Pith. sign in

REVIEW 3 major objections 6 minor 68 references

A generalized exponentiated gradient method cuts multi-class classification bias by up to 92% while keeping accuracy loss within 14%.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 17:42 UTC pith:HQACRTBE

load-bearing objection Real extension of EG to multi-class, but the fairness claims outrun the training objective. the 3 major comments →

arxiv 2603.21393 v2 pith:HQACRTBE submitted 2026-03-22 cs.LG stat.ML

A Generalised Exponentiated Gradient Approach to Enhance Fairness in Binary and Multi-class Classification Tasks

Maryam Boubekraoui , Giordano d'Aloisio , Antinisca Di Marco This is my paper
classification cs.LG stat.ML
keywords fairnessmulti-class classificationexponentiated gradientdemographic parityequalized oddsin-processing bias mitigationconstrained optimization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper extends the exponentiated gradient (EG) algorithm, originally designed for binary classification with a single fairness constraint, to multi-class classification with multiple simultaneous fairness constraints. It reformulates fair learning as a constrained optimization problem over randomized classifiers, using positive-label demographic parity and equalized odds as linear moment constraints. The resulting Generalized Exponentiated Gradient (GEG) algorithm alternates between a cost-sensitive learner and a dual auditor that reweights violated constraints. Across seven multi-class and three binary datasets, GEG reports fairness improvements up to 92% with accuracy decreases up to 14%, and outperforms six baselines in most fairness-effectiveness Pareto combinations.

Core claim

The central claim is that multi-class fair classification can be reduced to a saddle-point problem with linear fairness constraints defined on a single positive label, and that the exponentiated gradient scheme—learner and auditor playing a zero-sum game—solves this problem with an O(1/√T) convergence rate. The authors show that positive-label Demographic Parity and Equalized Odds can be written as linear moment inequalities, and that a Combined Parity constraint enforcing both at once remains linear. Empirically, they claim GEG reliably reduces positive-label and general multi-class fairness metrics, improves Pareto-optimal trade-offs compared to pre- and post-processing baselines, and work

What carries the argument

Positive-label moment constraints: instead of enforcing full multi-class Demographic Parity or Equalized Odds, GEG constrains P(h(X)=y_p | A=a) and P(h(X)=y_p | A=a,Y=y) for a chosen favorable class y_p. These are linear in the randomized classifier Q, making the Lagrangian convex-concave. GEG iterates between (1) a cost-sensitive learner that assigns each sample a signed weight combining misclassification cost and fairness violations, and (2) an exponentiated-gradient dual update over the simplex of constraints, with a duality-gap stopping criterion. Combined Parity stacks the SP and EO constraint matrices into one system.

Load-bearing premise

The paper assumes that enforcing fairness for a single chosen positive label is enough to reduce the general multi-class fairness metrics (SPD, EOD, AOD) on which it evaluates; this proxy does not always hold, as the paper's own tables show.

What would settle it

Compute full multi-class SPD and EOD on the Drug, Law, and Park datasets after applying GEG-SP and GEG-EO. If the full metrics are systematically worse than the positive-label metrics, the proxy assumption fails. A stronger falsifier: rerun GEG with constraints on all class labels instead of only y_p and compare full SPD/EOD to the reported results.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Practitioners can use GEG as an in-processing bias mitigation method for multi-class tasks without post-hoc reweighting or prediction changes.
  • GEG-CP can simultaneously improve demographic parity and equalized odds in binary classification, where the original EG typically handles one at a time.
  • When the base classifier is logistic regression and the number of classes is ≤4, GEG may increase accuracy while reducing bias, not just trade off.
  • GEG is compatible with stronger base classifiers such as random forests and gradient boosting, making it usable in high-performance pipelines.
  • The method adds no inference overhead, only training-time cost, so deployed systems keep their runtime latency.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The positive-label proxy is not a guaranteed surrogate for the full multi-class fairness metrics: the paper's own tables show GEG-SP worsening EOD on Drug and Law, and GEG-EO worsening EOD on Park, so claims about general DP/EO fairness are only as strong as the proxy holds.
  • A natural testable extension is to run GEG with constraints on every class label instead of a single y_p; if observed proxy failures disappear, the assumption of a single positive class is the limiting factor.
  • The convergence bound is standard no-regret theory; whether real-world gains persist with nonlinear or non-convex base models is an empirical question that the paper starts to answer with RF and GB results.
  • Since the auditor updates only the most violated constraints, adding more constraints (Combined Parity) may slow convergence and explain the lower effectiveness GEG-CP shows on some datasets.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The manuscript proposes GEG, an in-processing fairness method that extends the Exponentiated Gradient (EG) reduction of Agarwal et al. from binary to multi-class classification and from a single fairness constraint to multiple simultaneous constraints. The method optimizes positive-label Demographic Parity and Equalized Odds (Definitions 3–4) via linear moment constraints, solves the resulting saddle-point problem with multiplicative weights and a cost-sensitive learner, and includes a convergence theorem (Theorem 1). The authors evaluate three variants (SP, EO, CP) on seven multi-class and three binary datasets against several baselines, reporting Pareto-optimality counts and detailed fairness/effectiveness tables (Appendix C). The central claim is that GEG 'enhances fairness' under multiple fairness definitions with improvements up to 92%.

