Pith. sign in

REVIEW 3 major objections 6 minor 46 references

Functional Bilevel Optimization for Predictive Fairness

T0 review · 3 major / 6 minor · reviewed 2026-07-11 · grok-4.5

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

Pith's one-line read Mean demographic parity for continuous sensitive attributes is a functional bilevel problem that can be optimized with an exact, Hessian-free hypergradient.

desk verdict Clean functional-bilevel take on mean demographic parity for continuous A, with a checkable adjoint and solid methods work; the multi-dataset win claim is real but conditioned on a post-filtered 27/60 subset. read the letter →

arxiv 2607.05098 v1 pith:5NWH3GUI submitted 2026-07-06 cs.LG stat.ML

Ieva Petrulionyte , Julien Mairal , Michael Arbel This is my paper
classification cs.LGstat.ML
keywords demographicparityDPVarfunctionalbileveloptimizationhypergradientcontinuoussensitiveattributesfairregressionadversarialdebiasingHSIC
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

When the sensitive attribute is continuous and high-dimensional, full independence between predictions and that attribute is often too strong. This paper instead targets mean demographic parity: the average prediction should not swing systematically with the sensitive attribute. That idea is captured by DPVar, the variance of the conditional-mean prediction given the sensitive attribute. Estimating that conditional mean makes the accuracy–fairness objective a functional bilevel problem—the inner level fits a function, not a finite parameter vector. For squared loss the authors derive a closed-form adjoint that yields an exact hypergradient without Hessian-vector products (FBO); a second algorithm (ITD) unrolls inner steps and works beyond squared loss. On synthetic data and a new semi-synthetic suite built from 60 real tabular regression datasets, both methods achieve the lowest or near-lowest accuracy regret under DPVar budgets, matching or beating HSIC, adversarial, linear-dependence, and generalized-DP baselines.

What carries the argument

DPVar(ω)=Var_A(E[f_ω(X)|A]) together with the closed-form adjoint a*_ω(a)=−α(h*_ω(a)−μ_ω) for the squared-loss inner problem; this identity turns the outer hypergradient into a single expectation involving only the centered residual and ∂ωf_ω, eliminating Hessian-vector products.

What would settle it

On the same 27 non-degenerate tabular tasks, recompute true (or high-fidelity cross-fit) DPVar after training: if FBO/ITD no longer dominate under tight DPVar budgets once the inner estimator is replaced by a stronger or oracle conditional-mean fit, the claimed advantage collapses.

Watch

Extended reading notes

Core claim

Optimizing DPVar—the variance of the conditional-mean prediction given a continuous sensitive attribute—yields a functional bilevel program whose squared-loss case admits a closed-form adjoint and therefore an exact, Hessian-free hypergradient; the resulting FBO algorithm, together with the more general unrolled ITD method, produces the best or near-best fairness–accuracy trade-offs among strong dependence-penalty and adversarial baselines on synthetic and 60-dataset tabular benchmarks.

Load-bearing premise

A parametric neural net fit on a finite data split must faithfully estimate the true conditional mean of the prediction given a continuous high-dimensional sensitive attribute; if that estimate is wrong, the outer update steers an artifact rather than true mean parity.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes DPVar = Var_A(E[f_ω(X)|A]) as a mean demographic-parity criterion for continuous, high-dimensional sensitive attributes, formulates its optimization as a functional bilevel problem (inner conditional-mean regression, outer accuracy–fairness trade-off), and gives two algorithms: FBO, which uses a closed-form adjoint for squared-loss inner objectives to obtain a Hessian-free hypergradient (Prop. 3.1, App. B), and ITD, which differentiates through unrolled inner steps. Empirically, the methods are evaluated on a controlled synthetic interaction stress test and a semi-synthetic suite built from 60 tabular regression datasets with automatically constructed continuous A, against HSIC, adversarial debiasing, linear-dependence (R²), and GDP/GDP* baselines, with a unified cross-fitted DPVar diagnostic. The abstract and §4 claim that FBO and ITD achieve the lowest or near-lowest aggregate fairness–accuracy regret and match or outperform those baselines.

