Pith. sign in

REVIEW 4 major objections 5 minor 37 references

SOMtime the World Ain$'$t Fair: Violating Fairness Using Self-Organizing Maps

T0 review · 4 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read Purely unsupervised self-organizing maps recover withheld age and income as dominant embedding axes, with Spearman correlations up to 0.85.

desk verdict The SOMtime idea is worth a referee, but the headline claim is undermined by the paper's own use of the sensitive attribute in preprocessing, and the abstract promises a cluster-skew result the paper never reports. read the letter →

arxiv 2602.18201 v2 pith:LKJRK2YK submitted 2026-02-20 cs.AI cs.LG

Joseph Bingham , Netanel Arussy , Dvir Aran This is my paper · ORCID
classification cs.AIcs.LG
keywords fairnessthroughunawarenessself-organizingmapssensitiveattributeleakageunsupervisedrepresentationlearningglobalgeometricorderingdimensionalityreductionauditingordinalattributes
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

This paper claims that purely unsupervised, topology-preserving representations—specifically high-capacity self-organizing maps—arrange data along monotonic axes aligned with sensitive attributes such as age and income, even when those attributes are withheld from training. On the World Values Survey and Census-Income datasets, the method achieves Spearman correlations up to 0.85, while PCA, UMAP, t-SNE, and autoencoders stay below 0.34. This would mean the common assumption of fairness through unawareness fails at the representation level: any downstream use of such an embedding, whether clustering, visualization, or feature extraction, inherits demographic skew before a supervised task exists. The paper positions SOMtime as an audit tool that reveals this leakage, not as a fairness mitigation.

What carries the argument

The central object is a self-organizing map (SOM): a square lattice of prototype vectors trained by competitive learning with a Gaussian neighborhood function, scaled to high capacity with lattice size K = 5·N^0.54. Each observation is embedded in three dimensions using its best-matching unit's (x, y) lattice coordinates plus its quantization error z = ||x - w_bmu||, the distance to its prototype. A trajectory-adjacency algorithm links cluster centroids in this 3D space to recover the path along which a sensitive attribute is monotonically ordered, and leakage is quantified as the maximum absolute Spearman correlation between any single embedding axis and the withheld attribute.

What would settle it

Train SOMtime on data where the sensitive attribute is genuinely independent of all input features, or on synthetic data with known generative structure, with no feature selection or preprocessing informed by the attribute. If the Spearman correlation between the embedding axes and the withheld attribute stays near zero, the claim that SOMs systematically expose such attributes would be refuted. Alternatively, rerun the Census experiment on the full 40 columns without removing age- and income-correlated features; if correlations do not rise above baseline, the reported gap depends on preproces

Watch

Extended reading notes

Core claim

On two real-world tabular datasets, with age, income, and capital gains withheld from all training, a high-capacity self-organizing map arranges observations so that the withheld attributes vary monotonically along the learned lattice, reaching Spearman correlations of 0.85 for age on the WVS Canada subset and 0.83 for age on Census-Income. PCA, UMAP, t-SNE, and autoencoders, including a capacity-matched 1.4M-parameter autoencoder with near-perfect reconstruction, stay below 0.34. The authors interpret this as evidence that sensitive attributes are not merely extractable by a probing classifier but are dominant organizing axes of the representation, a form of ambient leakage that affects all

Load-bearing premise

The sensitive attribute is truly withheld from every step that shapes the representation; however, Section 4.2 admits that Census columns strongly correlated with age and income were removed and WVS questions were selected based on their age correlation, so the reported leakage is not purely emergent from unsupervised learning.

Editorial extensions

If this is right

  • Fairness auditing must extend beyond supervised predictors to unsupervised embeddings, because structured leakage exists before any task is defined.
  • Probing classifiers and SOMtime are complementary: probes measure worst-case extractability, while SOMtime measures whether a sensitive attribute is a dominant organizing principle of the representation.
  • Unsupervised segmentation of such embeddings produces demographically skewed clusters, so downstream clustering and recommendation systems inherit fairness risks without any supervised objective.
  • A capacity-matched autoencoder with near-perfect reconstruction does not reproduce the leakage, suggesting the SOM's topological inductive bias, not parameter count, drives the effect.
  • Practitioners should run representation-level leakage audits, including correlation checks and global ordering analysis, before deploying any unsupervised embedding in decision-making pipelines.