Significance. If correct, the contribution would be practically valuable: it would supply a principled in-processing algorithm for multi-class fairness with multiple constraints, backed by a no-regret analysis, and would provide a unified treatment of binary and multi-class tasks. The paper also includes an open-source replication package. However, the significance is substantially tempered by two gaps: (i) the training constraints target only the positive label, while evaluation uses stricter multi-class metrics and the paper's own tables show systematic worsenings on EOD; and (ii) the link between the Lagrangian and the implemented cost-sensitive learner is not derived, so the convergence guarantee does not demonstrably apply to the algorithm. The contribution is therefore best viewed as a heuristic extension with promising but overstated empirical support.

major comments (3)
  1. [Section 3.1 / Section 6.3 / Table C.2] The training constraints (Definitions 3–4) enforce parity only for the positive label y_p. The evaluation metrics (Section 4.3.1, Eqs. 25–27) are the stricter multi-class SPD/EOD/AOD. The paper asserts in Section 6.3 that predictions are 'also fair' with respect to these general definitions, but this is not implied and is contradicted by the paper's own results. In Table C.2, GEG-SP worsens EOD on Drug (0.011→0.095) and Law (0.037→0.187); GEG-EO worsens EOD on Park (0.163→0.549). In Table C.3, GEG-CP worsens EOD on Adult (0.098→0.109). Thus the abstract and Section 5.1 claim that 'all versions of GEG significantly enhance fairness across all datasets and fairness definitions' is unsupported. Improving positive-label parity can coexist with, or even cause, greater unfairness on other labels.
  2. [Section 3.3 / Algorithm 1 / Theorem 1] The cost-sensitive learner is not derived from the Lagrangian in Eq. (15). The per-sample weight w_j = γ_error_j + Σ_i λ_i γ_fair_i,j is used with an adjusted label, but γ_fair_i,j is equated to γ_i(h(x_j)), which is ill-defined because γ_i is a global moment, not a per-example quantity. The convergence theorem assumes a τ-approximate best-response oracle for L(h,λ), but no argument shows that the implemented cost-sensitive objective corresponds to minimizing L(h,λ) over H. Without this link, the theorem does not apply to the algorithm actually implemented. This is a load-bearing gap.
  3. [Section 4.3.1 / Eqs. (25)–(27)] The reported fairness metrics are not the multi-class definitions stated in Section 3.1. Eq. (26) defines EOD as max over y of P(h(X)=y|Y=y,A=1) − P(h(X)=y|Y=y,A=0), which only compares TPR per true label, not the full equalized-odds condition over all predicted labels (Definition 2). Eq. (27) defines AOD with FPR and TPR averages but also only over y. Thus the evaluation metrics are themselves restricted, positive-label-like variants; the paper should state this explicitly and should not describe them as 'implementing' Definitions 1–2. This also affects the statistical comparison and the Pareto-optimality counts derived from these metrics.
minor comments (6)
  1. [Section 3.1, Eqs. (4)–(5)] Use consistent formatting for y_p; the text currently shows 'y p' in several places.
  2. [Section 4.3.1, Eq. (27)] The phrase 'It from -1 to +1' should read 'It ranges from -1 to +1'.
  3. [Table C.2 / C.3 captions] The captions refer to grey shading, but the shading is not visible in the text version; please ensure the published version renders the shading or replace it with symbols.
  4. [Algorithm 1, line 13] The variables bγ_i(h_t) are computed but not defined in the algorithm; they should refer to the empirical constraint functions of Section 3.2.
  5. [Section 4.3.2] The description of the Wilcoxon test says the null hypothesis is 'the median between two dependent samples is different', but then states H0 as 'objective not improved'. Align the wording.
  6. [References] Reference [28] (Fairlearn documentation) is a web resource; provide proper access details.

