843, and 0.799 versus 0.813, 0.783, and 0.752 for Fair-OBNC on the bank account fraud dataset.
GFLC keeps the demographic parity ratio near 1.0 across the evaluated decision thresholds, while Fair-OBNC drops toward 0.05 (5% noise) and 0 (20% noise) at high thresholds.
Equal opportunity and equalized odds differences remain around 0.05 for GFLC across thresholds, indicating that the parity-focused correction does not degrade other fairness criteria on this dataset.
The method's benefit persists at higher noise: at 20% noise the performance-fairness trade-off remains favorable even though TPR behavior differs from the 5% and 10% cases.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A testable separation: train the classifier with only the margin and Laplacian terms (gamma=0) to isolate how much of the AUC gain comes from Ricci-flow reweighting versus the parity incentive; the paper does not ablate this.
The D=0.05 disparity tolerance is a fixed knob; in deployments where acceptable disparity is larger or smaller, the number of flips K+ and K− changes, so sensitivity to D would show how much accuracy the parity target costs.
If clean ground-truth group positive rates are unknown, GFLC's fairness target is not identifiable from data alone; a practitioner would need external knowledge that parity is the correct target before applying it.
The graph construction and Ricci flow operate on raw features; on non-Euclidean feature spaces, a different distance (e.g., cosine or radial basis) would change edge weights and could change the ranking of candidates.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The short version: GFLC is a sensible synthesis of known ingredients—margin-based confidence, graph Laplacian smoothing with Ricci-flow-adjusted weights, and a demographic-parity flip incentive. The combination is new, and the paper documents the algorithm completely enough to reimplement, including a worked derivation of the simplified Forman curvature. It also positions itself honestly against Fair-OBNC, which is the right baseline. So the contribution is real, though incremental: a pipeline, not a new principle.

The experiments do not yet support the headline claim of significant improvements. The stress test is correct and lands. Section 5.2 fixes GFLC's hyperparameters and reports a single run; Section 5.3 averages Fair-OBNC over 50 randomly sampled configurations. So the Table 1 AUC gains (0.061, 0.060, 0.047) compare one favorably configured run to an average that may include poor configurations. There are no error bars, no significance tests, and no variance for GFLC. That asymmetry alone would require a rewrite of the empirical section.

The fairness half of the trade-off is also partly enforced by construction. The flip-determination logic in Eqs. (13)-(15) targets D=0.05 demographic parity on the training labels, so the near-1.0 DP ratios in Figures 2c/6c are expected. What remains to be shown is whether the corrected labels improve the model on independent metrics—AUC and the equal-opportunity/equalized-odds plots—against equivalent effort. The current protocol cannot answer that. A no-correction control and at least one or two more datasets are needed. The clean-label DP assumption in Section 5.1.1 is also worth stating explicitly: the protocol shuffles the sensitive attribute before noise injection, which makes clean labels group-balanced, so the D constraint is appropriate here. Real-world clean labels need not satisfy that, and the paper should say when the method can be expected to transfer.

Minor points: the ordering sentence in Section 4 says scores are sorted descending and then ascending; that is a wording slip, not a substantive flaw. Hyperparameters alpha, beta, gamma, D, k, eta, and ricci_iter are all hand-set with no sensitivity analysis, which matters because the method has several interacting knobs. And the code is promised only upon acceptance, making the single-run results hard to check.

Who this is for: people working on fairness plus label noise, and on graph-based label correction. It deserves a serious referee, but with major revisions expected: rerun GFLC under the same 50-configuration protocol or otherwise report variance, add baselines and a no-correction control, release code, and discuss the clean-label DP assumption. As it stands, my verdict is conditional-to-negative on the empirical claims, not on the idea.

Referee Report

4 major / 7 minor

Summary. The paper proposes GFLC, a label-correction method for binary classification under group-dependent label noise. GFLC combines three scoring components: a prediction-margin term from an ensemble model, a graph-Laplacian regularizer computed on a k-NN graph whose edge weights are evolved by a discrete Ricci flow using Forman--Ricci curvature, and a demographic-parity incentive that scores how much flipping each label would change the DP ratio. The number of flips is determined by a rule that forces each sensitive group's positive rate into a band around the overall prevalence, controlled by a disparity tolerance D. The authors evaluate GFLC on the Bank Account Fraud Variant II dataset at 5%, 10%, and 20% group-dependent noise, comparing AUC and threshold-dependent fairness/performance metrics against Fair-OBNC. The paper reports AUC gains of 0.061, 0.060, and 0.047 over the baseline and DP ratios near 1.0 for GFLC across thresholds.

