Pith. sign in

REVIEW 4 major objections 5 minor 50 references

Analyzing Fairness of Computer Vision and Natural Language Processing Models

T0 review · 4 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read Stacking bias-mitigation algorithms across pre-, in-, and post-processing can sometimes beat any single algorithm on the accuracy-fairness trade-off.

desk verdict The sequential-vs-single mitigation benchmark is uninterpretable because the protected attribute is entangled with the prediction target in both the CV and NLP setups. read the letter →

arxiv 2412.09900 v3 pith:P35VVK54 submitted 2024-12-13 cs.LG cs.CL

Ahmed Rashed , Abdelkrim Kallich , Mohamed Eltayeb This is my paper · ORCID
classification cs.LGcs.CL
keywords machinelearningfairnessbiasmitigationcomputervisionnaturallanguageprocessingsequentialdemographicparityequalizedoddslibraries
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 tries to establish that applying bias-mitigation algorithms one after another across the machine-learning lifecycle—first at pre-processing, then during training, then on the outputs—can in some cases reduce demographic disparity while preserving model accuracy better than applying any single algorithm. It tests this claim on a computer-vision age-group classifier trained on UTKFace and a natural-language classifier that predicts urgent medical intervention from California Independent Medical Review text, using two fairness libraries. The headline numbers are mixed but, in the paper's telling, favorable on balance: the best sequential pipeline on the CV task cut the maximum demographic-parity difference from 55% to 5% with accuracy at 30.1%, whereas the best single algorithm reached 3.5% disparity only by dropping accuracy to 18%; on the NLP task the sequential pairing lowered disparity from 27.4% to 17.9%, beating the single-algorithm best of 21.4%. The same comparison with the second library on the CV task favored the individual algorithm, so the paper's conclusion is conditional: sequential application can be favorable, not that it always is.

What carries the argument

The central mechanism is the sequential mitigation pipeline: a pre-processing algorithm (Correlation Remover, Reweighing, or Disparate Impact Remover) alters the training data, an in-processing algorithm (Exponentiated Gradient, Adversarial Debiasing, or Learning Fair Representations) adds a fairness constraint during training, and a post-processing algorithm (Threshold Optimizer, Reject Option Classification, or Equalized Odds) adjusts the final predictions. The paper's comparison object is the difference between running one of these alone and running one from each stage in fixed order, with accuracy plus a disparity metric (demographic parity difference, equalized odds difference, or Theil index) measured before and after.

What would settle it

Relabel the California IMR records using only the doctor's findings, with gender and age excluded from the label, retrain the same NLP model, and recompute the demographic-parity and equalized-odds differences; if the disparities collapse or vanish, the reported NLP bias was an artifact of label construction rather than model bias.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is a recipe rather than a theorem: stack one mitigation algorithm at each lifecycle stage—pre-processing to clean the training data, in-processing to constrain the learner, post-processing to adjust decisions—and the stack sometimes dominates its best single component on the accuracy-versus-disparity trade-off. The evidence is three best-case comparisons. On the UTKFace age-group task, the first library's Correlation Remover plus Exponentiated Gradient reduced the largest demographic-parity difference from 55% to 5% and the largest equalized-odds difference from 31.1% to 7.8%, while accuracy fell from 66% to 30.1%; a single Threshold Optimizer achieved slightly lower disparity (3.5% and 3.4%) but at 18% accuracy. On the California IMR text task, the same sequential pairing reduced the largest demographic-parity difference from 27.4% to 17.9% with accuracy at 92.7%, while the single Exponentiated Gradient reached only 21.4% at 95.4% accuracy. The paper also records counterexamples: with the second library on the CV task, individual Disparate Impact Remover reduced the Theil index more than any sequential combination while holding accuracy at 70.9%, and on the NLP task that library's interventions left both fairness and accuracy essentially unchanged.

Load-bearing premise

The NLP experiment assumes the target variable ('whether a patient requires urgent intervention') is defined without reference to gender or age, so that gender-and-age disparities in predictions are the model's bias; the paper's own Section 5 says the target is built from the doctor's findings, patient gender, and age group, which would mean the fairness metrics are measuring the label, not the model.