Circularity Check

1 steps flagged

Minor circular overlap: positive-label fairness metrics mirror the optimized constraints; the central EG-reduction derivation remains independent.

specific steps
  1. self definitional [Section 3.1 (Defs. 3–4) and Section 4.3.1 (SPD-P/EOD-P/AOD-P, Eqs. 28–30)]
    "In the remainder of this section, we focus on Definitions 3 and 4 as training constraints ... SPD-P = P(h(X)=y_p|A=0)−P(h(X)=y_p|A=1)"

    The positive-label evaluation metrics are the same statistical quantities that GEG optimizes: Definition 3 requires P(h(X)=y_p|A=a)=P(h(X)=y_p) for all a, which for binary A is equivalent to SPD-P=0; Definition 4 (Eq. 5) is equivalently EOD-P/AOD-P=0. The Lagrangian (Eq. 15) and the dual updates in Algorithm 1 drive exactly these moments toward their tolerances. Thus reporting fairness improvements on SPD-P/EOD-P/AOD-P is a sanity check that the optimizer satisfied its own objective, not an independent prediction. The independent evidence is the general multi-class SPD/EOD/AOD metrics (Defs. 1–2), and the paper's own Tables C.2/C.3 show those can worsen (e.g., GEG-SP increases EOD on Drug 0.011→0.095; GEG-EO increases EOD on Park 0.163→0.549), so the blanket claim of enhancing fairness und

full rationale

The core methodological derivation is not circular: GEG is a faithful extension of Agarwal et al.'s external reduction-based EG framework, and Theorem 1 is a standard no-regret/convex-concave saddle-point argument that does not assume the conclusions it proves. The self-citations to [24], which share authors with this paper, are used mainly for the positive-label convention, dataset positive labels, and as a comparison baseline (DEMV); they are not invoked as the proof of GEG's convergence or as a uniqueness theorem. The only partial overlap is that the positive-label fairness metrics used in evaluation (SPD-P, EOD-P, AOD-P) are exactly the constraints optimized during training. Since evaluation is on held-out test folds, this overlap is not a strict identity, but it does mean improvement on those metrics is largely a check on the optimizer rather than an independent discovery. The paper does report the more general multi-class SPD/EOD/AOD metrics, which are not the training objective; however, the claim in Section 6.3 that positive-label fairness implies fairness under the general definitions is unsupported and contradicted by the paper's own tables. That is a validity concern, not a circularity of the derivation chain. Overall, the central algorithm and its convergence analysis remain self-contained, so the circularity score is low.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

GEG rests on the reduction machinery of Agarwal et al. (randomized classifiers, Lagrangian duality, exponentiated-gradient dual updates) and on practical assumptions: a meaningful positive class, a binary sensitive attribute with known privileged/unprivileged groups, and a cost-sensitive learner that approximates the assumed oracle. The 'Combined Parity' constraint is a new constraint combination but not a new entity. No invented physical or conceptual objects are introduced.

free parameters (4)
  • learning rate η = 1e-5
    Hand-set, copied from the original EG paper (Agarwal et al.). Controls dual update speed and thus convergence of the auditor.
  • tolerance δ = 0.05
    Hand-set, copied from EG. Sets the dual budget B=1/δ and the constraint slack used in the objective.
  • positive class y_p per dataset = varies by dataset
    Chosen from prior work (d'Aloisio et al., Chen et al.). The entire fairness constraint system and the reported positive-label metrics depend on this selection.
  • adaptive tolerance levels bε_i = unspecified
    Section 3.3 states that adaptive tolerance levels are allowed, but the algorithm does not specify how they are set; the M matrix uses ε=0. This is an under-specified modeling choice.
axioms (4)
  • standard math Sion's minimax theorem applies to the empirical Lagrangian over compact convex sets.
    Invoked in Section 3.4 to assert existence of a saddle point for the convex-concave Lagrangian.
  • standard math Exponentiated gradient over the simplex has no-regret guarantees for bounded linear losses.
    Used in Theorem 1 and Appendix B to derive the O(B sqrt(log(n+1)/T)) duality-gap bound.
  • domain assumption A single favorable class y_p exists and is meaningful for each multi-class problem.
    Required by Definitions 3-4 and Section 6.3. If no class is a natural positive outcome, the constraints and the evaluation metrics lose their interpretation.
  • domain assumption The cost-sensitive learner is a τ-approximate best response for the weighted 0-1 loss.
    Theorem 1 depends on this oracle; however, the implementation uses scikit-learn logistic regression optimizing a surrogate loss with sample weights, not 0-1 cost-sensitive loss. The approximation quality τ is unverified.