Significance. If the empirical claims were properly supported, the paper would address a timely and important problem: correcting labels while maintaining fairness under group-dependent noise. The proposed combination of Ricci-flow-optimized graph Laplacians with a fairness-aware correction score is novel relative to the cited literature, and the appendix provides a clear algebraic derivation of the simplified Forman curvature. The paper also gives credit to its closest baseline and uses a real, large-scale tabular dataset. However, the current manuscript does not substantiate the central claim of significant improvements: the evaluation is based on a single dataset, a single baseline, and an asymmetric comparison with no statistical support, and the demographic-parity benefit is largely enforced by construction. As a result, the significance of the contribution cannot be assessed from the evidence presented.

major comments (4)

[Section 5.2, Section 5.3, Table 1] The headline AUC gains in Table 1 are computed under an asymmetric experimental protocol: GFLC is evaluated with a single fixed hyperparameter configuration (Section 5.2), while Fair-OBNC is averaged over 50 randomly sampled configurations (Section 5.3). No standard deviation, confidence interval, or significance test is reported for GFLC. The reported gains of 0.061, 0.060, and 0.047 AUC are therefore not interpretable as reliable improvements; they may reflect a hand-picked favorable configuration rather than a systematic advantage. This is load-bearing for the paper's central claim of significant performance gains.
[Section 4.4 (Eqs. 13–15), Figures 2c, 4c, 6c] The demographic parity advantage of GFLC is substantially by construction. The flip determination logic in Eqs. (13)–(15) computes K+ and K- so that each group's positive rate is forced into [P(y)(1-D), P(y)(1+D)] with D=0.05. Consequently, the corrected training labels are guaranteed to satisfy demographic parity within tolerance, and the near-1.0 DP ratios in Figures 2c, 4c, and 6c are a direct consequence of this enforcement rather than an empirical outcome of the scoring function. The paper describes this as an improvement in the performance–fairness trade-off, but no comparison is made at controlled DP levels; the only unconstrained metric is AUC, which is already affected by the asymmetry identified in the previous comment.
[Section 5.1.1] The experimental protocol shuffles the sensitive attribute column before noise injection, guaranteeing that the clean labels are independent of the sensitive attribute and hence satisfy demographic parity by construction. The paper does not discuss or evaluate the behavior of GFLC when clean labels are not group-balanced, which is a likely real-world condition. In that case, the fairness incentive in Eq. (11) and the flip rules in Eqs. (13)–(15) would actively distort labels toward an inappropriate parity target and could reduce accuracy. The paper should either restrict its claims to the group-balanced clean-label setting or provide an analysis of the unbalanced case.
[Section 5.1 and Section 5.3] The empirical evaluation is limited to a single dataset (Bank Account Fraud Variant II) and a single baseline (Fair-OBNC). Given the paper's general claims about fair classification under group-dependent label noise, the absence of additional datasets or baselines makes it difficult to assess the method's generalizability. In addition, no ablation study isolates the contributions of the margin term, the Ricci-flow-optimized Laplacian, and the fairness term, so it is unclear which component drives any observed improvement.

minor comments (7)

[Section 4, after Eq. (7)] The text says instances are scored in descending order and then immediately states that instances are ordered by score in ascending order; the intended ordering for selecting the top candidates should be clarified.
[Throughout] There are several typographical errors, including 'Prposed' in the Section 4 heading, 'addtion' in Section 2, and 'obserave' in Section 3.4; these should be corrected.
[Section 5.2] The paper should state explicitly how many independent runs or seeds were used for GFLC; if the reported numbers come from a single run, this should be acknowledged and variance should be reported.
[Section 5.1] The manuscript should specify how the one-million-instance dataset is split into training, validation, and test sets, and whether the full dataset is used.
[Appendix A] The 'Final Adjustment' that multiplies the normalized curvature by w_uv is a heuristic modification; it should be clearly labeled as such rather than presented as a purely algebraic simplification of the original Forman formula.
[Section 4.2, Eq. (9)] The factor of 5 in the threshold objective TPR(τ) - 5·FPR(τ) is not justified; a sentence explaining this choice would support reproducibility.
[Figures 1–6] The curves for Fair-OBNC are averaged over 50 runs, but the figures show no error bands or confidence intervals; adding variance information would make the threshold-dependent comparisons much more informative.

