Mitigating Algorithmic Bias in Multiclass CNN Classifications Using Causal Modeling

Min Sik Byun; Wai Kwong Lau; Wendy Wan Yee Hui

REVIEW 3 major objections 5 minor 22 references

Mitigating Algorithmic Bias in Multiclass CNN Classifications Using Causal Modeling

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Causal post-processing cuts gender bias in multiclass emotion AI without hurting accuracy.

desk verdict A clean, simple extension of binary causal-model debiasing to multiclass, but the empirical evidence is too weak to support the headline claims. read the letter →

arxiv 2501.07885 v1 pith:YPLBFDWA submitted 2025-01-14 cs.LG cs.CV

Min Sik Byun , Wendy Wan Yee Hui , Wai Kwong Lau This is my paper · ORCID

classification cs.LGcs.CV

keywords algorithmicbiascausalmodelingfairnessmulticlassclassificationCNNpost-processingemotiongender

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

This paper claims that a post-processing causal-modeling step, previously limited to binary classification, can be extended to multiclass emotion classification and can reduce gender disparities in a CNN's predictions without sacrificing accuracy. The authors turn each emotion class into a one-vs-all binary problem, fit a linear regression per class to measure how much gender shifts the predicted probability, and then subtract that measured shift before choosing the predicted emotion. On a FairFace-derived test set with DeepFace-generated labels, they report gender gaps in true positive rates narrowing for every emotion while overall accuracy stays roughly the same or improves slightly. The underlying point is that fairness and accuracy need not be a trade-off when bias is corrected at the output layer of an already-trained model.

What carries the argument

The central mechanism is a set of one-vs-all causal models: for each emotion class, a linear regression of the CNN's predicted probability on the true class indicator and the protected attribute (gender), following the path-diagram approach used for binary classification by Hui and Lau. The gender regression coefficient provides the estimated effect to remove, and OvA is what lets this binary-classification correction be applied class by class in a multiclass setting. The corrected probabilities are then aggregated by maximum probability to produce a debiased label.

What would settle it

Take the same CNN predictions and apply the same debiasing procedure, but evaluate on a test set labeled by human annotators instead of by DeepFace; if the gender gaps do not shrink — or accuracy drops — on human labels, then the claimed fairness improvement is an artifact of the label generator rather than a genuine property of the CNN classifications.

Watch

Extended reading notes

Core claim

Using the one-vs-all (OvA) technique, the paper expands a binary causal-modeling bias-mitigation method to multiclass classification. For each of four coarse emotion classes (happy, neutral, sad, others), the CNN's predicted probability is regressed on the true class indicator and gender; the gender coefficient, statistically significant in every case, is interpreted as the direct effect of gender on that class's predicted probability. Debiasing removes this coefficient's contribution from the probability, and the class with the highest adjusted probability becomes the predicted emotion. The reported result is that gender gaps in true positive rates shrink across all emotions in both the causal-model training and test sets, while accuracy does not drop — on the causal-model test set it improves from 60.4% to 60.8%.

Load-bearing premise

The emotion labels used as ground truth are generated by the DeepFace pre-trained model rather than by human annotation, and the paper never checks whether those labels are themselves biased by gender or otherwise noisy.

Editorial extensions

If this is right

Post-processing can extend fairness corrections to multiclass deep models without retraining or altering the network architecture.
The debiasing is simple enough to implement with off-the-shelf linear regression, using only the model's predicted probabilities and a protected attribute.
Fairness and accuracy are not necessarily in tension: the paper reports that the correction can slightly improve test accuracy while narrowing gender gaps.
The same one-vs-all causal-model recipe could be applied to other protected attributes, other multiclass tasks, or other black-box classifiers.
Because the correction is applied after training, it is compatible with models whose internal parameters are unavailable or proprietary.