pith-pipeline@v1.3.0-alltime-deepseek · 36256 in / 19115 out tokens · 183754 ms · 2026-08-02T17:42:27.668652+00:00 · methodology

0 comments
read the original abstract

The widespread use of AI and ML models in sensitive areas raises significant concerns about fairness. While the research community has introduced various methods for bias mitigation in binary classification tasks, the issue remains under-explored in multi-class classification settings. To address this limitation, in this paper, we first formulate the problem of fair learning in multi-class classification as a multi-objective problem between effectiveness (i.e., prediction correctness) and multiple linear fairness constraints. Next, we propose a Generalised Exponentiated Gradient (GEG) algorithm to solve this task. GEG is an in-processing algorithm that enhances fairness in binary and multi-class classification settings under multiple fairness definitions. We conduct an extensive empirical evaluation of GEG against six baselines across seven multi-class and three binary datasets, using four widely adopted effectiveness metrics and three fairness definitions.

Figures

Figures reproduced from arXiv: 2603.21393 by Antinisca Di Marco, Giordano d'Aloisio, Maryam Boubekraoui.

Figure 1
Figure 1. Figure 1: Experimental process [24], we apply the pre-processing approach only on the training set. Similarly, the post-processing Blackbox approach is applied on the prediction from the ML model [50]. After training and testing the approaches, we report the mean and standard deviation of the metrics obtained. In the following, we describe the datasets employed (Section 4.1), the bench￾marks used in the evaluation (… view at source ↗
Figure 2
Figure 2. Figure 2: RQ1: Pareto optimality of baseline and GEG variations considering each pair of effectiveness and fairness metrics. Interestingly, improved fairness does not always reduce effectiveness. Signifi￾cant reductions in effectiveness scores are noted only for the Crime and Obesity datasets, likely due to their higher number of classes (6 and 5, respectively), which complicates the prediction task for the model. T… view at source ↗
Figure 3
Figure 3. Figure 3: RQ2: Pareto optimality of base LR, base EG approaches, and GEG-CP considering each pair of effectiveness and fairness metrics for binary classification. achieve a consistent optimal trade-off between fairness and effectiveness under all definitions considered. Notably, solutions by GEG-CP achieve the highest Pareto optimality under AOD POS, EOD, and AOD, across all effectiveness metrics. By looking at the … view at source ↗
Figure 4
Figure 4. Figure 4: RQ3: Pareto optimality of DEMV, Blackbox and GEG variations considering each pair of effectiveness and fairness metrics for multi-class classification. of GEG and the baseline DEMV and Blackbox approaches. From the plot, we observe how, while DEMV achieves better Pareto optimal solutions under the SPD POS definition, all GEG variations provide a consistent number of Pareto optimal solutions under all combi… view at source ↗
Figure 5
Figure 5. Figure 5: RQ4: Pareto optimality of RF and GEG variations considering each pair of effec￾tiveness and fairness metrics. Answer to RQ3: GEG overcomes Blackbox in terms of Pareto optimality under all effectiveness and fairness metrics combinations, while it overcomes DEMV in 18 out of 24 (%75) combinations. 5.4. RQ4: Different Base Classifiers [PITH_FULL_IMAGE:figures/full_fig_p026_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: RQ4: Pareto optimality of GB and GEG variations considering each pair of effec￾tiveness and fairness metrics. Answer to RQ4: GEG is effective in bias mitigation even when more complex base classifiers are employed, especially when the bias of the base classifier is relatively high. 6. Discussion In the following, we discuss the theoretical and practical implications derived from our study. 6.1. Theoretical… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

68 extracted references · 5 linked inside Pith

  1. [1]

    Abdi, H. (2010). Holm’s sequential bonferroni procedure.Encyclopedia of research design, 1(8):1–8

    2010

  2. [2]

