Pith. sign in

REVIEW 4 major objections 5 minor 38 references

Fairness in Semi-supervised Learning: Unlabeled Data Help to Reduce Discrimination

T0 review · 4 major / 5 minor · reviewed 2026-08-27 · deepseek-v4-flash

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

Pith's one-line read This paper argues that pseudo-labeling unlabeled data, re-sampling to balance protected-group by label cells, and bagging the resulting models reduces demographic-parity discrimination without sacrificing accuracy, and supports the…

desk verdict A sensible, new-enough pre-processing pipeline for fair SSL, but the title oversells it: the theory is a conditional restatement and the experiments are mixed without error bars. read the letter →

arxiv 2009.12040 v1 pith:HWWPMDN5 submitted 2020-09-25 cs.LG stat.ML

Tao Zhang , Tianqing Zhu , Jing Li , Mengde Han , Wanlei Zhou , Philip S. Yu This is my paper · ORCID
classification cs.LGstat.ML
keywords fairnesssemi-supervisedlearningpseudolabelingdemographicparityre-samplingbaggingbias-variance-noisedecompositiondiscrimination
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 sets out to turn semi-supervised learning into a fairness tool. It claims that when unlabeled data are pseudo-labeled, re-sampled so that every protected-attribute-by-label cell has equal size, and then bagged, the resulting ensemble has lower demographic-parity discrimination than training on the labeled data alone, without a corresponding drop in accuracy. The authors support this with a bias-variance-noise decomposition in which unlabeled data reduce discrimination in variance (larger effective training set plus bagging) while increasing discrimination in noise (pseudo-label mistakes), and a condition under which the reduction wins. Experiments on Health, Bank, and Adult data, plus synthetic data, show the pipeline beating uniform and preferential sampling baselines on the trade-off. The paper's contribution is a mechanism, not just a metric: it shows where unlabeled data enter the fairness calculus.

What carries the argument

The load-bearing object is the fairness-enhanced sampling (FS) pipeline in the pre-processing phase: (1) pseudo labeling trains a model on the labeled split, predicts labels for sampled unlabeled data, and concatenates them into a new training set; (2) fair re-sampling splits that set into the four protected-attribute by label groups $G_{PP}$, $G_{UP}$, $G_{PN}$, $G_{UN}$ and draws the same number of points from each, oversampling small cells; (3) fair ensemble learning baggs $K$ base models trained on independently re-sampled fair datasets. The theoretical carrier is a bias-variance-noise decomposition in which discrimination is the inter-group difference of expected losses; the proof rests on Lemma 2, which says the discrimination learning curve falls as training-set size grows, plus the claim that bias discrimination is unchanged when pseudo-labeled data are added.

What would settle it

On a dataset where the pseudo-labeling model is known to err more for one protected group, intentionally flip a fixed fraction of pseudo labels in that group, run the full FS pipeline with the paper's parameters ($K=200$, split rate $s=0.8$, sample size equal to the minimum cell size), and measure demographic parity on a balanced test set. If measured discrimination rises as the flip fraction increases faster than the variance-reduction estimate predicts, the theorem's condition is falsified in that regime.

Watch

Extended reading notes

Core claim

The central discovery is that discrimination in a classifier, measured as demographic-parity disparity $|\Pr(\hat{Y}=1|A=0) - \Pr(\hat{Y}=1|A=1)|$, can be decomposed into bias, variance, and noise components, and that semi-supervised data act on only two of them. Adding pseudo-labeled unlabeled data leaves the bias component unchanged, reduces the variance component because the effective training set grows and bagging averages away model instability, and adds a noise component because pseudo labels can be wrong in a group-dependent way. Theorem 1 states that the framework reduces discrimination whenever the variance reduction (including bagging's effect) is at least as large as the pseudo-label noise it introduces; the experiments then show this condition is met for logistic regression and SVM on the tested datasets. The paper also distinguishes representation discrimination, where minority regions of the input space are underrepresented, from evaluation bias, showing that the same fair model looks less fair on a skewed test set than on a balanced one.