Editorial extensions

If this is right

  • A practitioner facing a steep fairness-accuracy trade-off should test sequential pipelines before committing to a single aggressive post-processing fix, because the paper's CV results show a stack can reach similar disparity reduction at 30.1% accuracy where the single fix dropped to 18%.
  • Sequential pipelines are not a universal improvement: the second library's CV results show individual application reducing the Theil index from 58.6% to 47.1% while a four-algorithm stack only reached 53.3%, so the choice must be validated per library and task.
  • Fairness reporting that tracks the worst-off sensitive subgroup, as the paper does with 'highest DPD' and 'highest EOD,' can change which mitigation appears best; reporting only average disparity would miss the Threshold Optimizer's collapse in accuracy.
  • Baseline accuracy matters: the NLP model started at 98.1% accuracy, and on such nearly-saturated tasks the second library's stacks changed bias by less than one percentage point, suggesting sequential mitigation has little headroom when bias mainly lives in the label.

Reading between the lines

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

  • A natural next experiment the paper does not run is varying the order of the stages (pre+in, pre+post, in+post) to see whether the benefit comes from stacking per se or from always starting with pre-processing; the paper only tests the single order pre-processing first.
  • If the NLP label is indeed constructed with gender and age, the paper's NLP fairness numbers are better interpreted as measuring the dataset's labeling policy, and a corrected label would be the cleanest test of whether sequential mitigation helps text models at all.
  • The finding that the stack sacrifices less accuracy than the single strong post-processing fix suggests a testable mechanism: sequential interventions spread the fairness burden across stages, so no single stage is pushed hard enough to destroy the task signal; one could verify this by recording each stage's contribution to the disparity reduction.
  • The paper's own future-work direction—extending the pipeline to generative models—can be made concrete by treating the generated distribution as the post-processing output and applying the same re-ranking or threshold adjustments to it, which the existing libraries already support for classifiers.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper reports a comparative study of fairness mitigation applied to a computer vision model (UTKFace age-group classification) and an NLP model (California Independent Medical Review urgent-intervention prediction), using Fairlearn and AIF360. The authors compare applying mitigation algorithms individually at one stage of the ML lifecycle with applying them sequentially across preprocessing, in-processing, and post-processing stages, and they report demographic parity, equalized odds, and other fairness metrics before and after mitigation. The central claim is that sequential application of mitigation algorithms can reduce bias while preserving accuracy in some cases, based on best-case results in Tables 4-8.

Significance. If the reported results were valid, the paper would provide a useful practical comparison of two popular fairness libraries and an interesting empirical answer to whether sequential mitigation across ML lifecycle stages outperforms single-stage mitigation. The topic is relevant, and the use of public datasets and existing fairness libraries is a strength. However, the significance is conditional on the validity of the fairness measurements, and the manuscript has load-bearing confounds in both the CV and NLP analyses, lacks basic experimental reproducibility details (data splits, repeated runs, error bars, model details), and selects best-case outcomes across many algorithms. These issues prevent the results from supporting the central claim as stated.