    Agarwal, A., Beygelzimer, A., Dudik, M., Langford, J., and Wallach, H. (2018). A reductions approach to fair classification. In Dy, J. and Krause, A., editors,Proceedings of the 35th International Conference on Machine Learning, volume 80 ofProceedings of Machine Learning Research, pages 60–

    2018

  3. [3]

    Angwin, J., Larson, J., Mattu, S., and Kirchner, L. (2016). Machine bias. ProPublica, May, 23(2016):139–159

    2016

  4. [4]

    and Briand, L

    Arcuri, A. and Briand, L. (2011). A practical guide for using statistical tests to assess randomized algorithms in software engineering. InProceedings of the 33rd International Conference on Software Engineering, ICSE ’11, pages 1–10, New York, NY, USA. Association for Computing Machinery

    2011

  5. [5]

    A., Christopher, C

    Austin, K. A., Christopher, C. M., and Dickerson, D. (2016). Will i pass the bar exam: Predicting student success using lsat scores and law school performance.HofstrA l. rev., 45:753

    2016

  6. [6]

    and Ng, Y.-K

    Baskota, A. and Ng, Y.-K. (2018). A graduate school recommendation sys- tem using the multi-class support vector machine and knn approaches. In 2018 IEEE International Conference on Information Reuse and Integration (IRI), pages 277–284. IEEE

    2018

  7. [7]

    Bellamy, R. K. E., Dey, K., Hind, M., Hoffman, S. C., Houde, S., Kannan, K., Lohia, P., Martino, J., Mehta, S., Mojsilovi´ c, A., Nagar, S., Ramamurthy, K. N., Richards, J., Saha, D., Sattigeri, P., Singh, M., Varshney, K. R., and Zhang, Y. (2019). AI Fairness 360: An extensible toolkit for detecting and mitigating algorithmic bias.IBM Journal of Research...

    2019

  8. [8]

    Beutel, A., Chen, J., Doshi, T., Qian, H., Woodruff, A., Luu, C., Kreit- mann, P., Bischof, J., and Chi, E. H. (2019). Putting fairness principles into practice: Challenges, metrics, and improvements. InProceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society, pages 453–459

    2019

  9. [9]

    N., Benbelghit, A., and Belhaouari, S

    Bouanane, K., Medakene, A. N., Benbelghit, A., and Belhaouari, S. B. (2024). Faircolor: An efficient algorithm for the balanced and fair reviewer assignment problem.Information Processing & Management, 61(6):103865

    2024

  10. [10]

    and Gey, F

    Buckland, M. and Gey, F. (1994). The relationship between recall and precision.Journal of the American society for information science, 45(1):12–

    1994

  11. [11]

    and Zhu, F

    Cai, J. and Zhu, F. (2026). Learning fair representations without label- ing sensitive attribute via dynamic environment partitioning and invariant learning.Information Processing & Management, 63(2, Part B):104469

    2026

  12. [12]

    Canali, S., Schiaffonati, V., and Aliverti, A. (2022). Challenges and recom- mendations for wearable devices in digital health: Data quality, interoperabil- ity, health equity, fairness.PLOS Digital Health, 1(10):e0000104. Publisher: Public Library of Science

    2022

  13. [13]

    and Haas, C

    Caton, S. and Haas, C. (2023). Fairness in Machine Learning: A Survey. ACM Computing Surveys. Just Accepted

    2023

  14. [14]

    M., Hort, M., Sarro, F., and Harman, M

    Chen, Z., Zhang, J. M., Hort, M., Sarro, F., and Harman, M. (2022a). Fairness Testing: A Comprehensive Survey and Analysis of Trends. arXiv:2207.10223 [cs]

    Pith/arXiv arXiv

  15. [15]

    M., Sarro, F., and Harman, M

    Chen, Z., Zhang, J. M., Sarro, F., and Harman, M. (2022b). Maat: a novel ensemble approach to addressing fairness and performance bugs for machine learning software. InProceedings of the 30th ACM joint european software engineering conference and symposium on the foundations of software engi- neering, pages 1122–1134

  16. [16]

    M., Sarro, F., and Harman, M

    Chen, Z., Zhang, J. M., Sarro, F., and Harman, M. (2024). Fairness Im- provement with Multiple Protected Attributes: How Far Are We? Conference Name: 2024 IEEE/ACM 46th International Conference on Software Engi- neering (ICSE) Meeting Name: 2024 IEEE/ACM 46th International Con- ference on Software Engineering (ICSE) Place: Lisbon, Portugal Publisher: IEEE...

