Pith. sign in

REVIEW 5 major objections 4 minor 2 cited by

Fairness-Aware Grouping for Continuous Sensitive Variables: Application for Debiasing Face Analysis with respect to Skin Tone

T0 review · 5 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Partitioning skin tone by measured discrimination, not default bins, exposes the groups a model treats most unequally.

desk verdict Good method, solid synthetic validation, but the real-data story sits on an unvalidated proprietary skin-tone estimator and an evaluation metric that is partly the objective itself. read the letter →

arxiv 2507.11247 v1 pith:FMTORJFU submitted 2025-07-15 cs.CV cs.LG

Veronika Shilova , Emmanuel Malherbe , Giovanni Palma , Laurent Risser , Jean-Michel Loubes This is my paper
classification cs.CVcs.LG
keywords fairness-awaregroupingcontinuoussensitiveattributesskintoneestimationCIELABcolorspaceindividualtypologyanglesubgroupfairnesspost-hocdebiasingoptimaltransport
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

Group fairness is usually measured by slicing a population into predefined demographic bins, but when the protected attribute is continuous—skin tone, age, income—those bins can hide the subgroups that actually suffer discrimination. This paper claims that the right groups are the ones that maximize the variance of the per-group deviation $\Phi(k)=P(Y=1|S_P=k)-P(Y=1)$, the difference between a group's positive-outcome rate and the population rate, over all connected partitions of the attribute space. It contributes two algorithms: FairGroups, an exhaustive search over grid segments accelerated by dynamic-programming precomputation, and a K-Means variant that provably produces connected segments when fairness is monotonic in the attribute. On face images, using a proprietary CIELAB skin-tone estimator, the data-driven partitions find discrimination that Fitzpatrick-style bins and the usual light/dark split miss, remain stable between CelebA and FFHQ for a fixed model, and improve debiasing: the reported dependence between predictions and skin tone drops from 0.126 to 0.039 while accuracy moves only from 0.793 to 0.781.

What carries the argument

The machinery is the pair $(P, \Phi)$: $P$ is a partition of the continuous sensitive space $L$ into connected, non-overlapping groups $P_k$, and $\Phi(k)=P(Y=1|S_P=k)-P(Y=1)$ assigns each group its deviation from the population positive rate. The objective is to choose $P$ maximizing $\mathrm{Var}(\Phi(S_P))$, the probability-weighted inter-group variance of these deviations; for $K=2$ this equals $\pi(1-\pi)\mathit{DI}^2$. To solve it, FairGroups builds a grid $\Lambda$ on $L$, uses dynamic programming to fill an upper-triangular matrix with the values $\Psi_{j_1,j_2}=P(Y=1|L\in[\lambda_{j_1},\lambda_{j_2}])-P(Y=1)$ for all interval ranges in $O(M^2)$, and then enumerates all boundary combinations to maximize the objective exactly up to grid resolution. The K-Means heuristic clusters the diagonal entries $\psi_j=P(Y=1|L\in[\lambda_{j-1},\lambda_j])-P(Y=1)$; Proposition 2 ensures the resulting clusters are connected segments precisely when $P(Y=1|L)$ is monotonic in $L$. This same $\Phi$ machinery can be evaluated on ground-truth $Y$ or model output $\hat{Y}$, which is what lets the paper separate dataset-level discrimination from model amplification.

What would settle it

Take a public dataset with spectrophotometer-measured skin reflectance for the same faces, recompute ITA and the FairGroups partitions, and compare the group boundaries and post-processing HGR values; if they shift materially, the real-data conclusions rest on the proprietary estimator rather than on the partition method.

Watch

Extended reading notes

Core claim

The paper's central claim is that the partition of a continuous sensitive attribute most relevant to fairness is the one whose groups differ most in discrimination, measured by $\Phi(k)=P(Y=1|S_P=k)-P(Y=1)$, where $S_P$ is the group-membership variable induced by the partition. Maximizing the population-weighted variance of $\Phi$ over partitions is equivalent, when $K=2$, to maximizing $\pi(1-\pi)\mathit{DI}^2$, so the criterion balances how different the two groups are with how evenly the population is split. FairGroups is an exact-on-grid algorithm: it discretizes the attribute space into $M$ intervals, precomputes the positive-outcome rate for every interval range in $O(M^2)$ via dynamic programming, and exhaustively searches segment or rectangle boundaries. A cheaper K-Means variant clusters the same per-interval deviations and, under monotonic fairness ($P(Y=1|L)$ monotonic in $L$), provably returns connected segments. Empirically, on CelebA and FFHQ with a proprietary CIELAB skin-tone estimator, FairGroups partitions find more spread-out $\Phi$ values with narrower confidence intervals than Fitzpatrick or fixed light/dark groups, reproduce across datasets for a fixed model, and, when used with optimal-transport post-processing, reduce the dependence between predictions and ITA from 0.126 to 0.039 at an accuracy cost of 0.793 to 0.781.