major comments (4)
  1. [Section 5.1.1 and Tables 4-5] The CV fairness analysis is confounded. Section 5 states that for the CV model, the sensitive features are gender and ethnicity and the target variable is age group, but Section 5.1.1 then reports demographic parity and equalized odds 'for each subgroup of the sensitive feature (age group)', with Tables 4-5 identifying the max-bias subgroup as '30-80y' or '0-18y'. If the protected attribute is age group and the prediction target is age group, then any accurate classifier is definitionally 'biased' under demographic parity and equalized odds, and the reported reductions after mitigation are tautological rather than evidence of meaningful debiasing. If the intended sensitive features are gender and ethnicity, then the age-group subgroup metrics do not measure the stated bias. Either reading invalidates the CV headline numbers on which the conclusion rests.
  2. [Section 5, NLP target definition; Tables 7-8] The NLP target variable is defined in Section 5 as 'whether a patient requires urgent intervention, based on a combination of the doctor's findings, patient gender, and age group.' Since the protected attributes gender and age group are explicitly part of the label construction, the demographic parity and equalized odds differences reported in Tables 7-8 reflect the label-generation process, not model bias. The paper provides no alternative label definition or robustness check that would separate model-induced disparities from label-induced ones. The NLP results therefore cannot support the claim that sequential mitigation reduces bias while preserving accuracy.
  3. [Sections 5.1-5.2, Tables 4-8] No error bars, repeated runs, train/validation/test splits, or model training details are reported. All results are point estimates from what appear to be single runs, and the text repeatedly selects the 'best case scenario' algorithm when summarizing, e.g., Table 4 reports Threshold Optimizer with DPD=3.5% and accuracy=18% while Grid Search gives DPD=53.9% and accuracy=45%. This best-case selection across many algorithm combinations overstates the reliability and generality of the findings. The absence of variance estimates makes it impossible to judge whether differences between individual and sequential application are meaningful.
  4. [Section 7, conclusion] The central claim that 'applying mitigation algorithms in a sequential order can produce favorable results compared to using a single algorithm' is supported only by selected best-case outcomes, such as Correlation Remover + Exponentiated Gradient in Tables 5 and 7. Other sequential combinations perform worse than the best individual algorithm (e.g., Table 5 min-mitigation rows, Table 7 min-mitigation rows). Without a systematic comparison across all combinations, including measures of central tendency and dispersion, the conclusion overgeneralizes from favorable instances.
minor comments (5)
  1. [Table 3] The caption of Table 3 reads 'Results of Fairlearn + Computer Vision...' but the table content is a comparison of the pros and cons of the three libraries; the caption appears to be mislabeled.
  2. [Throughout] Accuracy values are given inconsistently as percentages and as fractions (e.g., 66% vs. 0.66, 70.5% vs. 0.7050, 98.1% vs. 0.981). Please standardize the notation.
  3. [References] Several references appear irrelevant to the claims, including Ref. [13] on CASPaxos and Ref. [14] on the Balmer spectrum, and Ref. [12] duplicates Ref. [44]; please revise the reference list.
  4. [Tables 2-3] The What-If Tool is introduced and listed in Tables 2-3 but is never used in the results; either include the corresponding analysis or remove it from the study description.
  5. [Section 5.2.1] The Fairlearn NLP results report DPD and EOD, but Table 7's 'Algorithm with Min Mitigation' rows report EOD values (58.3%, 41.3%) while the text and table header emphasize DPD; clarify which metric is used for the 'min mitigation' selection and why.

Circularity Check

3 steps flagged · score 6.0 of 10

The CV fairness numbers are definitional: the protected attribute used to compute DPD/EOD is the model's own age-group target. The NLP target is explicitly built from gender and age, and Fairlearn's best mitigation algorithms are evaluated on the exact metrics they optimize, so the headline 'bias reduction' is substantially produced by construction.

  1. self definitional [Section 5 (model setup) and Section 5.1.1 (Fairlearn CV results)]
    "For the CV model, the sensitive features analyzed for potential bias are gender and ethnicity, with the target variable being the prediction of a person's age group. ... In Figure 2, we present the values of the demographic parity and equalized odds metrics for each subgroup of the sensitive feature (age group)."

    The paper states the CV target is age-group prediction, yet the metrics are computed with age group as the sensitive feature. When the protected attribute equals the label, DPD and EOD are confusion-matrix statistics rather than measures of bias against an independent group. For class k, DPD compares the classifier's recall in the true age group k against its false-positive rate in other age groups, so an accurate classifier has a large 'disparity' by construction. The reported accuracy collapse from 66% to 18% after mitigation is the same phenomenon viewed another way: the only way to reduce the metric is to make predictions less correlated with the true age label. The claimed bias reductions therefore follow from the experimental setup itself.

  2. self definitional [Section 5 (NLP model setup)]
    "In the NLP model, the target variable is determining whether a patient requires urgent intervention, based on a combination of the doctor's findings, patient gender, and age group. The sensitive features here are gender and age group."

    The paper defines the NLP target label using the same attributes (gender, age group) that are later treated as sensitive in the fairness analysis. Any demographic-parity or equalized-odds difference can then inherit the label's own dependence on gender and age; it is not a clean measure of model bias. Mitigation algorithms that change predictions to reduce DPD/EOD are partly erasing the label's built-in association with protected attributes, so reported reductions such as DPD 27.4% to 17.9% are confounded and cannot alone support the headline claim that sequential mitigation reduces bias while preserving performance.

1 more flagged steps
  1. self definitional [Section 4 (methodology) and Section 5.1.1 (Fairlearn CV results)]
    "For F airlearn, accuracy was used to evaluate the overall predictive performance, while demographic parity difference served as a fairness metric ... Methods like the Threshold Optimizer adjusted predictions after model training to satisfy fairness criteria precisely ... After applying the mitigation algorithms individually, the highest DPD decreased to 3.5% (best case scenario) with the Threshold Optimizer algorithm."