Reading between the lines

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

  • The paper's own preprocessing (Section 4.2) removes Census columns that correlate strongly with age and income and selects WVS questions based on their correlation with age, meaning the reported leakage is partly guided by the sensitive attribute. A cleaner test that withholds the attribute before any feature selection would make the emergence claim much stronger.
  • The monotonic-ordering definition targets ordinal attributes like age and income; the same mechanism may produce categorical separation (e.g., race or gender) as discontinuous gradients, and a categorical auditing metric would test whether the effect extends beyond ordinals.
  • SOMs have a fixed lattice resolution, so the magnitude of leakage likely depends on sample size and lattice size; a systematic sensitivity analysis across map dimensions would show how robust the effect is.
  • If representation-level audits become standard, the paper's suggested mitigations—such as demographic entropy balancing per neuron or penalizing sensitive-attribute gradients across the lattice—could be tested within the same SOM framework, turning the audit tool into a fairness-aware representation learner.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes SOMtime, an auditing method based on high-capacity Self-Organizing Maps, and claims that purely unsupervised topology-preserving embeddings can recover withheld ordinal sensitive attributes (age, income, capital gains) as monotonic latent axes. On the WVS and Census-Income datasets, the authors report Spearman correlations up to 0.85 for SOMtime versus at most 0.34 for PCA, UMAP, t-SNE, and autoencoders. They conclude that fairness-through-unawareness fails at the representation level and that unsupervised representations should be routinely audited. The paper includes a trajectory-extraction algorithm (Algorithm 1) and reports recovery accuracy of age-group orderings.

Significance. If the central claim were established, the paper would make a useful contribution to fairness auditing by showing that a classical topology-preserving method can expose global monotonic structure aligned with sensitive attributes, a form of leakage that correlation-based probing may miss. The paper also provides code and reports stability across five runs with low standard deviations, which are strengths. However, the validity of the headline comparison rests on two conditions: that the sensitive attribute is truly withheld from every step shaping the representation, and that the evaluation is symmetric across methods. Both conditions are violated in the current manuscript, so the significance of the empirical findings is not yet demonstrated.

major comments (4)
  1. [§4.2, §3.1] The central premise that sensitive attributes are 'withheld from all representation learning' is contradicted by the dataset preprocessing. §3.1 states that sensitive attributes are used only for post-hoc auditing, but §4.2 describes removing Census columns with strong correlations to age/income to avoid 'trivial' recovery, and Table 1 justifies WVS question selection by correlation with age ('justifying our choice for their inclusion'). This means the input feature set was chosen using knowledge of the sensitive attribute. Consequently, the reported SOMtime correlations may reflect feature engineering rather than emergent leakage from purely unsupervised learning, and the abstract's claim that sensitive attributes 'emerge ... even when explicitly excluded from the input' is not supported.
  2. [§4.3, Table 2 caption] The evaluation is asymmetric. The Table 2 caption states 'Best values from ablations used,' meaning baseline hyperparameters were selected to maximize their reported correlations, while SOMtime uses fixed hyperparameters. Additionally, §4.3 states that for SOMtime only the Z-axis (activation) is examined, whereas for the other methods the maximum over every embedding dimension is reported. Since Table 2 reports 'maximum correlation between any single embedding axis,' the comparison is not apples-to-apples: baselines are given the advantage of 50 axes plus ablated hyperparameters, while SOMtime is restricted to one axis. This undermines the quantitative claim of a 3–8× improvement.
  3. [§3.5, §5.2] The method description is internally inconsistent about which axis is used. §3.5 says the maximum absolute correlation across the three embedding dimensions is reported, but Experiment 2 (§5.2) says for SOMtime 'only looking at the Z-axis,' while baselines use the 'strongest principal component.' The reader cannot determine whether Table 2 reports the max over (x,y,z) for SOMtime or only z, nor how the 'dominant 1D axis' was chosen for baselines. This needs clarification and a consistent protocol.
  4. [Abstract, §5.3, §6] The abstract and discussion claim that 'unsupervised segmentation of SOMtime embeddings produces demographically skewed clusters,' but no clustering experiment or quantitative cluster-skew result appears in the paper. Table 3 reports recovery of age-group ordering edges, not cluster demographic skew. This claim is unsupported and should either be removed or backed by an actual clustering experiment with a measurable imbalance metric.
minor comments (5)
  1. [§3.4, Algorithm 1] The description of the adjacency-graph construction is hard to follow: 'values of the graph are boolean identifiers of edge existence from row to column vertices' is vague, and the scaling operation in lines 14–15 uses sums over centroids without explaining why this normalization is appropriate. Pseudocode would benefit from concrete notation and a worked example.
  2. [References] Reference [31] has a malformed author list ('Yutaka Matsuo Yusuke Iwasawa, Kotaro Nakayama' appears to be three authors concatenated). Please correct.
  3. [Figure 2] The figure shows a qualitative visualization of age ordering but does not include a colorbar legend or axis labels. Since the paper emphasizes quantitative correlation, a version with the SOM grid coordinates and the recovered trajectory overlaid would be more informative.
  4. [§4.1] The sentence 'A selection criteria implemented here for determining questions to be used was that the data does not contain missing values' has subject-verb agreement errors and is unclear; presumably it means 'questions were selected for having no missing values.'
  5. [Section 7.1] The limitation paragraph correctly notes that only ordinal attributes are studied, but it does not mention the more serious limitation that feature selection used the sensitive attribute; this should be acknowledged.