Circularity Check

1 steps flagged · score 4.0 of 10

The demographic-parity result is largely the algorithm's own objective restated; the independent AUC claim is weakened by an asymmetric evaluation protocol, but no load-bearing self-citation or deeper circularity is present.

self definitional [Section 4.3-4.4 (Eqs. 11-15), Section 5.2 (D=0.05), Section 5.4.2, Section 6.2 (Fig. 2c)]
"The fairness term measures the impact of flipping label y_i on demographic parity ... and is defined as follows: ∆DP_i = DP_new − DP_original ... The resulting fairness term, ∆DP_i ... (i) ∆DP_i > 0 means that flipping y_i improves demographic parity ... First, we define the overall positive rate ... P_lower = P(y)·(1−D) and P_upper = P(y)·(1+D), where D ∈ [0,1] is the disparity tolerance parameter. ... To compute the Demographic Parity metric (Equation 16), we calculate the ratio between the lowest and highest predicted prevalence (Equation 12)."

The correction objective and the reported success metric are the same demographic-parity ratio (Eq. 12). Eqs. 13-15 explicitly push corrected training labels toward group positive rates within P(y)(1±D), with D=0.05 fixed in Sec. 5.2, and the score term γ∆DP_i prioritizes flips that increase this same ratio. Because the held-out test set is IID by construction (sensitive attribute shuffled before noise injection, Sec. 5.1.1), the near-1.0 DP curves in Fig. 2c are largely the propagated effect of the training-time parity constraint rather than an independent empirical discovery. The independent content lies in the AUC comparison and the secondary Equal Opportunity / Equalized Odds metrics; hence this is partial, not total, circularity.

full rationale

No load-bearing self-citation or imported uniqueness theorem is present; the only self-citation (Sulaiman et al., 2025) appears in a routine list of fairness methods and does not support any central premise. The margin and graph-Laplacian components are heuristics whose effects are evaluated independently, and the Ricci-flow construction is not used to derive the headline results by definition. The main circular aspect is that demographic parity is both optimized (Eqs. 11-15 with D=0.05) and evaluated (Eqs. 12/16), so the fairness half of the claimed trade-off is substantially by construction. The AUC half is independent in principle, but the comparison is not apples-to-apples: GFLC is reported as one fixed configuration (Sec. 5.2) while Fair-OBNC is averaged over 50 random configurations (Sec. 5.3); this is an experimental-protocol concern rather than circularity. The clean-label DP assumption from the shuffled sensitive attribute (Sec. 5.1.1) is a correctness risk/limitation, not a circular step. Overall, a moderate partial-circularity score is warranted, not a high one, because the central claim retains independent empirical content outside the DP metric.

Assumptions & free parameters 8 free parameters · 4 assumptions · 0 invented entities

The method depends on a set of hand-chosen hyperparameters, a graph-smoothness assumption, and a synthetic noise protocol where the clean data is deliberately made group-balanced. No new physical or model entities are introduced; the algorithmic components are standard.

free parameters (8)

alpha = 0.2
Weight of the margin term in score_i (Equation 7), set by hand in Section 5.2 with no sensitivity analysis.
beta = 0.6
Weight of the graph Laplacian term in Equation (7), set by hand in Section 5.2.
gamma = 0.2
Weight of the fairness incentive in Equation (7), set by hand in Section 5.2.
k = 10
Number of nearest neighbors for the k-NN graph (Section 3.1), chosen in Section 5.2.
eta = 0.1
Ricci flow step size in the edge weight update rule (Equation 6), set in Section 3.4.
ricci_iter = 2
Number of Ricci flow iterations applied to edge weights, set in Section 5.2.
D = 0.05
Disparity tolerance that determines flip counts K+ and K- and effectively forces group prevalence ratios toward 1 (Section 4.4).
tau+ and tau- = optimized via ROC
Margin thresholds selected by maximizing TPR - 5*FPR on the noisy training data (Equation 9), so they are fit to the observed labels.

assumptions (4)

domain assumption Graph smoothness: nearby points in feature space share labels.
The Laplacian term (Equation 8) treats high local label disagreement as evidence of noise; this is the standard manifold smoothness assumption adopted in Section 4.1.
domain assumption Clean labels satisfy demographic parity across sensitive groups because the experimental protocol shuffles the sensitive attribute before noise injection.
Section 5.1.1 removes correlations between the sensitive column, labels, and features, so the pre-noise labels have equal positive rates; the method then forces parity on corrected labels, which is only valid if true labels are group-balanced.
standard math The discretized Forman-Ricci curvature and discrete Ricci flow from the cited literature are valid for graph edge reweighting.
Sections 3.2-3.4 import these tools as background; the paper does not prove their applicability beyond the cited references.
ad hoc to paper The 'final adjustment' of multiplying the normalized curvature by w_uv is a reasonable heuristic for edge-centric weight updates.
Appendix A introduces this multiplication with the justification that strong edges should dominate, not by a derivation from Forman's original formula.