Significance. If the claims hold, the work is a clear methodological contribution: it cleanly connects a first-moment DP relaxation to functional bilevel optimization, supplies an explicit adjoint (App. B) that avoids Hessian-vector products in the squared-loss case, and ships a reusable continuous-A tabular protocol with a common diagnostic. The synthetic design (β_int knob) is a useful stress test of when mean-targeting should beat full-dependence penalties. The closed-form hypergradient and the dual FBO/ITD design are concrete strengths that other continuous-A fairness work can build on. The main significance risk is empirical scope: the headline multi-dataset superiority is currently summarized on a post-filtered subset rather than the full 60-dataset pool the abstract advertises.

major comments (3)
  1. Abstract and §4.2 / Fig. 5: the central multi-dataset claim (“lowest or near-lowest aggregate fairness-accuracy regret” on a benchmark “built from 60” datasets) is supported by win rates computed only on the 27 datasets with non-degenerate VAL ranges (≥0.1 on both MSE and DPVar). The other 33 are discarded after inspecting validation fronts because methods collapse. That is a post-hoc selection on the quantity being ranked. Please report the same constrained win-rate (or regret) analysis on the full 60-dataset pool, and/or restate the abstract claim as conditional on non-degenerate trade-offs, with a clear analysis of the 33 degenerate regimes (e.g., whether baselines are competitive there). Without this, the aggregate superiority statement is not fully supported.
  2. Prop. 3.1 vs Algorithms 1–2: Prop. 3.1 and App. B derive an exact functional hypergradient under h∈L²(P_A) with Hessian 2I. In practice FBO plugs a finite-capacity MLP h_φ into that formula and still describes the hypergradient as “exact.” When A is continuous/high-dimensional, h_φ is only an approximate conditional mean (flagged in Scope and limitations), so the implemented FBO step is an approximate functional hypergradient, not the exact parametric bilevel gradient. Please state this gap explicitly in §3 (when the functional formula is exact vs approximate), and ideally add a small ablation on inner fit quality (M, width of h_φ) vs achieved test DPVar so readers can see that the outer update is not steering an estimation artifact.
  3. §4 evaluation protocol: all methods are ranked solely by the DPVar diagnostic that FBO/ITD optimize, while HSIC, adversarial, R², and GDP optimize different dependence notions. Criterion-aligned comparison is legitimate for a DPVar paper, but the abstract’s “match or outperform … baselines” reading is stronger than “best under the DPVar metric.” Please either (i) also report at least one non-DPVar dependence measure (e.g., empirical HSIC or adversary accuracy) on the same Pareto points, or (ii) narrow the claim language to superiority under the DPVar diagnostic. This is especially important for GDP*, which is trained after a train-only PCA projection of A when dim(A)>2.
minor comments (6)
  1. Abstract / §4: “fairness-accuracy regret” is not defined; the reported summary is constrained win rate under Tight/Mid/Loose DPVar budgets. Align terminology with the actual metric.
  2. Fig. 4 caption is incomplete/garbled (“Each panel the final test Pareto front…”); fix wording and ensure axis labels are readable in the preprint.
  3. Notation: population weight α, training penalty γ, and selection weight ζ appear late (§4); introduce the distinction once in §3 when F(ω) is defined.
  4. Algorithm 2 FBO fairness gradient uses an IN-sum involving f_ω(x_i) rather than an explicit ∂_ωf factor written as in Prop. 3.1; a one-line note that autodiff implements the same expectation would help.
  5. Related work (Fig. 2 / App. A): OT fair-regression citations are well placed; a brief sentence on why post-processing Wasserstein methods are not run as baselines (discrete groups / post-hoc nature) would reduce reviewer questions.
  6. Appendix D: the 25% predictive-pool rule and corr thresholds are free design choices; a short sensitivity check (e.g., 15% vs 25%) in the appendix would strengthen the benchmark claim.

Circularity Check

1 steps flagged · score 1.0 of 10

No derivation circularity: closed-form hypergradient is derived from the bilevel objective; only minor non-load-bearing self-citation of the authors' functional-bilevel framework.