Load-bearing premise

The proof of Theorem 1 assumes that adding pseudo-labeled data leaves the bias component of discrimination unchanged ($|\bar{B}_{sl}| - |\bar{B}_{ssl}| = 0$); if the extra data shifts the model's typical prediction for either group, the variance-versus-noise accounting no longer covers the full change in discrimination.

Editorial extensions

If this is right

  • More unlabeled data in the pipeline generally raises accuracy: on the Adult dataset with logistic regression, accuracy climbs from about 0.728 at sample ratio $\rho=0.1$ to about 0.745 at $\rho=1$, while discrimination falls or stabilizes after an initial rise.
  • Ensemble size matters: accuracy and discrimination converge once $K$ is near 200; before that, both fluctuate, so the fairness benefit is statistical rather than guaranteed for any single fair dataset.
  • Because the framework is a pre-processing method, it can be stacked with in-processing or post-processing fairness methods to attack other sources of discrimination, as the paper itself notes.
  • The same model appears much fairer on a balanced test set than on a skewed one, implying that measured discrimination includes evaluation bias and that fair evaluation needs balanced or weighted test sets.
  • The choice of sample size trades off generalization against pseudo-label noise: a larger sample size can improve accuracy and reduce bias-related discrimination, but it can also add noise-related discrimination.

Reading between the lines

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

  • If pseudo-label noise is group-dependent and large, the theorem's condition $|\bar{V}_{a,sl}| - |\bar{V}_{a,ssl}| - \bar{N}_{a,p} \ge 0$ can fail; a practical test is to estimate each term on a validation set before committing to a sample ratio.
  • Forcing all four cells to equal size changes the training distribution's class priors, so downstream probability estimates would need recalibration; the paper does not discuss this effect.
  • Confidence-thresholded pseudo-labeling or a noise-robust loss would shrink the added noise term $\bar{N}_{a,p}$ and widen the regime where the method provably helps, a natural extension the paper leaves implicit.
  • The paper notes in its conclusion that labeled and unlabeled data are assumed to share a distribution; under covariate shift, pseudo labels inherit systematic errors, so the framework would need a shift correction.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Request a human review

A listed scientist reviews the paper for a fee and the review publishes here regardless of verdict. See the reviewers or get listed.

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 proposes a pre-processing framework for fair semi-supervised learning (SSL), called fairness-enhanced sampling (FS), which combines pseudo-labeling of unlabeled data, re-sampling to equalize the number of samples across protected-attribute/label groups, and bagging over the resulting fair datasets. The authors claim that unlabeled data can improve the accuracy–discrimination trade-off by reducing representation discrimination, and they support this with a bias–variance–noise decomposition of discrimination (Theorem 1) and experiments on three real-world datasets and synthetic data. The theoretical part is intended to show that discrimination decreases when the reduction in variance from more training data and ensembling outweighs the additional noise from pseudo-labeling. Experiments report accuracy and demographic-parity discrimination for logistic regression and SVM under different sample ratios, ensemble sizes, and sample sizes, and compare against original, uniform-sampling, and preferential-sampling baselines.

Significance. If the central claim were fully established, the paper would make a useful practical contribution: a simple pre-processing method that exploits abundant unlabeled data while controlling demographic-parity discrimination, with a decomposition-based explanation connecting data quantity, ensembling, and pseudo-label noise. The framework is clearly described and the experimental comparisons cover relevant baselines. However, the theoretical result is conditional on an unverified bias-invariance assumption and an unverified variance-reduction condition, and the experimental evidence does not consistently support the abstract’s universal phrasing: in several settings discrimination increases with the unlabeled-data ratio. The paper’s value currently lies more in the proposed pipeline and empirical exploration than in a validated theoretical guarantee. The bias–variance–noise framing is a useful lens, but the paper does not close the gap between the framework’s stated conditions and its conclusions.