Circularity Check

1 steps flagged · score 4.0 of 10

Target attribute was used in feature-selection preprocessing, contradicting the 'withheld' premise of the headline leakage claim.

  1. fitted input called prediction [§3.1 (Problem Setup) vs §4.2 (Datasets)]
    "§3.1: 'Sensitive attributes are used only for post-hoc auditing via correlation analysis.' §4.2: 'Although removing columns or features that are correlated with the sensitive features would require knowledge beforehand, we chose to include this portion of our preprocessing... We only ever use knowledge of the sensitive feature in this context.'"

    The paper's central claim is that sensitive attributes 'emerge' in a purely unsupervised embedding from which they are 'withheld.' But the input feature set was not constructed without s: WVS feature inclusion is justified in Table 1 by correlations with age, and Census columns strongly correlated with age/income were deliberately removed. s therefore shaped the representation's input distribution before any SOM was trained. The subsequent 'recovery' of s via Spearman correlation is not evidence of emergent leakage from an s-agnostic pipeline; it is in part a consequence of investigator knowledge of s being baked into the feature selection. The headline comparison (SOMtime 0.85 vs baselines <0.23) does not isolate unsupervised learning from target-informed preprocessing.

full rationale

No equation-level derivation is present, so the paper is not circular in the sense of Eq. X reducing to Eq. Y. The SOM training itself does not receive s, and the 0.85 correlation is not literally forced by the feature-selection steps (if anything, removing correlated Census columns makes recovery harder). However, the central load-bearing premise—that s is 'used only for post-hoc auditing'—is directly contradicted by §4.2, where s is used to choose and filter input features. This means the reported leakage is not independent of the target attribute at the preprocessing stage. Because the core empirical comparison still depends on actual unsupervised training and is not a parameter fit, I assign a moderate score of 4 rather than 6–8. The Z-axis/Algorithm 1 ordering is essentially a reformulation of the Z-axis Spearman correlation, but it is not circular because s is not used to define the path.

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

No new physical entities are introduced; the only named artifact is the SOMtime procedure, which is a method rather than an entity. The main hidden inputs are feature-selection decisions and ablation choices that use the sensitive attribute.