  1. self citation load bearing [§3 Prop. 3.1; App. B; citation [20]]
    "The proof, given in Appendix B, applies the functional implicit-differentiation and adjoint-sensitivity framework of Petrulionyte et al. [20]. The novelty here is the consequence: the squared-loss structure of the inner problem reduces the adjoint to a centered residual, and the hypergradient becomes a single expectation involving h⋆_ω − μ_ω and ∂ωfω(X)."

    The general functional adjoint identity is imported from the same authors' prior NeurIPS paper rather than re-proved from scratch. This is minor and not load-bearing for circularity: App. B still carries out the full DPVar-specific calculation (Hessian = 2I, adjoint = −α(h⋆−μ), total gradient), so the closed form is not assumed equal to the target by definition and does not force the empirical claims.

full rationale

The paper defines DPVar as Var_A(E[f_ω(X)|A]), casts its optimization as a functional bilevel problem (inner conditional-mean regression, outer accuracy–DPVar trade-off), and derives a closed-form adjoint for squared loss (Prop. 3.1, App. B). That derivation is algebraic: the inner Hessian collapses to 2I on L2, so the adjoint is the centered residual −α(h⋆_ω−μ_ω) and the hypergradient is an explicit expectation—not a quantity fitted to the target and renamed as a prediction. FBO/ITD then implement that hypergradient; empirical claims are comparisons against HSIC, adversarial, R², and GDP baselines under a shared DPVar diagnostic. Using the same diagnostic for all methods is criterion-aligned evaluation, not mathematical circularity. The only self-citation of note is Petrulionyte et al. [20] for the general functional adjoint-sensitivity identity; Appendix B re-derives the DPVar-specific gradient under stated assumptions (PA,in=PA,out, squared inner loss), so the central result is not forced by an unverified self-citation chain. Semi-synthetic A construction and the 27/60 non-degenerate filter are experimental-design choices, not definitional reductions of a claimed first-principles result. Score 1 for minor self-citation only.

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

The central claim rests on standard L2/bilevel calculus plus domain choices that define the fairness problem (DPVar as the target, parametric inner MLP, IN/OUT split, semi-synthetic A). Free parameters are the usual ML knobs (fairness weights, steps, architectures) selected on VAL; they do not invent new physical constants. Invented entities are methodological (DPVar scalar, FBO/ITD procedures, the benchmark construction), not new physical objects. Independent evidence for the methods is empirical comparison, not external formal verification.