    Threshold Optimizer is a post-processing method whose objective is to enforce demographic-parity and equalized-odds constraints, and the paper evaluates it with exactly those metrics. Selecting it as the 'Algorithm with Max Mitigation' (DPD 55% to 3.5%) is therefore a self-confirming result: the algorithm is constructed to minimize the target metric, so the reduction is by construction rather than an independent empirical discovery. Exponentiated Gradient similarly incorporates parity constraints during training. The individual-versus-sequential comparison retains some independent content, but the fairness numbers feeding that comparison are partly generated by optimizing the evaluation metric itself.

full rationale

The paper's sequential-versus-individual comparison is not itself circular, and its self-citations ([15], [38]) are descriptive rather than load-bearing. However, the CV arm's central fairness numbers are definitionally tied to the target: Section 5 says the CV target is age-group prediction, while Section 5.1.1 uses age group as the sensitive feature in the DPD/EOD calculations, so the reported disparities are class-confusion artifacts. The NLP arm similarly defines its target as 'based on a combination of the doctor's findings, patient gender, and age group,' making the protected attributes partly responsible for the label. In addition, Fairlearn's best-performing mitigation algorithms are evaluated on the exact metrics they are designed to optimize, so their reductions are partially tautological. These issues do not eliminate all empirical content—the ordering of individual versus sequential performance is a real comparison—but the headline conclusion that mitigation 'reduced bias' is substantially produced by construction. A score of 6 reflects this partial circularity, with the central claim still containing some independent comparative content.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The paper's central claims rest on the correctness of the external libraries, the representativeness of the two datasets, and an unexamined assumption about the NLP label. No new entities or hand-fitted constants are introduced, but the age bin boundaries and the max-subgroup aggregation rule are arbitrary choices that materially affect every reported number.

free parameters (2)
  • Age group boundaries = 0-18, 18-30, 30-80, 80+ years
    Hand-chosen bins for the CV classification target; not justified in the paper, and they determine the sensitive group definitions and all downstream fairness metrics.
  • Subgroup aggregation rule = max DPD/EOD across subgroups
    The paper reports the subgroup with maximum bias as the overall metric (Tables 4-5); this choice affects every comparison but is not justified or analyzed for stability.
assumptions (3)
  • domain assumption The fairness libraries (Fairlearn, AIF360) implement their documented metrics and mitigation algorithms correctly
    The paper relies on library outputs without independent validation (Section 4).
  • ad hoc to paper The California IMR dataset's target label is not itself a function of gender and age, so measured disparities are model-induced
    Section 5 states the target is based on doctor's findings, patient gender and age group; if gender/age influence the label, the fairness analysis is confounded.
  • domain assumption Model training (architectures, splits, hyperparameters) was performed correctly despite being unspecified
    Section 3 describes the datasets but never names the models or training procedure, so the accuracy and fairness numbers cannot be audited.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Analyzing Fairness of Computer Vision and Natural Language Processing Models." pith.science (2026). https://pith.science/paper/P35VVK54

@misc{pith2026241209900,
  author       = {Pith},
  title        = {Pith review of: Analyzing Fairness of Computer Vision and Natural Language Processing Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P35VVK54}},
  note         = {Machine review of arXiv:2412.09900}
}
read the original abstract

Machine learning (ML) algorithms play a critical role in decision-making across various domains, such as healthcare, finance, education, and law enforcement. However, concerns about fairness and bias in these systems have raised significant ethical and social challenges. To address these challenges, this research utilizes two prominent fairness libraries, Fairlearn by Microsoft and AIF360 by IBM. These libraries offer comprehensive frameworks for fairness analysis, providing tools to evaluate fairness metrics, visualize results, and implement bias mitigation algorithms. The study focuses on assessing and mitigating biases for unstructured datasets using Computer Vision (CV) and Natural Language Processing (NLP) models. The primary objective is to present a comparative analysis of the performance of mitigation algorithms from the two fairness libraries. This analysis involves applying the algorithms individually, one at a time, in one of the stages of the ML lifecycle, pre-processing, in-processing, or post-processing, as well as sequentially across more than one stage. The results reveal that some sequential applications improve the performance of mitigation algorithms by effectively reducing bias while maintaining the model's performance. Publicly available datasets from Kaggle were chosen for this research, providing a practical context for evaluating fairness in real-world machine learning workflows.

Figures

Figures reproduced from arXiv: 2412.09900 by the authors.