Reading between the lines

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

The reported bias may partly reflect biases in the label generator (DeepFace) rather than the CNN itself, since the ground-truth emotion labels were produced automatically and never validated against human annotation; if so, the debiasing could be correcting the label generator's biases as much as the CNN's.
Because the causal regressions are fit and evaluated on splits of the same automatically labeled test set, the measured fairness improvement may be optimistic; evaluating on an independently human-annotated test set would give a cleaner picture.
All gender coefficients are statistically significant, but that significance is likely inflated by the large sample size (n = 4,381); the practical significance is better judged by the reported gap reductions than by the p-values.
The method treats gender as the only protected attribute and uses coarse emotion groups, so interactions between gender and other protected attributes, or finer emotion distinctions, are not addressed; extending to intersectional fairness would require more elaborate causal models.

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 genuinely new thing here is the one-vs-all construction that lifts the binary causal-model debiasing from Hui and Lau into multiclass problems. That is a legitimate and useful extension, and the authors implemented it cleanly: fit a linear regression per class, subtract the gender coefficient, renormalize by maximum probability, and evaluate on a held-out split. Code and data are public, and the holdout is real—the debiasing coefficients are fit on a CM training set and applied to a separate CM test set. I agree with the reader that this deserves conditional acceptance rather than rejection.

Where the paper falls short is statistical credibility. The ground-truth emotion labels come from DeepFace with no validation against human annotation. That matters because every accuracy and fairness measurement is relative to those labels; if DeepFace is biased, the method is chasing its own reflection. The paper does not address this at all.

The second problem is uncertainty. The before/after comparisons in Tables X and XI have no confidence intervals or significance tests. The stress-test note is right: on the CM test set, the overall accuracy change is 4–5 samples, and the per-emotion TPR gaps are computed from small cell counts. For instance, the female sad TPR drops from 42/86 to 33/86—about 1.9 standard errors—and with four emotions and no multiple-comparison correction, seeing one or two nominally large reductions by chance is plausible. The claim that disparities were reduced 'across all emotions' overstates what the data support. Some reductions, like the 'others' gap on the CM training set, are negligible in magnitude anyway.

The causal assumptions are also simply borrowed from the authors' earlier binary paper without defense. The graph in Fig. 4 treats gender as a direct cause of the predicted probability and of the true class; that may be too simple for emotion classification, but the method does not depend on the graph being exactly right—it is a linear adjustment. Still, the authors should say more about what they are assuming and why.

What the paper does well is show a reproducible, easy-to-implement post-processing recipe that might help practitioners reduce gender gaps in multiclass CNNs. The regression coefficients are significant on the training set, and the test-set reductions, while not individually significant, are mostly in the expected direction. The accuracy does not collapse. That is enough to justify peer review.

I would send this to review, but the revision must include validation of at least a subset of labels against human annotation, confidence intervals or bootstrap estimates for the TPR gaps and accuracy changes, and a more measured conclusion. The method is worth having in the literature; the current evidence base is not.

Referee Report

3 major / 5 minor

Summary. The paper proposes a post-processing debiasing method for multiclass CNN emotion classification. Using the FairFace dataset with emotion labels generated by the DeepFace pretrained model, the authors train a custom CNN and then, for each emotion class, fit an OLS regression of the CNN's predicted probability on gender and the true class dummy (one-vs-all). The estimated gender coefficient is used to adjust the predicted probabilities, and the class with the highest adjusted probability is selected. The authors evaluate the method on an 80/20 split of the test set (the "CM training" and "CM test" sets), reporting overall accuracy and gender-wise true positive rates before and after debiasing. The central claim is that the debiased classifications reduce gender disparities across all emotions without sacrificing overall accuracy.