Load-bearing premise

The load-bearing premise is that the proprietary model used to extract skin tone yields unbiased CIELAB $L^*,a^*,b^*$ values for face images; the manuscript does not validate this estimator against physical measurements and states the values cannot be publicly shared, so every group boundary, fairness number, and debiasing result on CelebA and FFHQ depends on that premise.

Editorial extensions

If this is right

  • For any binary classifier with a continuous protected attribute, maximizing $\mathrm{Var}(\Phi(S_P))$ yields the attribute ranges where the model's predictions deviate most from the population base rate, so fairness reporting can target actual discrimination rather than predefined bins.
  • When fairness is monotonic in the attribute, the K-Means approximation provably produces connected segments, so cheap clustering can substitute for the exhaustive search.
  • Using FairGroups groups with the paper's optimal-transport post-processing drops the HGR dependence between predictions and ITA from 0.126 to 0.039, with accuracy moving only from 0.793 to 0.781; the resulting CDFs across groups are closer to aligned but remain modulated by the user parameter $\alpha$.
  • The same partition boundary learned on CelebA transfers to FFHQ with Rand Index 0.921 in 1D ITA and 0.883 in 2D $(L,h)$, indicating that, for a fixed model, the way discrimination is distributed over skin-tone space is stable across datasets.
  • Applying the ground-truth partition to model predictions shows that the darkest skin tones ($L<43$) become significantly more discriminated in the 'Attractive' predictions than in the underlying labels, a bias-amplification effect the paper's segmentation makes visible.

Reading between the lines

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

  • The $\Phi$ criterion is a statistical-parity measure; replacing it with per-group error-rate differences would produce partitions tailored to Equalized Odds, an extension the paper does not test but its machinery supports.
  • Because the variance objective grows with the number of groups, the choice of $K$ is not determined by the method itself; the paper leaves $K$ as a user input, so deployment would need a companion selection rule or stability check.
  • The paper's real-data boundaries are only as trustworthy as the proprietary skin-tone estimator; an equivalent public estimator trained on spectrophotometer ground truth is needed before the exact CelebA and FFHQ conclusions can be reproduced independently.
  • The connected-segment guarantee for K-Means relies on monotonic fairness; for non-monotonic attributes the method can output disconnected groups, so practitioners should first check monotonicity before using the cheap variant.
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

5 major / 4 minor

Summary. The paper proposes FairGroups, a method that partitions a continuous (possibly multidimensional) sensitive attribute into connected groups by maximizing Var(Phi(S_P)), where Phi(k)=P(Y=1|S_P=k)-P(Y=1). A K-Means alternative is proposed for the monotonic-fairness case. The method is validated on synthetic data with known ground-truth partitions, applied to skin-tone fairness on CelebA and FFHQ using a proprietary CIELAB skin-tone estimator, and the resulting partitions are used for optimal-transport-based post-processing debiasing. The paper claims that FairGroups reveals more nuanced discrimination patterns than predefined groups, that these patterns are stable across datasets, and that the debiasing step achieves a better fairness/accuracy trade-off.

Significance. If the claims hold, the paper addresses a real and timely gap: group-fairness auditing for continuous sensitive attributes. The formal objective is clearly stated, Proposition 1 is correct, and the synthetic experiments with known ground truth provide a useful sanity check, with Rand indices of 0.97-0.99 for FairGroups. The code is open access, which is a strength. However, the real-data conclusions are currently limited by two load-bearing issues: the proprietary, unvalidated skin-tone estimator is not independently checkable, and the main real-data evaluation metric Var(Phi(S_P)) is exactly the objective that FairGroups maximizes, making the reported advantage partly true by construction. With external validation of the skin-tone estimates and a stronger reliance on independent metrics (synthetic ground truth, downstream debiasing, cross-dataset stability), the approach could become a practical tool for fairness auditing.