Figure 1
Figure 1. The ML lifecycle and the application of the pre-processing, in-processing, and post-processing [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The results of applying Fairlearn to the computer vision model with applying the mitigation [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. The results of applying Fairlearn to the computer vision model with applying the mitigation [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: The results of applying AIF360 to the computer vision model with applying the mitigation [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]
Figure 5
Figure 5. Figure 5: The results of applying AIF360 to the computer vision model with applying the mitigation [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]
Figure 6
Figure 6. Figure 6: The results of applying Fairlearn to the NLP model with applying the mitigation algorithms [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: The results of applying Fairlearn to the NLP model with applying the mitigation algorithms [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]
Figure 8
Figure 8. Figure 8: The results of applying AIF360 to the NLP model with applying the mitigation algorithms [PITH_FULL_IMAGE:figures/full_fig_p020_8.png]
Figure 9
Figure 9. Figure 9: The results of applying AIF360 to the NLP model with applying the mitigation algorithms in [PITH_FULL_IMAGE:figures/full_fig_p020_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 11 canonical work pages

  1. [1]

    Amitabha Mukerjee, Rita Biswas, Kalyanmoy Deb, and Amrit P. Mathur. 2002. Multi–objective evolutionary algorithms for the risk–return trade–off in bank loan man- agement. Int. Trans. Oper. Res. 9, 5 (2002), 583–597

    2002

  2. [2]

    Miranda Bogen and Aaron Rieke. 2018. HelpWanted: An Examination of Hiring Algorithms, Equity and Bias. Technical Report. Upturn

    2018

  3. [3]

    Lipton, and Yishay Mansour

    Lee Cohen, Zachary C. Lipton, and Yishay Mansour. 2019. Efficient candidate screening under multiple tests and implications for fairness. arXiv:cs.LG/1905.11361 (2019)

    arXiv 2019

  4. [4]

    Shai Danziger, Jonathan Levav, and Liora Avnaim-Pesso. 2011. Extraneous factors in judicial decisions. Proc. Nat. Acad. Sci. 108, 17 (2011), 6889–6892

    work page 2011

  5. [5]

    Anne O’Keeffe and Michael McCarthy. 2010. The Routledge Handbook of Corpus Linguistics. Routledge

    2010

  6. [6]

    Julia Angwin, Jeff Larson, Surya Mattu, and Lauren Kirchner. 2019. Machine bias: There’s software used across the country to predict future criminals. and it’s bi- ased against blacks. https://www.propublica.org/article/machine-biasrisk-assessments- in-criminal-sentencing

    2019

  7. [7]

    2016.Weapons of Math Destruction: How Big Data Increases Inequal- ity and Threatens Democracy

    Cathy O’Neil. 2016.Weapons of Math Destruction: How Big Data Increases Inequal- ity and Threatens Democracy. Crown Publishing Group, New York, NY

    2016

  8. [8]

    M. A. Madaio, L. Stark, J. Wortman Vaughan, and H. Wallach. Co-Designing Check- lists to Understand Organizational Challenges and Opportunities around Fairness in AI. Chi 2020, pages 1–14, 2020

    2020

Show all 50 references
  1. [9]

    Buolamwini and T

    J. Buolamwini and T. Gebru. Gender shades: Intersectional accuracy disparities in commercial gender classification. In Conference on fairness, accountability and trans- parency, pages 77–91, 2018. 22

    2018

  2. [11]

    Rudin, C

    C. Rudin, C. Wang, and B. Coker. The age of secrecy and unfairness in recidivism prediction. pages 1–46, 2018. URL http://arxiv.org/abs/1811.00731

    2018 arXiv

  3. [12]

    Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Em- beddings

    Tolga Bolukbasi, Kai-Wei Chang, James Zou, Venkatesh Saligrama, Adam Kalai. “Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Em- beddings”, arXiv:1607.06520

    arXiv

  4. [13]

    CASPaxos: Replicated State Machines without logs

    Denis Rystsov. “CASPaxos: Replicated State Machines without logs”. arXiv:1802.07000

    arXiv

  5. [14]

    The Balmer spectrum of rational equivariant cohomology theo- ries

    J.P.C.Greenlees. “The Balmer spectrum of rational equivariant cohomology theo- ries”. arXiv:1706.07868

    arXiv

  6. [15]

    ”Analyzing Fairness of Classification Machine Learning Model with Structured Dataset”