free parameters (5)
  • fairness training weight γ (and population α / selection ζ)
    Scalar trade-off between MSE and DPVar; swept and selected on VAL to build Pareto fronts. Central comparative claims depend on these grids.
  • inner steps M, unrolling length K, stepsizes η_in / η_out
    Optimization hyperparameters for FBO/ITD; method-specific grids affect whether the hypergradient is accurate enough to realize the claimed trade-offs.
  • MLP architecture (2 hidden layers, width 64, tanh) for f_ω and h_φ
    Capacity of predictor and conditional-mean estimator is fixed by hand; high-dim A estimation quality depends on this choice.
  • sensitive-attribute construction fraction (25% of predictive pool) and correlation thresholds
    Appendix D procedure (top-200, |corr|≥0.02, proxy score s_j) defines A on every tabular dataset; results are conditional on this hand-designed construction.
  • synthetic unfairness knobs δ, β_int and nuisance dimension of A
    Control first-moment vs interaction-driven unfairness in §4.1; used to illustrate separation of methods, not fitted to real data.
assumptions (5)
  • domain assumption PA,in = PA,out so that the adjoint and hypergradient identities hold across IN/OUT splits
    Stated in §3; finite-sample splits only approximate this. Prop. 3.1 depends on it.
  • standard math Regularity conditions ensuring differentiability of ω ↦ h*_ω in the functional implicit-function setting of Petrulionyte et al. [20]
    Invoked for Prop. 3.1; not re-proved in full, imported from the functional bilevel framework.
  • standard math Inner squared-loss Hessian w.r.t. h is 2I on L2(PA,in), yielding closed-form adjoint a*_ω = −α(h*_ω − μ_ω)
    Core algebraic step in Appendix B; standard for L2 projection / squared regression.
  • domain assumption Mean demographic parity (flat conditional mean) is the appropriate fairness target for continuous high-dimensional A, rather than full independence
    Motivates DPVar throughout §1 and Scope; applications may require stronger distributional criteria.
  • ad hoc to paper Semi-synthetic A built from target-correlated features is a valid methodological proxy for continuous sensitive attributes
    Appendix D construction; paper disclaims legal/social sensitivity of selected coordinates but still uses them for the main benchmark claim.
invented entities (4)
  • DPVar = Var_A(E[f_ω(X)|A]) as operational continuous-A fairness criterion
    purpose: Scalar mean-parity violation that recovers weak DP for discrete A and is estimable for continuous/high-dim A
    Related to GDP/CoD but positioned as the training objective for neural predictors; independent evidence is conceptual/empirical, not a new physical quantity.
  • FBO algorithm (closed-form adjoint hypergradient for squared-loss DPVar)
    purpose: Exact Hessian-free outer updates for the functional bilevel DPVar problem
    New algorithmic object derived in Prop. 3.1 / Alg. 2; validated only within this paper’s experiments.
  • ITD algorithm for DPVar (unrolled inner differentiation)
    purpose: Approximate hypergradients that extend beyond squared-loss inner objectives
    Standard unrolling idea applied to this fairness bilevel problem; complementary to FBO.
  • 60-dataset semi-synthetic continuous-A fairness benchmark with unified DPVar diagnostic
    purpose: Standardized large-scale evaluation suite for continuous-attribute fairness methods
    New evaluation resource constructed in §4.2 / App. D; not an external public benchmark yet.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Functional Bilevel Optimization for Predictive Fairness." pith.science (2026). https://pith.science/paper/5NWH3GUI

@misc{pith2026260705098,
  author       = {Pith},
  title        = {Pith review of: Functional Bilevel Optimization for Predictive Fairness},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5NWH3GUI}},
  note         = {Machine review of arXiv:2607.05098}
}
abstract

When sensitive attributes are continuous and high-dimensional $-$ demographic score vectors, posteriors over attributes, age or income profiles $-$ enforcing full statistical independence is often too restrictive, and existing relaxations rely on indirect dependence penalties or adversarial schemes that do not directly target the fairness-accuracy trade-off. We instead consider mean demographic parity through DPVar, the variance of the conditional-mean prediction given the sensitive attribute, and show that optimizing it yields a functional bilevel problem. We propose two algorithms for this problem: FBO, which uses a closed-form adjoint we derive for the squared-loss case to obtain an exact hypergradient, and ITD, which differentiates through unrolled inner steps and extends beyond squared loss. On synthetic data and a new semi-synthetic benchmark built from 60 tabular regression datasets, both methods achieve the lowest or near-lowest aggregate fairness-accuracy regret, and consistently match or outperform strong HSIC, adversarial, linear-dependence, and generalized-DP baselines.

Figures

Figures reproduced from arXiv: 2607.05098 by the authors.

Figure 1
Figure 1. Conceptual illustration of DPVar. A predictor with higher DPVar has a conditional-mean [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Positioning of DPVar relative to demographic parity (DP) and existing fairness methods. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Test MSE-DPVar trade-off on synthetic data. Each point corresponds to a different DPVar penalty weight (chosen on VAL). Lower is better on both axes. Differences across methods become more visible as βint grows. 4.2 Semisynthetic benchmark on tabular datasets To complement the synthetic study with a large-scale evaluation on real data, we build a semi￾synthetic tabular benchmark from a curated pool of 60 regression … view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Seed-wise uncertainty of the fairness–accuracy frontier on six datasets. Each panel the final test Pareto front (seed 0) for each method per dataset (first 9 alphabetically). Lower DPVar means more fairness, lower MSE means more accuracy. Results. We use a win rate to …
Figure 5
Figure 5. Figure 5: How often each method is the dataset-level winner under constrained fairness budgets (Tight, Mid, Loose), across 27 tabular datasets. Each bar shows the proportion of datasets for which a method was that dataset-level winner at the corresponding budget. Error bars show…
Figure 6
Figure 6. Figure 6: Additional Pareto fronts on tabular regression datasets with semisynthetic continuous sensitive attributes. Each panel shows the test trade-off between prediction error (MSE; lower is better) and demographic-parity variance (DPVar; lower is better) for seed 0. Each cur…
Figure 7
Figure 7. Figure 7: Seed-wise uncertainty of the fairness–accuracy frontier on five datasets. Each panel corresponds to one dataset and summarizes 10 runs (seeds 0–9) of the final test Pareto front for each method. Lower DPVar budget means more fairness, lower MSE means more accuracy. For…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 1 canonical work pages