free parameters (4)
  • SOM lattice size constants (5 and 0.54 in K = 5·N^0.54) = 5, 0.54
    Chosen via a 'slight modification' of the Vesanto-Alhoniemi heuristic; determines SOM capacity and directly affects how much fine-grained structure, including sensitive structure, is preserved.
  • SOM training hyperparameters (sigma=0.7, learning rate=0.75) = 0.7, 0.75
    Fixed across experiments; they shape neighborhood smoothing and convergence, influencing whether monotonic gradients appear on the lattice.
  • Baseline hyperparameters selected by ablation = unspecified ('best values from ablations used')
    Table 2 caption says best values from ablations were used; selecting the configuration that maximizes each baseline's leakage turns the comparison into a search over free parameters.
  • Z-axis scaling in Algorithm 1 (multiplication by K) = K (lattice dimension)
    Line 15 scales activation by K before distance computation; this hand-specified weighting changes which trajectory edges are found and is not derived from data.
assumptions (4)
  • standard math SOM training yields a topology-preserving discretization of the data manifold.
    Relied on in §3.2; standard Kohonen SOM property.
  • domain assumption Dataset features contain recoverable signal about withheld age/income.
    The experiments require that survey/census responses encode age-related structure; Table 1 is used to justify chosen questions.
  • ad hoc to paper Sensitive attributes are used only for post-hoc auditing.
    Stated in §3.1 but directly contradicted by §4.2 preprocessing that removes features correlated with the sensitive attribute.
  • domain assumption Maximum per-axis correlation is an appropriate leakage measure.
    The paper defines leakage as the max |rho| over embedding axes (§3.5), assuming a single aligned axis constitutes a fairness risk.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SOMtime the World Ain$'$t Fair: Violating Fairness Using Self-Organizing Maps." pith.science (2026). https://pith.science/paper/LKJRK2YK

@misc{pith2026260218201,
  author       = {Pith},
  title        = {Pith review of: SOMtime the World Ain$'$t Fair: Violating Fairness Using Self-Organizing Maps},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LKJRK2YK}},
  note         = {Machine review of arXiv:2602.18201}
}
read the original abstract

Unsupervised representations are widely assumed to be neutral with respect to sensitive attributes when those attributes are withheld from training. We show that this assumption is false. Using SOMtime, a topology-preserving representation method based on high-capacity Self-Organizing Maps, we demonstrate that sensitive attributes such as age and income emerge as dominant latent axes in purely unsupervised embeddings, even when explicitly excluded from the input. On two large-scale real-world datasets (the World Values Survey across five countries and the Census-Income dataset), SOMtime recovers monotonic orderings aligned with withheld sensitive attributes, achieving Spearman correlations of up to 0.85, whereas PCA and UMAP typically remain below 0.23 (with a single exception reaching 0.31), and against t-SNE and autoencoders which achieve at most 0.34. Furthermore, unsupervised segmentation of SOMtime embeddings produces demographically skewed clusters, demonstrating downstream fairness risks without any supervised task. These findings establish that \textit{fairness through unawareness} fails at the representation level for ordinal sensitive attributes and that fairness auditing must extend to unsupervised components of machine learning pipelines. We have made the code available at~ https://github.com/JosephBingham/SOMtime

Figures

Figures reproduced from arXiv: 2602.18201 by the authors.

Figure 1
Figure 1. The SOMtime pipeline. (a) High-dimensional tabular data with sensitive attributes withheld. (b) A high-capacity [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Example of sensitive attribute emergence SOMtime. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 7 linked inside Pith

  1. [1]

    Census-Income (KDD)

    2000. Census-Income (KDD). UCI Machine Learning Repository. DOI: https://doi.org/10.24432/C5N30T

    doi:10.24432/c5n30t 2000

  2. [2]

    Farid Adilazuarda, Chen Cecilia Liu, Iryna Gurevych, and Alham Fikri Aji. 2025. From Surveys to Narratives: Rethinking Cultural Value Adaptation in LLMs. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, Christos Christodoulopoulos, Tanmoy Chakraborty, Carolyn Rose, and Violet Peng (Eds.). Association for Computati...

    doi:10.18653/v1/2025.emnlp-main.912 2025

  3. [3]

    Sara Ahmadian, Alessandro Epasto, Ravi Kumar, and Mohammad Mahdian. 2019. Clustering without Over-Representation. InProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining(Anchorage, AK, USA)(KDD ’19). Association for Computing Machinery, New York, NY, USA, 267–275. doi:10.1145/3292500.3330987

    arXiv 2019

  4. [4]

    Hyun-Jung Bae, Jong-Seong Park, Ji-Hyeok Choi, and Hyuk-Yoon Kwon. 2025. Learning model combined with data clustering and dimensionality reduction for short-term electricity load forecasting.Sci. Rep.15, 1 (Jan. 2025), 3575

    2025

  5. [5]

    Anshuman Chhabra, Karina Masalkovait ˙e, and Prasant Mohapatra. 2021. An Overview of Fairness in Clustering.IEEE Access9 (2021), 130698–130720. doi:10. 1109/ACCESS.2021.3114099

    arXiv 2021

  6. [6]