major comments (4)
  1. [Section 4, 'Discrimination in Bias' paragraph] The proof of Theorem 1 asserts that |\bar B(f)_sl| - |\bar B(f)_ssl| = 0 because "bias ... only depends on the model." This is not justified: \bar B_a(f) in Eq. (14) is an expectation over training sets D of B(y_m, x, a), and the main prediction y_m is itself defined by an expectation over D. Adding pseudo-labeled data and re-sampling to equalize group sizes changes the training-set distribution, so y_m and hence group-wise bias can shift. The equality needs a proof or an explicit condition under which it holds; as written it is a load-bearing unsupported assumption.
  2. [Section 4, 'Discrimination in Variance' paragraph and Theorem 1 statement] The variance-reduction step relies on Lemma 2, which is an imported general asymptotic inverse-power-law learning curve from [17], not a verified property of the specific re-sampling and pseudo-labeling scheme. The proof does not show that the sufficient condition (|\bar V_a(f)_sl| - |\bar V_a(f)_ssl|) - \bar N_{a,p} >= 0 actually holds for the proposed FS framework; the theorem is concluded by restating this condition. The paper should either prove this inequality under stated assumptions for FS or provide an empirical check that the condition is satisfied in the experiment settings.
  3. [Section 5.1.2, Figure 2 and accompanying text] The experimental results do not support the universal claim that unlabeled data reduce discrimination. In the panel (f) (SVM-Adult) the text states that the discrimination level "is steady and has a slight increase in SVM," and other panels also show non-monotone or increasing discrimination with the sample ratio. The abstract and Theorem 1 claim a general reduction; the paper should either restrict the claim to conditions where the variance-reduction condition holds or present evidence identifying when FS reduces discrimination. As it stands, the empirical evaluation is consistent with a conditional statement, not the paper's stated conclusion.
  4. [Section 4, Lemma 1 and Eq. (13)] The decomposition \bar\Gamma = |(\bar B_0 - \bar B_1) + (\bar V_0 - \bar V_1) + (\bar N_0 - \bar N_1)| is imported from [17] without stating the additional assumptions that make the bias–variance–noise decomposition valid for the demographic-parity discrimination measure under zero-one loss. Since the proof of Theorem 1 treats the three components independently, the paper should clarify which assumptions from [17] are being carried over and whether they are preserved under the FS re-sampling operation.
minor comments (5)
  1. [Abstract and Section 1] There are typos in the abstract and introduction: "an enormous of data" should be "an enormous amount of data," and "Y et" in the abstract should be "Yet."
  2. [Algorithm 1] The input line reads "Labled dataset" — should be "Labeled dataset." Also, Figure 1's caption writes "how to train the model" but the labels on the diagram say "phase" in one place and "phrase" in another; please standardize.
  3. [Section 4, final paragraph] The condition at the end of the proof is written as |\bar V_a(\hat Y )_{ssl} - \Delta \bar V_a(\hat Y )_{sl}| - \bar N_{a,p} >= 0, but the term \Delta is not defined and the notation differs from the earlier |\bar V_a(f)_sl| - |\bar V_a(f)_ssl|. Please make the notation consistent and define \Delta.
  4. [Section 5.1.2] The text says "LR has an accuracy of around 0.728 when \rho = 0.1 with the Adult dataset," but Figure 2's x-axis starts at \rho = 0.25. Please correct the value or the figure range.
  5. [Tables 1 and 2] Tables 1 and 2 report single accuracy and discrimination values without standard deviations, despite the experimental setup stating that results are averaged over 50 runs. Reporting variability would help assess whether the differences between ORI and FS are significant.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the framework's theoretical claim is an explicit conditional derived from an external bias-variance-noise decomposition, and the empirical claims are benchmarked against external datasets.