major comments (5)
  1. [§3, §6.2] The entire real-data analysis rests on L*, a*, b* values from a proprietary model [36] that the paper states cannot be publicly shared. Section 3 asserts that this algorithm 'aims to approximate physical measurements from a device' and is therefore free of annotation bias, but no validation against a reference standard is provided in this paper. A bias or miscalibration in the estimator, especially along the L* axis, would shift every group boundary, every Phi value, and all consequent variance, Rand index, and HGR results on CelebA and FFHQ. The central monotonic-fairness, cross-dataset-stability, and debiasing conclusions therefore cannot be independently verified. Please either validate the estimator against a reference (e.g., spectrophotometer or a public skin-tone benchmark) in this paper, or release the derived L*, a*, b* values as a supplementary dataset, or explicitly restrict the real-data claims to the proprietary estimator and discuss the associated risk.
  2. [§4.1, Tables 2, 3, 5] The headline evaluation metric Var(Phi(S_P)) is exactly the objective in Eq. (6) that FairGroups maximizes. Consequently, the larger variance values reported for FairGroups in Tables 2, 3, and 5 are partly true by construction; they do not by themselves establish that the partition is more meaningful. The synthetic Rand-index results and the debiasing experiment in Table 4 provide independent evidence, and the paper would be stronger if the real-data partition-quality claims were based on such independent metrics, or at least accompanied by an explicit discussion of this circularity.
  3. [§5.2, §6.5] The 2D FairGroups algorithm is not specified. Section 5.2 describes the exhaustive search over intervals of a 1D grid and Algorithms 1-2 implement 1D interval counting; Section 5 states that in the 2D case partitions are rectangles, but no algorithm is given for the exhaustive search over rectangles, nor is the grid resolution M for the 2D experiments reported. Section 6.5 then presents 2D FairGroups results on (L, h). Without a precise specification of the search space and the grid, the 2D experiments are not reproducible.
  4. [Appendix C, Eq. (18)-(19)] Equation (19) contains an algebraic error. The expression for Psi_{j1,j2} subtracts a second ratio P(Y=1|L in [lambda_{j1-1}, lambda_{j1}]) from the conditional probability, and the subsequent simplification to [sum_i 1{Y_i=1}1{L_i in [lambda_{j1}, lambda_{j2}]}] / [sum_i 1{L_i in [lambda_{j1}, lambda_{j2}]}] - P(Y=1) is invalid because A/B - C/D does not equal (A-C)/(B-D). Algorithm 2 appears to implement the correct computation (proba_by_range - N_positive/N), so this may be a typographical issue, but the derivation must be corrected because the dynamic programming precalculation is central to the method's efficiency claim.
  5. [Tables 2-5] The real-data tables report point estimates of Var(Phi(S_P)) and Rand indices without confidence intervals, standard errors, or repeated-run variability. Differences such as 0.092 vs. 0.089 in Table 2 may well be within sampling noise. Given finite samples and the grid approximation, please provide bootstrap confidence intervals or standard errors for all reported table metrics, and report the grid size M and number of repetitions used.
minor comments (4)
  1. [§3, §6.2] There are minor typographical issues: in §3 'litterature' should be 'literature', and in §6.2 'women face pictures' should probably be 'female face pictures'.
  2. [§5.2] The definition of the upper triangular matrix U^{Psi}_Lambda states i,j = 1, ..., M-1, but the matrix is said to be of size M x M and Algorithm 2 returns an M x M array; please clarify the indexing and dimensions.
  3. [§6.2] The grid size M used for the skin-tone partitions is not reported anywhere, although Eq. (15) is exact only up to the grid approximation; reporting M is necessary for reproducibility.
  4. [§8] The conclusion states that 'we observed the property of monotonic fairness w.r.t. skin tone' but no statistical test of monotonicity is provided; the claim is based on visual inspection of Figure 4.

Circularity Check

1 steps flagged · score 6.0 of 10

Variance-based evaluation is the optimized objective, so higher Var(Phi) for FairGroups is partly true by construction; independent Rand-index and HGR results prevent full circularity.

  1. self definitional [Sec. 4.1 Eq. (6); Sec. 6.5 Table 3]
    "Our objective is thus to find the partition P over all possible partitions that maximizes: arg max P Var(Φ(SP)). ... In Table 3, we compute the Var(Φ(SP)) values for the two partitions on CelebA and FFHQ, and observe significantly higher values for the one from FairGroups."