Significance. If the empirical claim were well supported, the paper would provide a simple, interpretable extension of the binary causal-modeling debiasing approach of Hui and Lau to multiclass CNN outputs, with public code and data. The holdout design is a genuine out-of-sample evaluation: the debiasing coefficients are fitted on the CM training set and then applied to the CM test set. The equations for the debiasing step are explicit and the code is publicly available, which are strengths.

major comments (3)

[§III-C, Table XI] The central claim that gender disparities are reduced 'across all emotions' on the CM test set is not statistically established. The overall accuracy change from 60.4% to 60.8% corresponds to roughly 4-5 samples out of 1,096. For the per-emotion TPR gaps, take the sad class: female TPR changes from 42/86 (0.488) to 33/86 (0.384), a drop of 0.104; the standard error of a proportion near 0.5 with n=86 is about 0.054, so the change is about 1.9 standard errors. The neutral gap reduction from 0.1738 to 0.1046 has a similar order of magnitude. With four emotion classes and no multiple-comparison correction, observing one or two nominally large gap reductions by chance is plausible. The paper should report confidence intervals or significance tests for the gap reductions, and the 'across all emotions' claim should be qualified accordingly.
[§III-A, Tables IV-XI] The ground-truth emotion labels are generated entirely by the DeepFace pretrained model, and the paper provides no validation against human annotations. All accuracy and fairness measurements, as well as the bias coefficients in Tables VI-IX, are computed relative to these automatically generated labels. If DeepFace has systematic gender-dependent errors, the detected 'bias' and the measured fairness improvement may be artifacts of the label generator rather than properties of the CNN. The authors should validate the emotion labels on a human-annotated subset and report per-gender agreement (e.g., Cohen's kappa), or at least discuss this limitation explicitly.
[§III-C, Fig. 4, Eqs. (1)-(2)] The term 'causal modeling' overstates what is done. Equations (1)-(2) are OLS regressions of predicted probabilities on gender and the true class dummy; no structural causal assumptions, confounding control, or counterfactual validation are provided. The debiasing step is a linear post-hoc shift of class probabilities. This is a legitimate post-processing heuristic, but the causal interpretation is not justified by the analysis. The authors should either state the identifying assumptions needed for a causal reading or frame the method as a heuristic adjustment.

minor comments (5)

[Tables VII-IX] The row label 'happy (yneutral)' in Tables VII, VIII, and IX appears to be a copy-paste error; it should read 'neutral (yneutral)', 'sad (ysad)', and 'others (yothers)' respectively.
[Index Terms] The index term 'Casual Modeling' is a typo and should read 'Causal Modeling'.
[§III-C(a)] The sentence 'The overall accuracies for the CM training and test sets are now, respectively, 60.6% and 60.4%, respectively' contains a duplicated 'respectively'; one occurrence should be removed.
[§III-C(c)] The text says 'As noted previously, the beta coefficients for "Male" are statistically significant across all emotions,' but the preceding discussion only reports p-values in the tables without explicitly drawing this conclusion; the cross-reference should be corrected or clarified.
[Data availability] The two Google Colab links are long and may be unstable; the authors should consider providing a persistent repository link or DOI for the code and data.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the debiased CM test evaluation is a genuine holdout, and the self-citation to the authors' prior method [5] is reimplemented rather than assumed.

full rationale

The paper's derivation chain is not circular. The debiasing coefficients in Eqs. (3)-(6) are estimated by linear regression on the CM training set (80% of the original test set, n=4,381) and then applied to the held-out CM test set (n=1,096), which the CNN never saw during training. The TPR-gap reductions in Table XI and Fig. 6 are therefore genuine out-of-sample predictions, not in-sample fits. The CM training-set improvement in Fig. 5 is in-sample, but the paper explicitly calls it the fitted result and then turns to cross-validation for a 'more meaningful assessment,' so it is not presented as a prediction. The method is attributed to the authors' prior work [5] (Hui and Lau), but all coefficients are re-estimated on the present data and the regression equations are stated in full, so the self-citation is a normal method attribution rather than a load-bearing circular premise. The principal caveat is that the emotion labels come from the DeepFace pre-trained model, making the fairness metrics relative to that label generator rather than to human annotation; this is a measurement-validity limitation, not a circularity in the derivation chain.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claim rests on fitted gender coefficients and on the validity of automated labels; no new physical or conceptual entities are introduced.