  1. [1]

    Solon Barocas and Andrew D. Selbst. Big data’s disparate impact.California Law Review, 104 (3):671–732, 2016

    2016

  2. [2]

    A survey on bias and fairness in machine learning.ACM Computing Surveys, 54:1–35, 2021

    Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, and Aram Galstyan. A survey on bias and fairness in machine learning.ACM Computing Surveys, 54:1–35, 2021

    2021

  3. [3]

    Generalized demographic parity for group fairness.International Conference on Learning Representations (ICLR), 2022

    Zhimeng Jiang, Xiaotian Han, Chao Fan, Fan Yang, Ali Mostafavi, and Xia Hu. Generalized demographic parity for group fairness.International Conference on Learning Representations (ICLR), 2022

    2022

  4. [4]

    Equality of opportunity in supervised learning

    Moritz Hardt, Eric Price, and Nati Srebro. Equality of opportunity in supervised learning. Advances in Neural Information Processing Systems (NIPS), 2016

    2016

  5. [5]

    Kusner, Joshua Loftus, Chris Russell, and Ricardo Silva

    Matt J. Kusner, Joshua Loftus, Chris Russell, and Ricardo Silva. Counterfactual fairness. In Advances in Neural Information Processing Systems (NIPS), 2017

    2017

  6. [6]

    MIT Press, 2023

    Solon Barocas, Moritz Hardt, and Arvind Narayanan.Fairness and Machine Learning: Limita- tions and Opportunities. MIT Press, 2023

    2023

  7. [7]

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

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

    2017

  8. [8]

    Inherent trade-offs in the fair determination of risk scores

    Jon Kleinberg, Sendhil Mullainathan, and Manish Raghavan. Inherent trade-offs in the fair determination of risk scores. InInnovations in Theoretical Computer Science (ITCS 2017), volume 67 ofLeibniz International Proceedings in Informatics (LIPIcs), pages 43:1–43:23, 2017

    2017

Show all 46 references
  1. [9]

    Fairness-aware classifier with prejudice remover regularizer.Machine Learning and Knowledge Discovery in Databases, pages 35–50, 2012

    Toshihiro Kamishima, Shotaro Akaho, Hideki Asoh, and Jun Sakuma. Fairness-aware classifier with prejudice remover regularizer.Machine Learning and Knowledge Discovery in Databases, pages 35–50, 2012

    2012

  2. [10]

    Teo, Le Song, Bernhard Schölkopf, and Alex J

    Arthur Gretton, Kenji Fukumizu, Choon H. Teo, Le Song, Bernhard Schölkopf, and Alex J. Smola. A kernel statistical test of independence.Advances in Neural Information Processing Systems (NIPS), 2008

    2008

  3. [11]

    Mitigating unwanted biases with adversarial learning.Proceedings of the Conference on AI, Ethics, and Society (AIES), pages 335–340, 2018

    Brian Hu Zhang, Blake Lemoine, and Margaret Mitchell. Mitigating unwanted biases with adversarial learning.Proceedings of the Conference on AI, Ethics, and Society (AIES), pages 335–340, 2018

    2018

  4. [12]

    Censoring representations with an adversary.International Conference on Learning Representations (ICLR), 2016

    Harrison Edwards and Amos Storkey. Censoring representations with an adversary.International Conference on Learning Representations (ICLR), 2016

    2016

  5. [13]

    Fairness-aware learning for continuous attributes and treatments.Proceedings of Machine Learning Research (PMLR), 97: 4382–4391, 2019