    2024

  17. [17]

    Chouldechova, A. (2017). Fair prediction with disparate impact: A study of bias in recidivism prediction instruments.Big data, 5(2):153–163

    2017

  18. [18]

    Corbett-Davies, S., Pierson, E., Feller, A., Goel, S., and Huq, A. (2017). Algorithmic decision making and the cost of fairness. InProceedings of the 23rd acm sigkdd international conference on knowledge discovery and data mining, pages 797–806. 39

    2017

  19. [19]

    Publisher: Wiley Online Library

  20. [20]

    Cortez, P., Cerdeira, A., Almeida, F., Matos, T., and Reis, J. (2009). Modeling wine preferences by data mining from physicochemical properties. Decision support systems, 47(4):547–553

    2009

  21. [21]

    Denis, C., Elie, R., Hebiri, M., and Hu, F. (2024). Fairness guarantees in multi-class classification with demographic parity.Journal of Machine Learning Research, 25(130):1–46

    2024

  22. [22]

    Dutta, S., Wei, D., Yueksel, H., Chen, P.-Y., Liu, S., and Varshney, K. (2020). Is there a trade-off between fairness and accuracy? a perspective using mismatched hypothesis testing. InInternational conference on machine learning, pages 2803–2813. PMLR

    2020

  23. [23]

    Dwork, C., Hardt, M., Pitassi, T., Reingold, O., and Zemel, R. (2012). Fairness through awareness. InProceedings of the 3rd Innovations in The- oretical Computer Science Conference, ITCS ’12, pages 214–226, New York, NY, USA. Association for Computing Machinery

    2012

  24. [24]

    d’Aloisio, G., Di Sipio, C., Di Marco, A., and Di Ruscio, D. (2025a). How fair are we? from conceptualization to automated assessment of fairness def- initions.Software and Systems Modeling, pages 1–27

  25. [25]

    d’Aloisio, G., D’Angelo, A., Di Marco, A., and Stilo, G. (2023). Debiaser for Multiple Variables to enhance fairness in classification tasks.Information Processing & Management, 60(2):103226

    2023

  26. [26]

    D., Marco, A

    d’Aloisio, G., Sipio, C. D., Marco, A. D., and Ruscio, D. D. (2025b). To- wards early detection of algorithmic bias from dataset’s bias symptoms: An empirical study.Information and Software Technology, 188:107905

  27. [27]

    EU AI Act: first regulation on artificial intelligence |News|European Parliament

    European Union (2023). EU AI Act: first regulation on artificial intelligence |News|European Parliament

    2023

  28. [28]

    Fabris, A., Messina, S., Silvello, G., and Susto, G. A. (2022). Algorithmic fairness datasets: the story so far.Data Mining and Knowledge Discovery, 36(6):2074–2152

    2022

  29. [29]

    Fairlearn documentation

    Fairlearn (2022). Fairlearn documentation

    2022

  30. [30]

    A., Zhang, J., Hu, Z., Qi, Y., Nickleach, S., Socolinsky, D., Sengamedu, S., and Faloutsos, C

    Fang, X., Xu, W., Tan, F. A., Zhang, J., Hu, Z., Qi, Y., Nickleach, S., Socolinsky, D., Sengamedu, S., and Faloutsos, C. (2024). Large Language Models(LLMs) on Tabular Data: Prediction, Generation, and Understanding – A Survey. arXiv:2402.17944 [cs]

    Pith/arXiv arXiv 2024

  31. [31]

    K., Mirkes, E

    Fehrman, E., Muhammad, A. K., Mirkes, E. M., Egan, V., and Gorban, A. N. (2017). The Five Factor Model of Personality and Evaluation of Drug Consumption Risk. In Palumbo, F., Montanari, A., and Vichi, M., editors, Data Science, Studies in Classification, Data Analysis, and Knowledge Orga- nization, pages 231–242, Cham. Springer International Publishing. 40

    2017

  32. [32]

    and Schapire, R

    Freund, Y. and Schapire, R. E. (1997). A decision-theoretic generalization of on-line learning and an application to boosting

    1997

  33. [33]

    A., Scheidegger, C., Venkatasubramanian, S., Choudhary, S., Hamilton, E

    Friedler, S. A., Scheidegger, C., Venkatasubramanian, S., Choudhary, S., Hamilton, E. P., and Roth, D. (2019). A comparative study of fairness- enhancing interventions in machine learning. InProceedings of the Conference on Fairness, Accountability, and Transparency, pages 329–338. ACM