full rationale

The derivation chain is not circular. The central theoretical statement (Theorem 1) is explicitly conditional: it concludes that unlabeled data reduce discrimination only if the reduction in variance discrimination outweighs the noise discrimination introduced by pseudo-labeling. The proof obtains this from the bias-variance-noise decomposition of Chen et al. and Domingos, which are external sources, and the antecedent of the theorem is not disguised as a derived prediction. The empirical evaluation compares the proposed FS framework against ORI, uniform sampling, and preferential sampling on real and synthetic datasets, so the reported accuracy/discrimination trade-off is externally benchmarked rather than enforced by construction. The proof's assertion that discrimination in bias is unchanged (|B_sl|-|B_ssl|=0) and the claim that variance reduction dominates are substantive assumptions and are legitimate rigor/correctness concerns, but they are not circular: the paper does not define bias in terms of the theorem's conclusion, nor does it fit a parameter and then relabel the fit as a prediction. The only self-citation is to related work by Noroozi et al. (including one co-author), and it is not load-bearing for the paper's method or proof. No step in the manuscript reduces, by definition or by self-citation, to its own inputs.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

No new entities are introduced. The theoretical contribution is a conditional decomposition; the main burden is on transfer of known results and an unverified inequality.

free parameters (4)
  • sample ratio ρ = values 0.1 to 1.0 explored; main comparison uses ρ=1
    Controls how much unlabeled data is pseudo-labeled and added to the training set; the accuracy-discrimination trade-off depends on this choice (Section 5.1.2).
  • ensemble size K = 200 (main), explored to 500
    Number of fair datasets/base models in bagging; the paper reports performance stabilizes around K=200 (Section 5.1.3).
  • sample size ns = minimum group size in each dataset
    Number of samples drawn per protected-group/label cell; determines how much data is retained and affects accuracy and discrimination (Section 5.1.4).
  • split rate s = 0.8
    Fraction of labeled data used for training versus test; affects pseudo-labeler quality and all reported numbers (Section 5.1.1.2).
assumptions (5)
  • standard math Bias-variance-noise decomposition of expected loss (Def 3) applies to demographic-parity discrimination via misclassification-rate decomposition.
    Imported from Domingos (2000) and Chen et al. (2018); used without proof in Section 4.
  • domain assumption Lemma 2 (discrimination decreases as training set size n grows) holds for pseudo-labeled SSL training sets.
    Invoked in Theorem 1's proof to argue variance discrimination drops with augmented data; no evidence given that the learning-curve result transfers to pseudo-labeled data (Section 4).
  • domain assumption Labeled and unlabeled data follow the same distribution.
    Explicitly stated as an assumption in Section 7; required for pseudo-labels to be trustworthy.
  • ad hoc to paper Discrimination in bias is unchanged when the training set is augmented with pseudo-labeled data (|B_sl|-|B_ssl|=0).
    Asserted in the proof of Theorem 1 without justification; changing the training distribution generally changes the main prediction and hence bias.
  • domain assumption Protected attribute is binary and demographic parity is the chosen fairness criterion.
    Section 2.2; limits scope but is acknowledged.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fairness in Semi-supervised Learning: Unlabeled Data Help to Reduce Discrimination." pith.science (2026). https://pith.science/paper/HWWPMDN5

@misc{pith2026200912040,
  author       = {Pith},
  title        = {Pith review of: Fairness in Semi-supervised Learning: Unlabeled Data Help to Reduce Discrimination},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HWWPMDN5}},
  note         = {Machine review of arXiv:2009.12040}
}
read the original abstract

A growing specter in the rise of machine learning is whether the decisions made by machine learning models are fair. While research is already underway to formalize a machine-learning concept of fairness and to design frameworks for building fair models with sacrifice in accuracy, most are geared toward either supervised or unsupervised learning. Yet two observations inspired us to wonder whether semi-supervised learning might be useful to solve discrimination problems. First, previous study showed that increasing the size of the training set may lead to a better trade-off between fairness and accuracy. Second, the most powerful models today require an enormous of data to train which, in practical terms, is likely possible from a combination of labeled and unlabeled data. Hence, in this paper, we present a framework of fair semi-supervised learning in the pre-processing phase, including pseudo labeling to predict labels for unlabeled data, a re-sampling method to obtain multiple fair datasets and lastly, ensemble learning to improve accuracy and decrease discrimination. A theoretical decomposition analysis of bias, variance and noise highlights the different sources of discrimination and the impact they have on fairness in semi-supervised learning. A set of experiments on real-world and synthetic datasets show that our method is able to use unlabeled data to achieve a better trade-off between accuracy and discrimination.