    FairGroups is defined by Eq. (6) as the partition maximizing Var(Φ(SP)), and the exhaustive search in Eq. (15) optimizes exactly that objective (up to the grid approximation). Section 6.5 then reports that FairGroups has 'significantly higher' Var(Φ(SP)) as evidence that it reveals how discrimination manifests. This comparison is forced by construction: the chosen partition maximizes the very metric used to evaluate it, so the variance comparisons in Tables 2, 3, and 5 are not independent confirmations. The paper's synthetic Rand-index and debiasing HGR results do provide independent content, but the variance-based evaluation itself is self-definitional.

full rationale

The paper's core algorithm is defined as the maximizer of Var(Φ(SP)), and a substantial part of the real-data evaluation (Tables 2, 3, and 5) compares partitions on exactly that quantity. Reporting that FairGroups attains the highest variance is therefore true by construction rather than an independent empirical finding. This is a genuine but partial circularity: the method also receives independent support from synthetic experiments with known ground truth (Rand index up to 0.99), from cross-dataset stability measured by Rand index, and from debiasing results where HGR is reduced after post-processing. Those evaluations do not reduce to the optimized objective. The proprietary skin-tone estimator from [30,36] is a load-bearing external input whose unavailability raises validity and reproducibility concerns, but it is not a circularity: it is an assumption about the input data, not a claim derived from itself. Overall, the central variance-based validation is partly tautological, so a moderate score is warranted, but the method is not entirely reducible to its own objective.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The central claim rests on the user-chosen number of groups K, the grid discretization, the validity of the proprietary skin-tone measure, and the statistical-parity definition of fairness. No new physical or formal entities are postulated.

free parameters (3)
  • Number of groups K = 5 (synthetic), 6 (ITA), 4 (2D), 2 (appendix)
    K is user-defined and set differently per experiment; no data-driven selection criterion is provided, and all reported results depend on it.
  • Grid size M = not reported
    The exhaustive search and K-Means operate on a discretization grid Lambda of M intervals; M is never specified, so the approximation error of the 'exact up to grid' claim cannot be assessed.
  • Debiasing parameter alpha = not reported
    Equation (16) includes user parameter alpha controlling the fairness-accuracy trade-off; the value used in Table 4 is not given.
assumptions (6)
  • standard math Observations (X, L, Y) are i.i.d. from distribution D.
    Section 4 states this; needed for the variance formulas, delta-method confidence intervals, and the population-level interpretation of Equation (7).
  • domain assumption P(Y=1 | L) is monotonic in L for the skin-tone case (monotonic fairness).
    Definition 1 and Proposition 2 require monotonicity for K-Means groups to be connected segments; the paper says it is observed for the 'Attractive' target on CelebA but does not provide a statistical test of the assumption.
  • ad hoc to paper The proprietary skin-tone estimation algorithm [30,36] yields unbiased L*,a*,b* values.
    Section 6.2 states skin tone is extracted with a proprietary model and the values cannot be shared; no validation of the estimator is given in this paper, so all real-data group boundaries rest on this external tool.
  • domain assumption The one-versus-all difference P(Y=1 | S_P=k) - P(Y=1) is an adequate fairness measure.
    Section 4.2 defines Phi this way, aligning with statistical parity; it ignores error-rate disparities such as Equalized Odds, so the method only targets base-rate differences.
  • domain assumption The grid Lambda is fine enough that maximizing over grid intervals approximates the continuous optimum.
    Section 5.2 says the solution is exact only 'up to the approximation of the partitions search space induced by the grid Lambda'; M is not reported.
  • domain assumption The optimal transport post-processing framework [19,44] achieves statistical parity with minimal accuracy loss.
    Section 7 relies on the Wasserstein-barycenter result from [19] to justify the debiasing step and the fairness-accuracy trade-off.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fairness-Aware Grouping for Continuous Sensitive Variables: Application for Debiasing Face Analysis with respect to Skin Tone." pith.science (2026). https://pith.science/paper/FMTORJFU

@misc{pith2026250711247,
  author       = {Pith},
  title        = {Pith review of: Fairness-Aware Grouping for Continuous Sensitive Variables: Application for Debiasing Face Analysis with respect to Skin Tone},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FMTORJFU}},
  note         = {Machine review of arXiv:2507.11247}
}
read the original abstract