    Jeremie Mary, Clément Calauzènes, and Noureddine El Karoui. Fairness-aware learning for continuous attributes and treatments.Proceedings of Machine Learning Research (PMLR), 97: 4382–4391, 2019

    2019

  6. [14]

    Insung Kong, Kunwoong Kim, and Yongdai Kim. Fair representation learning for continuous sensitive attributes using expectation of integral probability metrics.IEEE Transactions on Pattern Analysis & Machine Intelligence, 47:3784–3795, 2025

    2025

  7. [15]

    Fair Bilevel Neural Network (FairBiNN): On Balancing Fairness and Accuracy via Stackelberg Equilibrium.Advances in Neural Information Processing Systems (NeurIPS), 2024

    Mehdi Yazdani-Jahromi, Ali Khodabandeh Yalabadi, AmirArsalan Rajabi, Aida Tayebi, Ivan Garibay, and Ozlem Garibay. Fair Bilevel Neural Network (FairBiNN): On Balancing Fairness and Accuracy via Stackelberg Equilibrium.Advances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  8. [16]

    Nonconvex optimization for regression with fairness constraints

    Junpei Komiyama, Akiko Takeda, Junya Honda, and Hajime Shimao. Nonconvex optimization for regression with fairness constraints. InProc. International Conference on Machine Learning (ICML), 2018

    2018

  9. [17]

    Mitigating discrimination in insurance with wasserstein barycenters.PKDD/ECML Workshops, 2023

    Arthur Charpentier, François Hu, and Philipp Ratz. Mitigating discrimination in insurance with wasserstein barycenters.PKDD/ECML Workshops, 2023

    2023

  10. [18]

    Should bank stress tests be fair?Management Science, 71(1): 262–278, 2024

    Paul Glasserman and Mike Li. Should bank stress tests be fair?Management Science, 71(1): 262–278, 2024. doi: 10.1287/mnsc.2022.02060

    2024 doi

  11. [19]

    FairJob: A Real-World Dataset for Fairness in Online Systems.Advances in Neural Information Processing Systems (NeurIPS), 2024

    Mariia Vladimirova, Eustache Diemert, and Federico Pavone. FairJob: A Real-World Dataset for Fairness in Online Systems.Advances in Neural Information Processing Systems (NeurIPS), 2024. 10

    2024

  12. [20]

    Functional bilevel optimization for machine learning.Advances in Neural Information Processing Systems (NeurIPS), 2024

    Ieva Petrulionyte, Julien Mairal, and Michael Arbel. Functional bilevel optimization for machine learning.Advances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  13. [21]

    Muhammad Bilal Zafar, Isabel Valera, Manuel Gomez Rodriguez, and Krishna P. Gummadi. Fairness constraints: Mechanisms for fair classification.Proceedings of Machine Learning Research (PMLR), 54:962–970, 2017

    2017

  14. [22]

    A reductions approach to fair classification.Proceedings of Machine Learning Research (PMLR), 80:60–69, 2018

    Alekh Agarwal, Alina Beygelzimer, Miroslav Dudik, John Langford, and Hanna Wallach. A reductions approach to fair classification.Proceedings of Machine Learning Research (PMLR), 80:60–69, 2018

    2018

  15. [23]

    Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Reingold, and Richard S. Zemel. Fairness through awareness.Proceedings of the Innovations in Theoretical Computer Science Conference (ITCS), 2012

    2012

  16. [24]

    Fair kernel learning

    Adrián Pérez-Suay, Valero Laparra, Gonzalo Mateo-García, Jordi Muñoz-Marí, Luis Gómez- Chova, and Gustau Camps-Valls. Fair kernel learning. InMachine Learning and Knowledge Discovery in Databases - European Conference, ECML PKDD 2017, volume 10534 ofLecture Notes in Computer S...