Figures

Figures reproduced from arXiv: 2009.12040 by the authors.

Figure 1
Figure 1. The three phases of the fairness-enhanced sampling framework: 1) where to sample, 2) how to sample and 3) how to train [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. The trade-off between accuracy (Red) and discrimination [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 4
Figure 4. The impact of sample size on accuracy (Red) and dis [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Comparison with original scheme (ORI), uniform sam [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 34 canonical work pages

  1. [37]

    Lever- aging labeled and unlabeled data for consistent fair binary classi- fication,

    E. Chzhen, C. Denis, M. Hebiri, L. Oneto, and M. Pontil, “Lever- aging labeled and unlabeled data for consistent fair binary classi- fication,” in Advances in Neural Information Processing Systems 32 , 2019, pp. 12 739–12 750

    work page 2019

  2. [38]

    Leveraging Semi-Supervised Learning for Fairness using Neural Networks

    V . Noroozi, S. Bahaadini, S. Sheikhi, N. Mojab, and P . S. Yu, “Leveraging semi-supervised learning for fairness using neural networks,” arXiv preprint arXiv:1912.13230, 2019. Tao Zhang works towards his Ph.D degree with the school of Computer Science in the University of Technology Sydney, Australia. His research interests include privacy preserving, al...

    work page Pith review arXiv 1912

  3. [17]

    Why is my classifier dis- criminatory?

    I. Chen, F. D. Johansson, and D. Sontag, “Why is my classifier dis- criminatory?” in Advances in Neural Information Processing Systems 31, 2018, pp. 3539–3550

    work page 2018

  4. [1]

    Fair prediction with disparate impact: A study of bias in recidivism prediction instruments,

    A. Chouldechova, “Fair prediction with disparate impact: A study of bias in recidivism prediction instruments,”Big data, vol. 5, no. 2, pp. 153–163, 2017

    2017

  5. [2]

    Dissecting racial bias in an algorithm used to manage the health of populations,

    Z. Obermeyer, B. Powers, C. Vogeli, and S. Mullainathan, “Dissecting racial bias in an algorithm used to manage the health of populations,” Science, vol. 366, no. 6464, pp. 447– 453, 2019. [Online]. Available: https://science.sciencemag.org/ content/366/6464/447

    work page 2019

  6. [3]

    Building classifiers with independency constraints,

    T. Calders, F. Kamiran, and M. Pechenizkiy, “Building classifiers with independency constraints,” in 2009 IEEE International Confer- ence on Data Mining Workshops. IEEE, 2009, pp. 13–18

    work page 2009

  7. [4]

    Fairness Constraints: Mechanisms for Fair Classification,

    M. B. Zafar, I. Valera, M. G. Rodriguez, and K. P . Gummadi, “Fairness Constraints: Mechanisms for Fair Classification,” in Pro- ceedings of the 20th International Conference on Artificial Intelligence and Statistics, vol. 54, 20–22 Apr 2017, pp. 962–970

    work page 2017

  8. [5]

    Equality of opportunity in supervised learning,

    M. Hardt, E. Price, N. Srebro et al. , “Equality of opportunity in supervised learning,” in Advances in neural information processing systems, 2016, pp. 3315–3323

    work page 2016

Show all 38 references
  1. [6]

    Fairness through awareness,

    C. Dwork, M. Hardt, T. Pitassi, O. Reingold, and R. Zemel, “Fairness through awareness,” in Proceedings of the 3rd innovations in theoretical computer science conference. ACM, 2012, pp. 214–226

    2012

  2. [7]

    The variational fair autoencoder,

    C. Louizos, K. Swersky, Y. Li, M. Welling, and R. Zemel, “The variational fair autoencoder,”arXiv preprint arXiv:1511.00830, 2015

    2015 arXiv

  3. [8]