    Flavio Chierichetti, Ravi Kumar, Silvio Lattanzi, and Sergei Vassilvitskii. 2017. Fair clustering through fairlets. InProceedings of the 31st International Conference on Neural Information Processing Systems(Long Beach, California, USA)(NIPS’17). Curran Associates Inc., Red Hook, NY, USA, 5036–5044

    2017

  7. [7]

    Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Reingold, and Richard Zemel. 2012. Fairness through awareness. InProceedings of the 3rd Innovations in Theoretical Computer Science Conference(Cambridge, Massachusetts)(ITCS ’12). Association for Computing Machinery, New York, NY, USA, 214–226. doi:10. 1145/2090236.2090255

    arXiv 2012

  8. [8]

    Michael D Ekstrand and Daniel Kluver. 2021. Exploring author gender in book rating and recommendation.User Model. User-adapt Interact.31, 3, 377–420

    2021

Show all 37 references
  1. [9]

    Mehrdad Ghadiri, Samira Samadi, and Santosh Vempala. 2021. Socially Fair k-Means Clustering. InProceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency(Virtual Event, Canada)(FAccT ’21). Association for Computing Machinery, New York, NY, USA, 438–44...

    2021 doi

  2. [10]

    Axel Guérin, Pierre Chauvet, and Frédéric Saubion. 2024. A Survey on Recent Advances in Self-Organizing Maps. arXiv:2501.08416 [cs.NE] https://arxiv.org/ abs/2501.08416

    2024 arXiv

  3. [11]

    Christian Haerpfer, Ronald Inglehart, Alejandro Moreno, Christian Welzel, Kseniya Kizilova, Jaime Diez-Medrano, Marta Lagos, Pippa Norris, Eduard Ponarin, and Bi Puranen. 2020. World Values Survey wave 7 (2017-2020) cross- national data-set

    2020

  4. [12]

    G E Hinton and R R Salakhutdinov. 2006. Reducing the dimensionality of data with neural networks.Science313, 5786 (July 2006), 504–507

    2006

  5. [13]

    Bo Kang, Darío García García, Jefrey Lijffijt, Raúl Santos-Rodríguez, and Tijl de Bie. 2021. Conditional t-SNE: More informative t-SNE embeddings. In2021 IEEE 8th International Conference on Data Science and Advanced Analytics (DSAA). 1–2. doi:10.1109/DSAA53316.2021.9564212

    2021

  6. [14]

    Kleinberg, Sendhil Mullainathan, and Manish Raghavan

    Jon M. Kleinberg, Sendhil Mullainathan, and Manish Raghavan. 2016. Inherent Trade-Offs in the Fair Determination of Risk Scores.CoRRabs/1609.05807 (2016). arXiv:1609.05807 http://arxiv.org/abs/1609.05807

    2016 arXiv

  7. [15]

    Matthäus Kleindessner, Samira Samadi, Pranjal Awasthi, and Jamie Morgenstern

  8. [16]

    T. Kohonen. 1990. The self-organizing map.Proc. IEEE78, 9 (1990), 1464–1480. doi:10.1109/5.58325

    1990 doi

  9. [17]

    Cheng Li, Mengzhuo Chen, Jindong Wang, Sunayana Sitaram, and Xing Xie. 2024. CultureLLM: Incorporating Cultural Differences into Large Language Models. InAdvances in Neural Information Processing Systems, A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and...

    2024 doi

  10. [18]

    Shuhan Liu, Zheyun Qin, Xuan Hou, Yining Wang, Ziwen Wang, and Zhao- hui Peng. 2026. Fairness-aware graph representation learning through bias disentanglement.Information and Software Technology193 (2026), 108034. doi:10.1016/j.infsof.2026.108034

    2026

  11. [19]

    Christos Louizos, Kevin Swersky, Yujia Li, Max Welling, and Richard Zemel

  12. [20]

    Andrzej Maćkiewicz and Waldemar Ratajczak. 1993. Principal components analysis (PCA).Comput. Geosci.19, 3 (March 1993), 303–342

    1993

  13. [21]