Within a legal framework, fairness in datasets and models is typically assessed by dividing observations into predefined groups and then computing fairness measures (e.g., Disparate Impact or Equality of Odds with respect to gender). However, when sensitive attributes such as skin color are continuous, dividing into default groups may overlook or obscure the discrimination experienced by certain minority subpopulations. To address this limitation, we propose a fairness-based grouping approach for continuous (possibly multidimensional) sensitive attributes. By grouping data according to observed levels of discrimination, our method identifies the partition that maximizes a novel criterion based on inter-group variance in discrimination, thereby isolating the most critical subgroups. We validate the proposed approach using multiple synthetic datasets and demonstrate its robustness under changing population distributions - revealing how discrimination is manifested within the space of sensitive attributes. Furthermore, we examine a specialized setting of monotonic fairness for the case of skin color. Our empirical results on both CelebA and FFHQ, leveraging the skin tone as predicted by an industrial proprietary algorithm, show that the proposed segmentation uncovers more nuanced patterns of discrimination than previously reported, and that these findings remain stable across datasets for a given model. Finally, we leverage our grouping model for debiasing purpose, aiming at predicting fair scores with group-by-group post-processing. The results demonstrate that our approach improves fairness while having minimal impact on accuracy, thus confirming our partition method and opening the door for industrial deployment.

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. OT-FairBoost: Optimal Transport-Guided Gradient Boosting for Fairness Regularization on Tabular Data

    math.ST 2026-07 conditional novelty 6.0 of 10

    Embedding discrete Wasserstein-2 gradients and diagonal Hessians into LightGBM yields stronger accuracy–fairness trade-offs than prior in- and post-processing baselines on classification, regression, and multi-group tasks.

  2. Discovering Geometric Biases in 3D Face Reconstruction: A Curvature-Aware Spectral Framework for Fairness Evaluation

    cs.CV 2026-07 conditional novelty 6.0 of 10

    A curvature-aware spectral framework using the Laplace-Beltrami operator detects demographic biases in 3D face reconstruction that Euclidean metrics miss, validated by a user study showing 73.6% perceptual accuracy.

Reference graph

Works this paper leans on

47 extracted references · 39 canonical work pages · cited by 2 Pith papers

  1. [36]

    Robin, T

    K. Robin, T. Loïc, E. Malherbe, and M. Perrot. Beyond color correction: Skin color estimation in the wild through deep learning. Electronic Imaging, 32:1–8, 2020

    work page 2020

  2. [1]

    R. Berk, H. Heidari, S. Jabbari, M. Kearns, and A. Roth. Fairness in criminal justice risk assessments: The state of the art. Sociological Methods & Research, 50(1):3–44, 2021

    work page 2021

  3. [2]

    Besse, E

    P. Besse, E. del Barrio, P. Gordaliza, J.-M. Loubes, and L. Risser. A survey of bias in machine learning through the prism of statistical parity. The American Statistician, 76(2), 2022

    work page 2022

  4. [3]

    Buolamwini and T

    J. Buolamwini and T. Gebru. Gender shades: Intersectional accuracy disparities in commercial gender classification. In Conference on fair- ness, accountability and transparency, pages 77–91. PMLR, 2018

    2018

  5. [4]

    Caton and C

    S. Caton and C. Haas. Fairness in machine learning: A survey. ACM Computing Surveys, 56(7):1–38, 2024

    work page 2024

  6. [5]

    Chardon, I

    A. Chardon, I. Cretois, and C. Hourseau. Skin colour typology and suntanning pathways. International journal of cosmetic science, 1991

    work page 1991

  7. [6]

    J. Cho, A. Zala, and M. Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generation models. InProceedings of the IEEE/CVF International Conference on Computer Vision, 2023

    work page 2023

  8. [7]

    Chouldechova and A

    A. Chouldechova and A. Roth. A snapshot of the frontiers of fairness in machine learning. Communications of the ACM, 63(5):82–89, 2020

    work page 2020

Show all 47 references
  1. [8]

    Corbett-Davies, E

    S. Corbett-Davies, E. Pierson, A. Feller, S. Goel, and A. Huq. Algo- rithmic decision making and the cost of fairness. In Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining, pages 797–806, 2017

    2017

  2. [9]

    De Rigal, M.-L

    J. De Rigal, M.-L. Abella, F. Giron, L. Caisey, and M. A. Lefebvre. Development and validation of a new skin color chart®. Skin Research and Technology, 13(1):101–109, 2007

    2007

  3. [10]

    Del Barrio, P

    E. Del Barrio, P. Gordaliza, and J.-M. Loubes. Review of mathe- matical frameworks for fairness in machine learning. arXiv preprint arXiv:2005.13755, 2020

    2005 arXiv

  4. [11]

    Dwork, M

    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, pages 214–226, 2012

    2012

  5. [12]

    A. E. Ezugwu, A. M. Ikotun, O. O. Oyelade, L. Abualigah, J. O. Agushaka, C. I. Eke, and A. A. Akinyelu. A comprehensive survey of clustering algorithms: State-of-the-art machine learning applications, taxonomy, challenges, and future research prospects. Engineering Ap- plicati...