    2017

  17. [25]

    Kernel dependence reg- ularizers and Gaussian processes with applications to algorithmic fairness.Pattern Recognition, 132:108922, 2022

    Zhu Li, Adrián Pérez-Suay, Gustau Camps-Valls, and Dino Sejdinovic. Kernel dependence reg- ularizers and Gaussian processes with applications to algorithmic fairness.Pattern Recognition, 132:108922, 2022

    2022

  18. [26]

    Fair regression with Wasserstein barycenters

    Evgenii Chzhen, Christophe Denis, Mohamed Hebiri, Luca Oneto, and Massimiliano Pontil. Fair regression with Wasserstein barycenters. InAdvances in Neural Information Processing Systems (NIPS), 2020

    2020

  19. [27]

    Projection to fairness in statistical learning.arXiv preprint 2005.11720, 2020

    Thibaut Le Gouic, Jean-Michel Loubes, and Philippe Rigollet. Projection to fairness in statistical learning.arXiv preprint 2005.11720, 2020

    2005 arXiv

  20. [28]

    A minimax framework for quantifying risk-fairness trade-off in regression.The Annals of Statistics, 50(4):2416–2442, 2022

    Evgenii Chzhen and Nicolas Schreuder. A minimax framework for quantifying risk-fairness trade-off in regression.The Annals of Statistics, 50(4):2416–2442, 2022

    2022

  21. [29]

    Fair learning with Wasserstein barycenters for non-decomposable performance measures.International Conference on Artifi- cial Intelligence and Statistics (AISTATS), 2023

    Solenne Gaucher, Nicolas Schreuder, and Evgenii Chzhen. Fair learning with Wasserstein barycenters for non-decomposable performance measures.International Conference on Artifi- cial Intelligence and Statistics (AISTATS), 2023

    2023

  22. [30]

    Fairness-aware neural Rényi min- imization for continuous features.International Joint Conference on Artificial Intelligence (IJCAI), 2020

    Vincent Grari, Sylvain Lamprier, and Marcin Detyniecki. Fairness-aware neural Rényi min- imization for continuous features.International Joint Conference on Artificial Intelligence (IJCAI), 2020

    2020

  23. [31]

    Veronika Shilova, Emmanuel Malherbe, Giovanni Palma, Laurent Risser, and Jean-Michel Loubes. Fairness-aware grouping for continuous sensitive variables: Application for debiasing face analysis with respect to skin tone.Proceedings of the 28th European Conference on Artificial ...

    2025

  24. [32]

    Fairbatch: Batch selection for model fairness.International Conference on Learning Representations (ICLR), 2021

    Yuji Roh, Kangwook Lee, Steven Euijong Whang, and Changho Suh. Fairbatch: Batch selection for model fairness.International Conference on Learning Representations (ICLR), 2021

    2021

  25. [33]

    Fair machine learning under limited demographically labeled data.Workshop on Socially Responsible Machine Learning (SRML), 2022

    Mustafa Safa Ozdayi, Murat Kantarcioglu, and Rishabh Iyer. Fair machine learning under limited demographically labeled data.Workshop on Socially Responsible Machine Learning (SRML), 2022

    2022

  26. [34]

    Fairness-informed pareto optimization : An efficient bilevel framework.arXiv preprint arXiv:2601.13448, 2026

    Sofiane Tanji, Samuel Vaiter, and Yassine Laguel. Fairness-informed pareto optimization : An efficient bilevel framework.arXiv preprint arXiv:2601.13448, 2026

    2026

  27. [35]

    Hyperparameter optimization with approximate gradient.Proceedings of Machine Learning Research (PMLR), 48:737–746, 2016

    Fabian Pedregosa. Hyperparameter optimization with approximate gradient.Proceedings of Machine Learning Research (PMLR), 48:737–746, 2016

    2016

  28. [36]

    UCI machine learning repository, 2019

    Dheeru Dua and Casey Graff. UCI machine learning repository, 2019

    2019

  29. [37]

    van Rijn, Bernd Bischl, and Luis Torgo

    Joaquin Vanschoren, Jan N. van Rijn, Bernd Bischl, and Luis Torgo. Openml: Networked science in machine learning.ACM SIGKDD Explorations Newsletter, 15(2):49–60, 2014

    2014

  30. [38]

    Mutual information neural estimation.Proceedings of Machine Learning Research (PMLR), 80:531–540, 2018

    Mohamed Ishmael Belghazi, Aristide Baratin, Sai Rajeshwar, Sherjil Ozair, Yoshua Bengio, Aaron Courville, and R Devon Hjelm. Mutual information neural estimation.Proceedings of Machine Learning Research (PMLR), 80:531–540, 2018

    2018

  31. [39]

    Domain-adversarial training of neural networks.Journal of Machine Learning Research (JMLR), 17(59):1–35, 2016

    Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks.Journal of Machine Learning Research (JMLR), 17(59):1–35, 2016. 11 A Additional Related Wo...