how reviews work

0 comments

Cite this review

Pith. "Pith review of GFLC: Graph-based Fairness-aware Label Correction for Fair Classification." pith.science (2026). https://pith.science/paper/DJIAIUTB

@misc{pith2026250615620,
  author       = {Pith},
  title        = {Pith review of: GFLC: Graph-based Fairness-aware Label Correction for Fair Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DJIAIUTB}},
  note         = {Machine review of arXiv:2506.15620}
}

read the original abstract

Fairness in machine learning (ML) has a critical importance for building trustworthy machine learning system as artificial intelligence (AI) systems increasingly impact various aspects of society, including healthcare decisions and legal judgments. Moreover, numerous studies demonstrate evidence of unfair outcomes in ML and the need for more robust fairness-aware methods. However, the data we use to train and develop debiasing techniques often contains biased and noisy labels. As a result, the label bias in the training data affects model performance and misrepresents the fairness of classifiers during testing. To tackle this problem, our paper presents Graph-based Fairness-aware Label Correction (GFLC), an efficient method for correcting label noise while preserving demographic parity in datasets. In particular, our approach combines three key components: prediction confidence measure, graph-based regularization through Ricci-flow-optimized graph Laplacians, and explicit demographic parity incentives. Our experimental findings show the effectiveness of our proposed approach and show significant improvements in the trade-off between performance and fairness metrics compared to the baseline.

Figures

Figures reproduced from arXiv: 2506.15620 by the authors.

**Figure 1.** Performance Metrics Comparison Across Different Thresholds at 5% Noise Rate Demographic Parity (Figure 2c): The demographic parity ratio results in Figure 2c reveal the most dramatic performance differences between the methods, where higher values indicate better demographic parity. GFLC maintains exceptional demographic parity with ratios consistently near 1.0 across all threshold values, indicating optimal demogra… view at source ↗

**Figure 2.** Fairness Metrics Comparison Across Different Thresholds at 5% Noise Rate In contrast, Fair-OBNC exhibits significantly worse fairness performance, with peaks reaching approximately 0.18 at threshold values around 0.05 and 0.3. Moreover, FairOBNC shows good Equal Opportunity around 0.04 approximately at threshold 0.11. Furthermore, Fair-OBNC shows a problematic bimodal distribution with substantial fairness degradat… view at source ↗

**Figure 3.** Performance Metrics Comparison Across Different Thresholds at 10% Noise Rate (a) Equal Opportunity ↓ (b) Equalized odds ↓ (c) Demographic Parity ↑ [PITH_FULL_IMAGE:figures/full_fig_p019_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Fairness Metrics Comparison Across Different Thresholds at 10% Noise Rate [PITH_FULL_IMAGE:figures/full_fig_p019_4.png]

**Figure 5.** Figure 5: Performance Metrics Comparison Across Different Thresholds at 20% Noise Rate [PITH_FULL_IMAGE:figures/full_fig_p020_5.png]

**Figure 6.** Figure 6: Fairness Metrics Comparison Across Different Thresholds at 20% Noise Rate Demographic Parity (Figure 6c): The demographic parity ratio shown in Figure 6c indicates that the GFLC method maintains exceptional demographic parity, with ratios consistently close to 1.0 acro…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

6 extracted references · 3 canonical work pages

[1]

Algan, G., Ulusoy, I. (2021). Image classification with deep learning in the presence of noisy labels: A survey,Knowledge-Based Systems215, 106771. Ando, R., Zhang, T. (2006). Learning on graph with laplacian regularization,Advances in neural information processing systems19. Canalli, Y ., Braida, F., Alvim, L., Zimbr ˜ao, G. (2024). Fair transition loss:...

work page 2021
[26]

Weber, M., Saucan, E., Jost, J. (2017). Characterizing complex networks with forman-ricci cur- vature and associated geometric flows,Journal of Complex Networks5(4), 527–550. Wu, S., Gong, M., Han, B., Liu, Y ., Liu, T. (2022). Fair classification with instance-dependent label noise,Conference on Causal Learning and Reasoning, PMLR, pp. 927–943. Zafar, M....

work page 2017
[262]

Sulaiman, M., Mahmoud, N. T. A., Roy, K. (2025). Advancing equal opportunity fairness and group robustness through group-level cost-sensitive deep learning.,Baltic Journal of Modern Computing13(1). Tjandra, D., Wiens, J. (2023). Leveraging an alignment set in tackling instance-dependent label noise,Conference on Health, Inference, and Learning, PMLR, pp. ...