free parameters (4)

Gender coefficient for happy class = -0.058
Fitted in the linear regression for 'happy' probability (Table VI) and used directly in the debiasing equation (3).
Gender coefficient for neutral class = 0.052
Fitted in the linear regression for 'neutral' probability (Table VII) and used directly in the debiasing equation (4).
Gender coefficient for sad class = 0.025
Fitted in the linear regression for 'sad' probability (Table VIII) and used directly in the debiasing equation (5).
Gender coefficient for others class = 0.026
Fitted in the linear regression for 'others' probability (Table IX) and used directly in the debiasing equation (6).

assumptions (3)

domain assumption Causal graph in Fig. 4: gender a and true emotion y cause predicted probability y_hat through linear paths, with no unmeasured confounders.
Invoked in Section III-C when Eqs. (1)-(2) are specified; without this, the regression coefficient on gender cannot be interpreted as a causal effect.
domain assumption DeepFace-generated labels are a valid ground truth for emotion and for fairness measurement.
Section III-A uses DeepFace to label the whole dataset; the paper never checks against human labels, so this assumption underlies every accuracy and TPR calculation.
domain assumption The one-vs-all calibrated probabilities from [20] are well-calibrated and suitable for linear regression.
Section III-C applies OvA to 'improve multiclass predicted probabilities'; the calibration details are not provided.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Mitigating Algorithmic Bias in Multiclass CNN Classifications Using Causal Modeling." pith.science (2026). https://pith.science/paper/YPLBFDWA

@misc{pith2026250107885,
  author       = {Pith},
  title        = {Pith review of: Mitigating Algorithmic Bias in Multiclass CNN Classifications Using Causal Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YPLBFDWA}},
  note         = {Machine review of arXiv:2501.07885}
}

read the original abstract

This study describes a procedure for applying causal modeling to detect and mitigate algorithmic bias in a multiclass classification problem. The dataset was derived from the FairFace dataset, supplemented with emotional labels generated by the DeepFace pre-trained model. A custom Convolutional Neural Network (CNN) was developed, consisting of four convolutional blocks, followed by fully connected layers and dropout layers to mitigate overfitting. Gender bias was identified in the CNN model's classifications: Females were more likely to be classified as "happy" or "sad," while males were more likely to be classified as "neutral." To address this, the one-vs-all (OvA) technique was applied. A causal model was constructed for each emotion class to adjust the CNN model's predicted class probabilities. The adjusted probabilities for the various classes were then aggregated by selecting the class with the highest probability. The resulting debiased classifications demonstrated enhanced gender fairness across all classes, with negligible impact--or even a slight improvement--on overall accuracy. This study highlights that algorithmic fairness and accuracy are not necessarily trade-offs. All data and code for this study are publicly available for download.

Figures

Figures reproduced from arXiv: 2501.07885 by the authors.

**Figure 1.** A Sample of the Labelled Data. B. The CNN Classification Model The FairFace dataset supplemented with emotion labels from DeepFace was split into training, validation, and test sets, as shown in Table I. The splitting was performed in a stratified manner to maintain the proportion of gender and 1https://colab.research.google.com/drive/1CxRZ7QxX2b09ee0pF3GbwA0 lZ syooeQ?usp=sharing 2https://colab.research.google.com/… view at source ↗

**Figure 2.** Gender Fairness of the CNN Model (CM Training). [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Gender Fairness of the CNN Model (CM Test). [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Causal Model for Binary Classification. [5] [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

**Figure 5.** Figure 5: Gender Fairness of the CNN Model (CM Training, Debiased). [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

**Figure 6.** Figure 6: Gender Fairness of the CNN Model (CM Test, Debiased). [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 16 canonical work pages

[1]