    2022

  6. [13]

    Feldman, S

    M. Feldman, S. A. Friedler, J. Moeller, C. Scheidegger, and S. Venkata- subramanian. Certifying and removing disparate impact. In proceed- ings of the 21th ACM SIGKDD international conference on knowledge discovery and data mining, pages 259–268, 2015

    2015

  7. [14]

    T. B. Fitzpatrick. The validity and practicality of sun-reactive skin types i through vi. Archives of dermatology, 124(6):869–871, 1988

    1988

  8. [15]

    Garcia and M

    D. Garcia and M. Abascal. Colored perceptions: Racially distinctive names and assessments of skin color. American Behavioral Scientist , 60(4):420–441, 2016

    2016

  9. [16]

    Gevers, A

    T. Gevers, A. Gijsenij, J. Van de Weijer, and J.-M. Geusebroek.Color in computer vision: fundamentals and applications . John Wiley & Sons, 2012

    2012

  10. [17]

    Giuliani, E

    L. Giuliani, E. Misino, and M. Lombardi. Generalized disparate impact for configurable fairness solutions in ml. In International Conference on Machine Learning, pages 11443–11458. PMLR, 2023

    2023

  11. [18]

    Gordaliza, E

    P. Gordaliza, E. Del Barrio, G. Fabrice, and J.-M. Loubes. Obtaining fairness using optimal transport theory. In International conference on machine learning, pages 2357–2365. PMLR, 2019

    2019

  12. [19]

    T. L. Gouic, J.-M. Loubes, and P. Rigollet. Projection to fairness in statistical learning. arXiv preprint arXiv:2005.11720, 2020

    2005 arXiv

  13. [20]

    Gustafson, C

    L. Gustafson, C. Rolland, N. Ravi, Q. Duval, A. Adcock, C.-Y . Fu, M. Hall, and C. Ross. Facet: Fairness in computer vision evaluation benchmark. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20370–20382, 2023

    2023

  14. [21]

    K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

    2016

  15. [22]

    Kamishima, S

    T. Kamishima, S. Akaho, and J. Sakuma. Fairness-aware learning through regularization approach. In 2011 IEEE 11th international con- ference on data mining workshops, pages 643–650. IEEE, 2011

    2011

  16. [23]

    Karkkainen and J

    K. Karkkainen and J. Joo. Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. InProceed- ings of the IEEE/CVF winter conference on applications of computer vision, pages 1548–1558, 2021

    2021

  17. [24]

    Kearns, S

    M. Kearns, S. Neel, A. Roth, and Z. S. Wu. Preventing fairness ger- rymandering: Auditing and learning for subgroup fairness. In Interna- tional conference on machine learning. PMLR, 2018

    2018

  18. [25]

    N. M. Kinyanjui, T. Odonga, C. Cintas, N. C. Codella, R. Panda, P. Sat- tigeri, and K. R. Varshney. Fairness of classifiers across skin tones in dermatology. In International Conference on Medical Image Comput- ing and Computer-Assisted Intervention. Springer, 2020

    2020

  19. [26]

    X. Li, Z. Chen, J. M. Zhang, F. Sarro, Y . Zhang, and X. Liu. Dark-skin individuals are at more risk on the street: Unmasking fairness issues of autonomous driving systems. arXiv preprint arXiv:2308.02935, 2023

    2023 arXiv

  20. [27]

    Z. Liu, P. Luo, X. Wang, and X. Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), December 2015

    2015

  21. [28]

    B. C. K. Ly, E. B. Dyer, J. L. Feig, A. L. Chien, and S. Del Bino. Re- search techniques made simple: cutaneous colorimetry: a reliable tech- nique for objective skin color measurement. Journal of Investigative Dermatology, 140(1):3–12, 2020

    2020

  22. [29]