    2019

  34. [34]

    Grinsztajn, L., Oyallon, E., and Varoquaux, G. (2022). Why do tree-based models still outperform deep learning on typical tabular data? In Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., and Oh, A., editors, Advances in Neural Information Processing Systems, volume 35, pages 507–

    2022

  35. [35]

    T., and Yoo, S

    Harman, M., McMinn, P., De Souza, J. T., and Yoo, S. (2011). Search based software engineering: Techniques, taxonomy, tutorial.Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 7007 LNCS:1–59

    2011

  36. [36]

    Hardt, M., Price, E., and Srebro, N. (2016). Equality of opportunity in supervised learning.Advances in neural information processing systems, 29:3315–3323

    2016

  37. [37]

    M., Sarro, F., and Harman, M

    Hort, M., Zhang, J. M., Sarro, F., and Harman, M. (2021). Fairea: A model behaviour mutation approach to benchmarking bias mitigation methods. In Proceedings of the 29th ACM joint meeting on European software engineering conference and symposium on the foundations of software engineering, pages 994–1006

    2021

  38. [38]

    M., Harman, M., and Sarro, F

    Hort, M., Chen, Z., Zhang, J. M., Harman, M., and Sarro, F. (2024). Bias mitigation for machine learning classifiers: A comprehensive survey.ACM Journal on Responsible Computing, 1(2):1–52

    2024

  39. [39]

    Kamiran, F., Calders, T., and Pechenizkiy, M. (2010). Discrimination aware decision tree learning. In2010 IEEE International Conference on Data Min- ing, pages 869–874. IEEE

    2010

  40. [40]

    and Calders, T

    Kamiran, F. and Calders, T. (2012). Data preprocessing techniques for classification without discrimination.Knowledge and Information Systems, 33(1):1–33

    2012

  41. [41]

    Kleinberg, J., Mullainathan, S., and Raghavan, M. (2016). Inherent trade- offs in the fair determination of risk scores.arXiv preprint arXiv:1609.05807

    Pith/arXiv arXiv 2016

  42. [42]

    and Warmuth, M

    Kivinen, J. and Warmuth, M. K. (1997). Exponentiated gradient versus gradient descent for linear predictors.information and computation, 132(1):1– 63

    1997

  43. [43]

    Kozodoi, N., Jacob, J., and Lessmann, S. (2022). Fairness in credit scoring: Assessment, implementation and profit implications.European Journal of Operational Research, 297(3):1083–1094. Publisher: North-Holland

    2022

  44. [44]

    and others (1996)

    Kohavi, R. and others (1996). Scaling up the accuracy of naive-bayes clas- sifiers: A decision-tree hybrid. InKdd, volume 96, pages 202–207. 41

    1996

  45. [45]

    Ma, W., Chen, W., Lu, L., and Fan, X. (2024). Integrating learners’ knowl- edge background to improve course recommendation fairness: A multi-graph recommendation method based on contrastive learning.Information Process- ing & Management, 61(4):103750

    2024

  46. [46]

    Lim, T.-S., Loh, W.-Y., and Shih, Y.-S. (2000). A comparison of pre- diction accuracy, complexity, and training time of thirty-three old and new classification algorithms.Machine learning, 40(3):203–228

    2000

  47. [47]

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

    2021

  48. [48]

    Meenachi, L., Ramakrishnan, S., Sivaprakash, M., Thangaraj, C., and Sethupathy, S. (2022). Multi Class Ensemble Classification for Crop Rec- ommendation. In2022 International Conference on Inventive Computation Technologies (ICICT), pages 1319–1324. ISSN: 2767-7788

    2022

  49. [49]

    Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M., and Duchesnay, E. (2011). Scikit-learn: Machine learning in Python.Journal of Machine Learn- ing Research, 12:2825–2830

    2011

  50. [50]

    Palechor, F. M. and Manotas, A. d. l. H. (2019). Dataset for estimation of obesity levels based on eating habits and physical condition in individuals from Colombia, Peru and Mexico.Data in Brief, 25:104344

    2019

  51. [51]

    Quadros, A., Magalh˜ aes, S., Mol, D., Lima, J., Vieira, A., and Brand˜ ao, W. (2025). Multi-class bias mitigation methods for classification without discrimination.SN Computer Science, 6(8):1–17