A survey on bias and fairness in machine learning,

N. Mehrabi, F. Morstatter, N. Saxena, K. Lerman, and A. Galstyan, “A survey on bias and fairness in machine learning,” ACM Computing Surveys, https://dl.acm.org/doi/10.1145/3457607. Last accessed 1 December 2024

doi:10.1145/3457607 2024
[2]

Fairness and bias in artificial intelligence: A brief survey of sources, impacts, and mitigation strategies,

E. Ferrara, “Fairness and bias in artificial intelligence: A brief survey of sources, impacts, and mitigation strategies,” arXiv.org, https://arxiv. org/abs/2304.07683. Last accessed 1 December 2024

arXiv 2024
[3]

Mahoney, K

T. Mahoney, K. R. Varshney, and M. Hind, AI Fairness: How to Measure and Reduce Unwanted Bias in Machine Learning, O’Reilly, 2020

work page 2020
[4]

Hardt, E

M. Hardt, E. Price, and N. Srebro, ”Equality of opportunity in supervised learning,” Advances in Neural Information Processing Systems 2016, 29, pp. 3315–3323

work page 2016
[5]

Detecting and mitigating algorithmic bias in binary classification using causal modeling,

W. Hui and W. K. Lau, “Detecting and mitigating algorithmic bias in binary classification using causal modeling,” 4th International Confer- ence on Computer Communication and Information Systems, Phuket, Thailand, February 27-29, 2024. Available: https://doi.org/10.1109/CC CIS63483.2024.00016

work page arXiv 2024
[6]

Putzel and S

P. Putzel and S. Lee, ”Blackbox post-processing for multiclass fairness,”

work page
[7]

Mitigating Nonlinear Algorithmic Bias in Binary Classification,

W. Hui and W. K. Lau, “Mitigating Nonlinear Algorithmic Bias in Binary Classification,” IEEE Conference on Artificial Intelligence, 2024. Available: https://ieeexplore.ieee.org/document/10605518/

work page arXiv 2024
[8]

Data augmentation for discrimination prevention and bias disambiguation,

S. Sharma, Y . Zhang, J. M. R. Aliaga, D. Bouneffouf, V . Muthusamy, and K. R. Varshney, “Data augmentation for discrimination prevention and bias disambiguation,” Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, 2020, pp.358-364

work page 2020

Show all 22 references

[9]

AI fairness via domain adaptation,

N. Joshi and P. Burlina, “AI fairness via domain adaptation,” Available: https://arxiv.org/abs/2104.01109. Last accessed 1 December 2024. 6

2024 arXiv
[10]

Classification with fairness constraints: A meta-algorithm with provable guarantees,

L. E. Celis, L. Huang, V . Keswani, and N. K. Vishnoi, “Classification with fairness constraints: A meta-algorithm with provable guarantees,” Proceedings of the 2019 Conference on Fairness, Accountability, and Transparency, Atlanta, GA, USA, pp. 29–31

2019
[11]

Beutel, J

A. Beutel, J. Chen, Z. Zhao, and E. H. Chi, ”Data decisions and theoretical implications when adversarially learning fair representations,” Proceedings of Workshop on Fairness, Accountability, and Transparency in Machine Learning, Halifax, Canada, August 2017. Available: https...

2017 arXiv
[12]

W. Paul, A. Hadzic, N. Joshi, F. Alajaji, and P. Burlina, ”TARA: Training and representation alteration for AI fairness and domain generalization,” Neural Computation, 34 (3), 2022, 716-753

2022
[13]

Direct and indirect effects,

J. Pearl, “Direct and indirect effects,” Proceedings of the Seventeenth Conference on Uncertainty in Artificial Intelligence, 2001, pp. 411-420

2001
[14]

Pearl, Causality

J. Pearl, Causality. Cambridge University Press, 2009

2009
[15]

Counterfactual fairness,

M. J Kusner, J. Loftus, C. Russell, and R. Silva, “Counterfactual fairness,” Advances in Neural Information Processing Systems, USA, 2017, pp. 4066–4076

2017
[16]

Fairness through causal awareness: Learning causal latent-variable models for biased data,

D. Madras, E. Creager, T. Pitassi, and R. Zemel, “Fairness through causal awareness: Learning causal latent-variable models for biased data,” Proceedings of the Conference on Fairness, Accountability, and Transparency, January 2019, pp. 349-358. Available: https://doi.org/10 ....