    MacQueen

    J. MacQueen. Some methods for classification and analysis of multi- variate observations. In Proceedings of Berkeley Symposium on Mathe- matical Statistics and Probability/University of California Press, 1967

    1967

  23. [30]

    Malherbe, M

    E. Malherbe, M. Remise, S. Zhang, and M. Perrot. Skin tone diagnosis in the wild: Towards more robust and inclusive user experience using oriented aleatoric uncertainty. In Proceedings of the Asian Conference on Computer Vision, pages 1–19, 2022

    2022

  24. [31]

    J. Mary, C. Calauzenes, and N. El Karoui. Fairness-aware learning for continuous attributes and treatments. In International Conference on Machine Learning, pages 4382–4391. PMLR, 2019

    2019

  25. [32]

    B. Maze, J. Adams, J. A. Duncan, N. Kalka, T. Miller, C. Otto, A. K. Jain, W. T. Niggel, J. Anderson, J. Cheney, et al. Iarpa janus benchmark- c: Face dataset and protocol. In 2018 international conference on bio- metrics (ICB), pages 158–165. IEEE, 2018

    2018

  26. [33]

    Merler, N

    M. Merler, N. Ratha, R. S. Feris, and J. R. Smith. Diversity in faces. arXiv preprint arXiv:1901.10436, 2019

    1901 arXiv

  27. [34]

    W. M. Rand. Objective criteria for the evaluation of clustering methods. Journal of the American Statistical association, 66(336), 1971

    1971

  28. [35]

    Risser, A

    L. Risser, A. G. Sanz, Q. Vincenot, and J.-M. Loubes. Tackling algo- rithmic bias in neural-network classifiers using wasserstein-2 regular- ization. Journal of Mathematical Imaging and Vision, 64(6), 2022

    2022

  29. [37]

    Schumann, F

    C. Schumann, F. Olanubi, A. Wright, E. Monk, C. Heldreth, and S. Ricco. Consensus and subjectivity of skin tone annotation for ml fairness. Advances in Neural Information Processing Systems, 36, 2024

    2024

  30. [38]

    R. Sève. Science de la couleur: Aspects physiques et perceptifs . Cha- lagam éd., 2009

    2009

  31. [39]

    L. Sweeney. Discrimination in online ad delivery. Communications of the ACM, 56(5):44–54, 2013

    2013

  32. [40]

    Thong, P

    W. Thong, P. Joniak, and A. Xiang. Beyond skin tone: A multidimen- sional measure of apparent skin color. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4903–4913, 2023

    2023

  33. [41]

    X. Wang, Y . Zhang, and R. Zhu. A brief review on algorithmic fairness. Management System Engineering, 1(1):7, 2022

    2022

  34. [42]

    I. L. Weatherall and B. D. Coombs. Skin color measurements in terms of cielab color space values. Journal of investigative dermatology, 1992

    1992

  35. [43]

    Wilkes, C

    M. Wilkes, C. Y . Wright, J. L. du Plessis, and A. Reeder. Fitzpatrick skin type, individual typology angle, and melanin index in an african population: steps toward universally applicable skin photosensitivity as- sessments. JAMA dermatology, 151(8):902–903, 2015

    2015

  36. [44]

    R. Xian, L. Yin, and H. Zhao. Fair and optimal classification via post- processing. In International conference on machine learning , pages 37977–38012. PMLR, 2023

    2023

  37. [45]

    Xu and D

    R. Xu and D. Wunsch. Survey of clustering algorithms. IEEE Transac- tions on neural networks, 16(3):645–678, 2005

    2005

  38. [46]

    M. Zhu. Recall, precision and average precision. Department of Statis- tics and Actuarial Science, University of Waterloo, Waterloo , 2(30):6, 2004. A Relation with Disparate Impact in the Binary Case Detailed Proof of Proposition 1. Proof. When SP is binary, that is, there ar...

    2004

  39. [47]

    predefined

    × (M − 1) with ones on the main diagonal and in all elements above it. Therefore, we have Ψj1,j2 ≈ PN i=1 1{Yi=1}1{Li∈[λj1 −1,λj2 ]} PN i=1 1{Li∈[λj1 −1,λj2 ]} − − PN i=1 1{Yi=1}1{Li∈[λj1 −1,λj1 ]} PN i=1 1{Li∈[λj1 −1,λj1 ]} − PN i=1 1{Yi=1} N = = PN i=1 1{Yi=1} 1{Li∈[λj1 −1,λ...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.