    Eliciting and enforcing subjective individual fairness,

    C. Jung, M. Kearns, S. Neel, A. Roth, L. Stapleton, and Z. S. Wu, “Eliciting and enforcing subjective individual fairness,” arXiv preprint arXiv:1905.10660, 2019

    1905 arXiv

  4. [9]

    Counterfactual fairness,

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

    2017

  5. [10]

    Avoiding discrimination through causal reasoning,

    N. Kilbertus, M. R. Carulla, G. Parascandolo, M. Hardt, D. Janz- ing, and B. Sch ¨olkopf, “Avoiding discrimination through causal reasoning,” in Advances in Neural Information Processing Systems , 2017, pp. 656–666. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 11

    2017

  6. [11]

    Data preprocessing techniques for classification without discrimination,

    F. Kamiran and T. Calders, “Data preprocessing techniques for classification without discrimination,” Knowledge and Information Systems, vol. 33, no. 1, pp. 1–33, 2012

    2012

  7. [12]

    Learning fair representations,

    R. Zemel, Y. Wu, K. Swersky, T. Pitassi, and C. Dwork, “Learning fair representations,” in International Conference on Machine Learn- ing, 2013, pp. 325–333

    2013

  8. [13]

    Learning ad- versarially fair and transferable representations,

    D. Madras, E. Creager, T. Pitassi, and R. Zemel, “Learning ad- versarially fair and transferable representations,” arXiv preprint arXiv:1802.06309, 2018

    2018 arXiv

  9. [14]

    Learning controllable fair representations,

    J. Song, P . Kalluri, A. Grover, S. Zhao, and S. Ermon, “Learning controllable fair representations,” in Proceedings of the 22nd Inter- national Conference on Artificial Intelligence and Statistics (AISTATS) 2019,, vol. 89, 16–18 Apr 2019, pp. 2164–2173

    2019

  10. [15]

    Fairness- aware classifier with prejudice remover regularizer,

    T. Kamishima, S. Akaho, H. Asoh, and J. Sakuma, “Fairness- aware classifier with prejudice remover regularizer,” in Joint Eu- ropean Conference on Machine Learning and Knowledge Discovery in Databases. Springer, 2012, pp. 35–50

    2012

  11. [16]

    Guarantees for spectral clustering with fairness constraints,

    M. Kleindessner, S. Samadi, P . Awasthi, and J. Morgenstern, “Guarantees for spectral clustering with fairness constraints,” in Proceedings of the 36th International Conference on Machine Learning , vol. 97, Long Beach, California, USA, 09–15 Jun 2019, pp. 3458– 3467

    2019

  12. [18]

    A framework for understanding unintended consequences of machine learning,

    H. Suresh and J. V . Guttag, “A framework for understanding unintended consequences of machine learning,” arXiv preprint arXiv:1901.10002, 2019

    1901 arXiv

  13. [19]

    A unified bias-variance decomposition,

    P . Domingos, “A unified bias-variance decomposition,” in Proceed- ings of 17th International Conference on Machine Learning , 2000, pp. 231–238

    2000

  14. [20]

    Pseudo-label: The simple and efficient semi- supervised learning method for deep neural networks,

    D.-H. Lee, “Pseudo-label: The simple and efficient semi- supervised learning method for deep neural networks,” in Work- shop on Challenges in Representation Learning, ICML , vol. 3, 2013, p. 2

    2013

  15. [21]

    Bagging predictors,

    L. Breiman, “Bagging predictors,” Machine learning, vol. 24, no. 2, pp. 123–140, 1996

    1996

  16. [22]

    Semi-supervised learning literature survey,

    X. J. Zhu, “Semi-supervised learning literature survey,” University of Wisconsin-Madison Department of Computer Sciences, Tech. Rep., 2005

    2005

  17. [23]

    Optimized pre-processing for discrimination preven- tion,

    F. Calmon, D. Wei, B. Vinzamuri, K. N. Ramamurthy, and K. R. Varshney, “Optimized pre-processing for discrimination preven- tion,” in Advances in Neural Information Processing Systems , 2017, pp. 3992–4001

    2017

  18. [24]