2019
[17]

Fairness in algorithmic decision making: An excursion through the lens of causality,

A. Khademi, S. Lee, D. Foley, and V . Honavar, “Fairness in algorithmic decision making: An excursion through the lens of causality,” The World Wide Web Conference, May 2019, pp. 2907-2914

2019
[18]

Available: https://huggingface.co/datasets/HuggingF aceM4/FairFace

FairFace Dataset. Available: https://huggingface.co/datasets/HuggingF aceM4/FairFace
[19]

Available: https://github.com/serengil/deepface

DeepFace. Available: https://github.com/serengil/deepface
[20]

Transforming classifier scores into accurate multiclass probability estimates,

B. Zadrozny and C. Elkan, “Transforming classifier scores into accurate multiclass probability estimates,” Proceedings of the 8th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2002, pp. 694–699. https://doi.org/10.1145/775047.775151

2002
[21]

Zhang and L

Z. Zhang and L. Wang, Advanced Statistics Using R. Granger, IN: ISDSA Press. ISBN: 978-1-946728-01-2. 7
[2022]

Last accessed 4 October 2023

Available: https://arxiv.org/abs/2201.04461. Last accessed 4 October 2023

2023 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

A survey on bias and fairness in machine learning,

N. Mehrabi, F. Morstatter, N. Saxena, K. Lerman, and A. Galstyan, “A survey on bias and fairness in machine learning,” ACM Computing Surveys, https://dl.acm.org/doi/10.1145/3457607. Last accessed 1 December 2024

doi:10.1145/3457607 2024

[2] [2]

Fairness and bias in artificial intelligence: A brief survey of sources, impacts, and mitigation strategies,

E. Ferrara, “Fairness and bias in artificial intelligence: A brief survey of sources, impacts, and mitigation strategies,” arXiv.org, https://arxiv. org/abs/2304.07683. Last accessed 1 December 2024

arXiv 2024

[3] [3]

Mahoney, K

T. Mahoney, K. R. Varshney, and M. Hind, AI Fairness: How to Measure and Reduce Unwanted Bias in Machine Learning, O’Reilly, 2020

work page 2020

[4] [4]

Hardt, E

M. Hardt, E. Price, and N. Srebro, ”Equality of opportunity in supervised learning,” Advances in Neural Information Processing Systems 2016, 29, pp. 3315–3323

work page 2016

[5] [5]

Detecting and mitigating algorithmic bias in binary classification using causal modeling,

W. Hui and W. K. Lau, “Detecting and mitigating algorithmic bias in binary classification using causal modeling,” 4th International Confer- ence on Computer Communication and Information Systems, Phuket, Thailand, February 27-29, 2024. Available: https://doi.org/10.1109/CC CIS63483.2024.00016

work page arXiv 2024

[6] [6]

Putzel and S

P. Putzel and S. Lee, ”Blackbox post-processing for multiclass fairness,”

work page

[7] [7]

Mitigating Nonlinear Algorithmic Bias in Binary Classification,

W. Hui and W. K. Lau, “Mitigating Nonlinear Algorithmic Bias in Binary Classification,” IEEE Conference on Artificial Intelligence, 2024. Available: https://ieeexplore.ieee.org/document/10605518/

work page arXiv 2024

[8] [8]

Data augmentation for discrimination prevention and bias disambiguation,

S. Sharma, Y . Zhang, J. M. R. Aliaga, D. Bouneffouf, V . Muthusamy, and K. R. Varshney, “Data augmentation for discrimination prevention and bias disambiguation,” Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, 2020, pp.358-364

work page 2020