    Ahmed Rashed, Abdelkrim Kallich, Mohamed Eltayeb. ”Analyzing Fairness of Classification Machine Learning Model with Structured Dataset”. arXiv:2412.09896 [cs.LG]

    arXiv

  7. [16]

    Computer Vision Dataset: https://www.kaggle.com/datasets/jangedoo/utkface-new NLP Dataset: https://www.kaggle.com/datasets/prasad22/ca-independent-medical-review

  8. [17]

    ”A survey of privacy risks and mitigation strategies in the Artificial Intelligence life cycle”

    Shahriar S, Allana S, Hazratifard SM, Dara R. ”A survey of privacy risks and mitigation strategies in the Artificial Intelligence life cycle”. IEEE Access. 2023 Jun 19;11:61829-54

    2023

  9. [18]

    Mehrabi, N., Morstatter, F., Saxena, N., Lerman, K., & Galstyan, A. (2021). A Survey on Bias and Fairness in Machine Learning. ACM Computing Surveys, 54(6), 1-35

    2021

  10. [19]

    Suresh, H., & Guttag, J. V. (2021). A Framework for Understanding Unintended Consequences of Machine Learning. Communications of the ACM, 64(8), 62-71

    2021

  11. [20]

    ”Mitigating Disparate Im- pact of Differential Privacy in Federated Learning through Robust Clustering”

    Saber Malekmohammadi, Afaf Taik, Golnoosh Farnadi. ”Mitigating Disparate Im- pact of Differential Privacy in Federated Learning through Robust Clustering”. arXiv:2405.19272 [cs.LG]

    arXiv

  12. [21]

    ”Fairness in Algorithmic Recourse Through the Lens of Substantive Equality of Oppor- tunity”

    Andrew Bell, Joao Fonseca, Carlo Abrate, Francesco Bonchi, Julia Stoyanovich. ”Fairness in Algorithmic Recourse Through the Lens of Substantive Equality of Oppor- tunity”. arXiv:2401.16088 [cs.LG]

    arXiv

  13. [22]

    ”On the (In)Compatibility between Group Fairness and Individual Fairness”

    Shizhou Xu, Thomas Strohmer. ”On the (In)Compatibility between Group Fairness and Individual Fairness”. arXiv:2401.07174 [math.ST]

    arXiv

  14. [23]

    Binns, R. (2018). Fairness in Machine Learning: Lessons from Political Philoso- phy. Proceedings of the 2018 Conference on Fairness, Accountability, and Transparency (F AT)

    2018

  15. [24]

    ”The effects of data preprocessing on probability of default model fairness”

    Di Wu. ”The effects of data preprocessing on probability of default model fairness”. arXiv:2408.15452 [econ.EM]

    arXiv

  16. [25]

    H., Lemoine, B., & Mitchell, M

    Zhang, B. H., Lemoine, B., & Mitchell, M. (2018). Mitigating Unwanted Biases with Adversarial Learning. Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society (AIES)

    2018

  17. [26]

    Chouldechova, A. (2017). Fair Prediction with Disparate Impact: A Study of Bias in Recidivism Prediction Instruments. Big Data, 5(2), 153-163

    2017

  18. [27]

    Angwin, J., Larson, J., Mattu, S., & Kirchner, L. (2016). Machine Bias. ProPublica

    2016

  19. [28]

    Xu, D., Yuan, S., Zhang, L., & Wu, X. (2020). FairGAN: Fairness-aware Genera- tive Adversarial Networks. Proceedings of the 2020 International Joint Conference on Artificial Intelligence (IJCAI)

    2020

  20. [29]

    ”Bridging Interpretability and Robustness Using LIME-Guided Model Refinement”

    Navid Nayyem, Abdullah Rakin, Longwei Wang. ”Bridging Interpretability and Robustness Using LIME-Guided Model Refinement”. arXiv:2412.18952 [cs.LG] 23

    arXiv

  21. [30]

    M., & Lee, S.-I

    Lundberg, S. M., & Lee, S.-I. (2017). A Unified Approach to Interpreting Model Pre- dictions. Proceedings of the 31st Conference on Neural Information Processing Systems (NeurIPS)

    2017

  22. [31]

    Buolamwini, J., & Gebru, T. (2018). Gender Shades: Intersectional Accuracy Dis- parities in Commercial Gender Classification. Proceedings of the 2018 Conference on Fairness, Accountability, and Transparency (F AT)