    Learning fair representations via an adversarial framework,

    R. Feng, Y. Yang, Y. Lyu, C. Tan, Y. Sun, and C. Wang, “Learning fair representations via an adversarial framework,” arXiv preprint arXiv:1904.13341, 2019

    1904 arXiv

  19. [25]

    Mitigating un- wanted biases with adversarial learning,

    B. H. Zhang, B. Lemoine, and M. Mitchell, “Mitigating un- wanted biases with adversarial learning,” in Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society , 2018, pp. 335–340

    2018

  20. [26]

    Costs and benefits of fair representation learning,

    D. McNamara, C. S. Ong, and R. C. Williamson, “Costs and benefits of fair representation learning,” in Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society , 2019, pp. 263–270

    2019

  21. [27]

    Fairness beyond disparate treatment & disparate impact: Learn- ing classification without disparate mistreatment,

    M. B. Zafar, I. Valera, M. Gomez Rodriguez, and K. P . Gummadi, “Fairness beyond disparate treatment & disparate impact: Learn- ing classification without disparate mistreatment,” in Proceedings of the 26th International Conference on World Wide Web , 2017, pp. 1171–1180

    2017

  22. [28]

    Empirical risk minimization under fairness constraints,

    M. Donini, L. Oneto, S. Ben-David, J. S. Shawe-Taylor, and M. Pon- til, “Empirical risk minimization under fairness constraints,” in Advances in Neural Information Processing Systems , 2018, pp. 2791– 2801

    2018

  23. [29]

    A reductions approach to fair classification,

    A. Agarwal, A. Beygelzimer, M. Dud ´ık, J. Langford, and H. Wal- lach, “A reductions approach to fair classification,” arXiv preprint arXiv:1803.02453, 2018

    2018 arXiv

  24. [30]

    Optimization with non-differentiable constraints with applications to fairness, recall, churn, and other goals,

    A. Cotter, H. Jiang, M. Gupta, S. Wang, T. Narayan, S. You, and K. Sridharan, “Optimization with non-differentiable constraints with applications to fairness, recall, churn, and other goals,” Journal of Machine Learning Research, vol. 20, no. 172, pp. 1–59, 2019

    2019

  25. [31]

    Two-player games for efficient non-convex constrained optimization,

    A. Cotter, H. Jiang, and K. Sridharan, “Two-player games for efficient non-convex constrained optimization,” CoRR, 2018

    2018

  26. [32]

    Nonconvex optimization for regression with fairness constraints,

    J. Komiyama, A. Takeda, J. Honda, and H. Shimao, “Nonconvex optimization for regression with fairness constraints,” in Interna- tional conference on machine learning, 2018, pp. 2737–2746

    2018

  27. [33]

    Learning optimal and fair decision trees for non-discriminative decision-making,

    S. Aghaei, M. J. Azizi, and P . Vayanos, “Learning optimal and fair decision trees for non-discriminative decision-making,” arXiv preprint arXiv:1903.10598, 2019

    1903 arXiv

  28. [34]

    Multiaccuracy: Black-box post-processing for fairness in classification,

    M. P . Kim, A. Ghorbani, and J. Zou, “Multiaccuracy: Black-box post-processing for fairness in classification,” in Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society, 2019, pp. 247– 254

    2019

  29. [35]

    Bias mitigation post-processing for individual and group fairness,

    P . K. Lohia, K. N. Ramamurthy, M. Bhide, D. Saha, K. R. Varshney, and R. Puri, “Bias mitigation post-processing for individual and group fairness,” in Icassp 2019-2019 ieee international conference on acoustics, speech and signal processing (icassp). IEEE, 2019, pp. 2847– 2851

    2019

  30. [36]

    Learning non-discriminatory predictors,

    B. Woodworth, S. Gunasekar, M. I. Ohannessian, and N. Srebro, “Learning non-discriminatory predictors,” in Conference on Learn- ing Theory, 2017, pp. 1920–1953

    2017

Pith tools

Reviewed August 27, 2026 · model on record in the stance chip above.