arXiv 2025
[2336]

Bochner’s method for cell complexes and combinatorial ricci curvature,Discrete & Computational Geometry29, 323–374

Forman (2003). Bochner’s method for cell complexes and combinatorial ricci curvature,Discrete & Computational Geometry29, 323–374. Hamilton, R. S. (1982). Three-manifolds with positive ricci curvature,Journal of Differential geometry17(2), 255–306. Han, B., Yao, Q., Liu, T., Niu, G., Tsang, I. W., Kwok, J. T., Sugiyama, M. (2020). A survey of label-noise ...

arXiv 2003
[8650]

Silva, I. O. e., Soares, C., Sousa, I., Ghani, R. (2023). Systematic analysis of the impact of label noise correction on ml fairness,Australasian Joint Conference on Artificial Intelligence, Springer, pp. 173–184. Sreejith, R., Mohanraj, K., Jost, J., Saucan, E., Samal, A. (2016). Forman curvature for complex networks,Journal of Statistical Mechanics: The...

work page 2023
[9984]

G., Athalye, A., Mueller, J

Northcutt, C. G., Athalye, A., Mueller, J. (2021). Pervasive label errors in test sets destabilize machine learning benchmarks,arXiv preprint arXiv:2103.14749. Saleiro, P., Kuester, B., Hinkson, L., London, J., Stevens, A., Anisfeld, A., Rodolfa, K. T., Ghani, R. (2018). Aequitas: A bias and fairness audit toolkit,arXiv preprint arXiv:1811.05577. Samal, A...

arXiv 2021

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Algan, G., Ulusoy, I. (2021). Image classification with deep learning in the presence of noisy labels: A survey,Knowledge-Based Systems215, 106771. Ando, R., Zhang, T. (2006). Learning on graph with laplacian regularization,Advances in neural information processing systems19. Canalli, Y ., Braida, F., Alvim, L., Zimbr ˜ao, G. (2024). Fair transition loss:...

work page 2021

[2] [26]

Weber, M., Saucan, E., Jost, J. (2017). Characterizing complex networks with forman-ricci cur- vature and associated geometric flows,Journal of Complex Networks5(4), 527–550. Wu, S., Gong, M., Han, B., Liu, Y ., Liu, T. (2022). Fair classification with instance-dependent label noise,Conference on Causal Learning and Reasoning, PMLR, pp. 927–943. Zafar, M....

work page 2017

[3] [262]

Sulaiman, M., Mahmoud, N. T. A., Roy, K. (2025). Advancing equal opportunity fairness and group robustness through group-level cost-sensitive deep learning.,Baltic Journal of Modern Computing13(1). Tjandra, D., Wiens, J. (2023). Leveraging an alignment set in tackling instance-dependent label noise,Conference on Health, Inference, and Learning, PMLR, pp. ...

arXiv 2025

[4] [2336]

Bochner’s method for cell complexes and combinatorial ricci curvature,Discrete & Computational Geometry29, 323–374

Forman (2003). Bochner’s method for cell complexes and combinatorial ricci curvature,Discrete & Computational Geometry29, 323–374. Hamilton, R. S. (1982). Three-manifolds with positive ricci curvature,Journal of Differential geometry17(2), 255–306. Han, B., Yao, Q., Liu, T., Niu, G., Tsang, I. W., Kwok, J. T., Sugiyama, M. (2020). A survey of label-noise ...

arXiv 2003

[5] [8650]

Silva, I. O. e., Soares, C., Sousa, I., Ghani, R. (2023). Systematic analysis of the impact of label noise correction on ml fairness,Australasian Joint Conference on Artificial Intelligence, Springer, pp. 173–184. Sreejith, R., Mohanraj, K., Jost, J., Saucan, E., Samal, A. (2016). Forman curvature for complex networks,Journal of Statistical Mechanics: The...

work page 2023

[6] [9984]

G., Athalye, A., Mueller, J

Northcutt, C. G., Athalye, A., Mueller, J. (2021). Pervasive label errors in test sets destabilize machine learning benchmarks,arXiv preprint arXiv:2103.14749. Saleiro, P., Kuester, B., Hinkson, L., London, J., Stevens, A., Anisfeld, A., Rodolfa, K. T., Ghani, R. (2018). Aequitas: A bias and fairness audit toolkit,arXiv preprint arXiv:1811.05577. Samal, A...

arXiv 2021