    2018

  23. [32]

    Kearns, M., Neel, S., Roth, A., & Wu, Z. S. (2018). Preventing Fairness Ger- rymandering: Auditing and Learning for Subgroup Fairness. Proceedings of the 35th International Conference on Machine Learning (ICML)

    2018

  24. [33]

    Barocas, S., Hardt, M., & Narayanan, A. (2019). Fairness and Accountability in Machine Learning. Proceedings of the 2019 Conference on Fairness, Accountability, and Transparency

    2019

  25. [34]

    Angwin, J., Larson, J., Mattu, S., & Kirchner, L. (2016). Machine Bias: There’s Software Used Across the Country to Predict Future Criminals. And it’s Biased Against Blacks. ProPublica

    2016

  26. [35]

    Fairlearn by Microsoft. (n.d.). Retrieved from https://fairlearn.org/

  27. [36]

    AIF360 by IBM. (n.d.). Retrieved from https://aif360.mybluemix.net/

  28. [37]

    What-If Tool by Google. (n.d.). Retrieved from https://github.com/google/tf-what- if

  29. [38]

    Github of the project: https://github.com/mohammad2012191/Fairness-in-Machine- Learning-Identifying-and-Mitigation-of-Bias/

  30. [39]

    Equality of opportunity in supervised learning

    Moritz Hardt, Eric Price, and Nati Srebro. Equality of opportunity in supervised learning. In NeurIPS, 3315–3323. 2016. URL: https://proceedings.neurips.cc/paper/2016/hash/9d2682367c3935defcb1f9e247a97c0d- Abstract.html

    2016

  31. [40]

    Does the end justify the means? on the moral justification of fairness-aware machine learning

    Hilde Weerts, Lamb` er Royakkers, and Mykola Pechenizkiy. Does the end justify the means? on the moral justification of fairness-aware machine learning. arXiv preprint arXiv:2202.08536, 2022

    2022 arXiv

  32. [41]

    The unfairness of fair ma- chine learning: levelling down and strict egalitarianism by default

    Brent Mittelstadt, Sandra Wachter, and Chris Russell. The unfairness of fair ma- chine learning: levelling down and strict egalitarianism by default. arXiv preprint arXiv:2302.02404, 2023

    2023 arXiv

  33. [42]

    Barocas, S., Hardt, M., & Narayanan, A. (2017). Fairness and machine learning. Online resource

    2017

  34. [43]

    Bellamy, R. K. E., et al. (2019). AI Fairness 360: An extensible toolkit for detecting, understanding, and mitigating unwanted algorithmic bias. IBM Journal of Research and Development. DOI:10.1147/JRD.2019.2942287

    2019

  35. [44]

    Bolukbasi, T., et al. (2016). Man is to computer programmer as woman is to home- maker? Debiasing word embeddings. NeurIPS 2016. DOI:10.48550/arXiv.1607.06520

    2016 doi arXiv

  36. [45]

    A., et al

    Friedler, S. A., et al. (2019). A comparative study of fairness-enhancing interven- tions in machine learning. Proceedings of F AT 2019. DOI:10.48550/arXiv.1802.04422

    2019 doi arXiv

  37. [46]

    Mehrabi, N., et al. (2021). A survey on bias and fairness in machine learning. ACM Computing Surveys. DOI:10.1145/3457607

    2021 doi

  38. [47]

    Rajkomar, A., et al. (2018). Ensuring fairness in machine learning to advance health equity. Annals of Internal Medicine. DOI:10.7326/M18-1990

    2018 doi

  39. [48]

    Xu, D., et al. (2018). FairGAN: Fairness-aware generative adversarial networks. 2018 IEEE Big Data. DOI:10.1109/BigData.2018.8622626

    2018

  40. [49]

    B., et al

    Arrieta, A. B., et al. (2020). Explainable artificial intelligence (XAI): Concepts, taxonomies, opportunities, and challenges toward responsible AI. Information Fusion. 24 DOI:10.1016/j.inffus.2019.12.012

    2020 doi

  41. [50]

    Doshi-Velez, F., & Kim, B. (2017). Towards a rigorous science of interpretable machine learning. arXiv preprint. DOI:10.48550/arXiv.1702.08608 25

    2017 doi arXiv

  42. [2019]

    URL http://arxiv.org/abs/1907.02227

    1907 arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.