    2016

  32. [40]

    We work in L2(PAout) the space of square integrable functions wrt

    Functional gradient in h.We note h(A), φ(A), and EA∼PAout as h, φ, and E for readability. We work in L2(PAout) the space of square integrable functions wrt. the measure PAout equipped with the inner product ⟨u, v⟩L2(PAout ) :=E A∼PAout [u(A)v(A)]. For any direction φ∈L 2(PAout...

  33. [41]

    The adjointa ⋆ ω ∈L 2(PAin )solves ∂2 hhLin(ω, h⋆ ω)a⋆ ω =−∂ hLout(ω, h⋆ ω), hence 2a⋆ ω(a) =−2α(h ⋆ ω(a)−µ ω) a⋆ ω(a) =−α(h ⋆ ω(a)−µ ω), µ ω :=E A∼PAout [h⋆ ω(A)]

    Adjoint.The inner objective is Lin(ω, h) =E (X,A)∼Pin (fω(X)−h(A)) 2 , whose Hessian w.r.t.his∂ 2 hhLin = 2IonL 2(PAin ). The adjointa ⋆ ω ∈L 2(PAin )solves ∂2 hhLin(ω, h⋆ ω)a⋆ ω =−∂ hLout(ω, h⋆ ω), hence 2a⋆ ω(a) =−2α(h ⋆ ω(a)−µ ω) a⋆ ω(a) =−α(h ⋆ ω(a)−µ ω), µ ω :=E A∼PAout [h⋆ ω(A)]

  34. [42]

    Compute each term •Direct outer term: ∂ωLout(ω, h⋆ ω) = 2E out [(fω(X)−Y)∂ ωfω(X)]

    Total gradientBy the functional adjoint identity, ∇ωF(ω) =∂ ωLout(ω, h⋆ ω) +∂ 2 ωhLin(ω, h⋆ ω)[a⋆ ω]. Compute each term •Direct outer term: ∂ωLout(ω, h⋆ ω) = 2E out [(fω(X)−Y)∂ ωfω(X)]. •Implicit (adjoint) term: ∂hLin(ω, h)[φ] =−2E in [(fω(X)−h(A))φ(A)], differentiating w.r.t....

  35. [43]

    We compute cj =|corr(X j, y)| for every coordinate and rank coordinates in decreasing order ofc j

    Rank coordinates by target informativeness.Let Xj denote the jth standardized feature coordinate and letybe the standardized training target. We compute cj =|corr(X j, y)| for every coordinate and rank coordinates in decreasing order ofc j

  36. [44]

    If this threshold leaves the pool empty, we fall back to the top-ranked coordinates without thresholding

    Form a predictive candidate pool.We retain the top 200 ranked coordinates, after discard- ing coordinates with cj <0.02 . If this threshold leaves the pool empty, we fall back to the top-ranked coordinates without thresholding

  37. [45]

    Concretely, we sample up to 128 other coordinates and define pj = max c̸=j |corr(Xc, Xj)|2, where the maximum is taken over the sampled coordinates

    Prefer coordinates that are also proxy-like.For each candidate coordinate Xj, we compute a simple proxy score that measures how predictable it is from the remaining features. Concretely, we sample up to 128 other coordinates and define pj = max c̸=j |corr(Xc, Xj)|2, where the ...

  38. [46]

    These selected coordinates are removed fromXbefore training the predictor

    Define the sensitive attribute and remove it from the predictor input.The sensitive attributeAis formed by taking the top 25% of this candidate pool according tos j, clipped so that at least one predictive coordinate remains in the predictor input. These selected coordinates a...

Pith tools

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