    2025

  52. [52]

    and Lee, S

    Putzel, P. and Lee, S. (2022). Blackbox Post-Processing for Multiclass Fairness.arXiv:2201.04461 [cs]. arXiv: 2201.04461

    Pith/arXiv arXiv 2022

  53. [53]

    Ratanamahatana, C. A. and Gunopulos, D. (2002). Scaling up the naive bayesian classifier: Using decision trees for feature selection

    2002

  54. [54]

    Radovanovi´ c, S., Petrovi´ c, A., Delibaˇ si´ c, B., and Suknovi´ c, M. (2020). En- forcing fairness in logistic regression algorithm. In2020 International Con- ference on INnovations in Intelligent SysTems and Applications (INISTA), pages 1–7. IEEE

    2020

  55. [55]

    and Fitzpatrick-Lins, K

    Rosenfield, G. and Fitzpatrick-Lins, K. (1986). A coefficient of agreement as a measure of thematic classification accuracy.Photogrammetric Engineering and Remote Sensing, 52(2):223–227

    1986

  56. [56]

    and Baveja, A

    Redmond, M. and Baveja, A. (2002). A data-driven software tool for en- abling cooperative information sharing among police departments.European Journal of Operational Research, 141(3):660–678. 42

    2002

  57. [57]

    Sion, M. (1958). On general minimax theorems.Pacific Journal of Math- ematics, 8:171–176

    1958

  58. [58]

    Sarro, F., Petrozziello, A., and Harman, M. (2016). Multi-objective soft- ware effort estimation. InProceedings of the 38th International Conference on Software Engineering, ICSE ’16, pages 619–630, New York, NY, USA. Association for Computing Machinery

    2016

  59. [59]

    Suchithra, M. S. and Pai, M. L. (2018). Improving the Performance of Sig- moid Kernels in Multiclass SVM Using Optimization Techniques for Agricul- tural Fertilizer Recommendation System. In Zelinka, I., Senkerik, R., Panda, G., and Lekshmi Kanthan, P. S., editors,Soft Computing Systems, Commu- nications in Computer and Information Science, pages 857–868,...

    2018

  60. [60]

    P., Singla, A., Weller, A., and Zafar, M

    Speicher, T., Heidari, H., Grgic-Hlaca, N., Gummadi, K. P., Singla, A., Weller, A., and Zafar, M. B. (2018). A unified approach to quantifying al- gorithmic unfairness: Measuring individual &group unfairness via inequality indices. InProceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 2239–2248

    2018

  61. [61]

    Tsanas, A., Little, M., McSharry, P., and Ramig, L. (2009). Accurate telemonitoring of Parkinson’s disease progression by non-invasive speech tests. Nature Precedings, pages 1–1. Publisher: Nature Publishing Group

    2009

  62. [62]

    Tizpaz-Niari, S., Kumar, A., Tan, G., and Trivedi, A. (2022). Fairness-aware configuration of machine learning libraries.arXiv preprint arXiv:2202.06196

    Pith/arXiv arXiv 2022

  63. [63]

    Woolson, R. F. (2005). Wilcoxon signed-rank test.Encyclopedia of Bio- statistics, 8. Publisher: Wiley Online Library

    2005

  64. [64]

    Wang, Y., Zhang, R., Yang, Q., Zhou, Q., Zhang, S., Fan, Y., Huang, L., Li, K., and Zhou, F. (2024). Faircare: Adversarial training of a heteroge- neous graph neural network with attention mechanism to learn fair represen- tations of electronic health records.Information Processing & Management, 61(3):103682

    2024

  65. [65]

    Zhang, G., Cheng, D., Yuan, G., and Zhang, S. (2024). Learning fair representations via rebalancing graph structure.Information Processing & Management, 61(1):103570. 43

    2024

  66. [66]

    M., Ezz, M., and Almuayqil, S

    Yanes, N., Mostafa, A. M., Ezz, M., and Almuayqil, S. N. (2020). A machine learning-based recommender system for improving students learning experiences.IEEE Access, 8:201218–201235

    2020

  67. [68]

    P., and Zaiane, O

    Zhang, J., Cao, P., Gross, D. P., and Zaiane, O. R. (2013). On the applica- tion of multi-class classification in physical therapy recommendation.Health Information Science and Systems, 1(1):15. 44

    2013

  68. [520]

    Curran Associates, Inc