    David Madras, Elliot Creager, Toniann Pitassi, and Richard S. Zemel. 2018. Learn- ing Adversarially Fair and Transferable Representations.CoRRabs/1802.06309 (2018). arXiv:1802.06309 http://arxiv.org/abs/1802.06309

    2018 arXiv

  14. [22]

    Leland McInnes, John Healy, and James Melville. 2020. UMAP: Uni- form Manifold Approximation and Projection for Dimension Reduction. arXiv:1802.03426 [stat.ML] https://arxiv.org/abs/1802.03426

    2020 arXiv

  15. [23]

    Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, and Aram Galstyan. 2021. A Survey on Bias and Fairness in Machine Learning.ACM Comput. Surv.54, 6, Article 115 (July 2021), 35 pages. doi:10.1145/3457607

    2021 doi

  16. [24]

    Daniel Moyer, Shuyang Gao, Rob Brekelmans, Greg Ver Steeg, and Aram Galstyan

  17. [25]

    Jaakko Peltonen, Wen Xu, Timo Nummenmaa, and Jyrki Nummenmaa. 2023. Fair Neighbor Embedding. InProceedings of the 40th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 202), Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Enge...

    2023

  18. [26]

    Ricardo Trainotti Rabonato and Lilian Berton. 2024. A systematic review of fairness in Machine Learning.AI and Ethics5, 3 (Sep 2024), 1943–1954. doi:10. 1007/s43681-024-00577-5

    2024

  19. [27]

    Shauli Ravfogel, Yanai Elazar, Hila Gonen, Michael Twiton, and Yoav Goldberg

  20. [28]

    Samira Samadi, Uthaipon Tao Tantipongpipat, Jamie Morgenstern, Mohit Singh, and Santosh S. Vempala. 2018. The Price of Fair PCA: One Extra Dimension. CoRRabs/1811.00103 (2018). arXiv:1811.00103 http://arxiv.org/abs/1811.00103

    2018 arXiv

  21. [29]

    Laurens van der Maaten and Geoffrey Hinton. 2008. Visualizing Data using t-SNE.Journal of Machine Learning Research9, 86 (2008), 2579–2605. http: //jmlr.org/papers/v9/vandermaaten08a.html

    2008

  22. [30]

    Vesanto and E

    J. Vesanto and E. Alhoniemi. 2000. Clustering of the self-organizing map.IEEE Transactions on Neural Networks11, 3 (2000), 586–600. doi:10.1109/72.846731

    2000 doi

  23. [31]

    Yutaka Matsuo Yusuke Iwasawa, Kotaro Nakayama. 2018. Censoring Representa- tions with Multiple-Adversaries over Random Subspaces. https://openreview. net/forum?id=ByuP8yZRb

    2018

  24. [32]

    Rich Zemel, Yu Wu, Kevin Swersky, Toni Pitassi, and Cynthia Dwork. 2013. Learning Fair Representations. InProceedings of the 30th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 28), Sanjoy Dasgupta and David McAllester (Eds.). PMLR...

    2013

  25. [33]

    Wenlong Zhao, Debanjan Mondal, Niket Tandon, Danica Dillion, Kurt Gray, and Yuling Gu. 2024. WorldValuesBench: A Large-Scale Benchmark Dataset for Multi-Cultural Value Awareness of Language Models. arXiv:2404.16308 [cs.CL] https://arxiv.org/abs/2404.16308 KDD ’25, August 3–7, ...

    2024 arXiv

  26. [2017]

    arXiv:1511.00830 [stat.ML] https: //arxiv.org/abs/1511.00830

    The Variational Fair Autoencoder. arXiv:1511.00830 [stat.ML] https: //arxiv.org/abs/1511.00830

    arXiv

  27. [2018]

    InProceedings of the 32nd International Conference on Neural Information Processing Systems(Montréal, Canada)(NIPS’18)

    Invariant representations without adversarial training. InProceedings of the 32nd International Conference on Neural Information Processing Systems(Montréal, Canada)(NIPS’18). Curran Associates Inc., Red Hook, NY, USA, 9102–9111

  28. [2019]

    InInterna- tional Conference on Machine Learning

    Guarantees for Spectral Clustering with Fairness Constraints. InInterna- tional Conference on Machine Learning. https://api.semanticscholar.org/CorpusID: 59291890

  29. [2020]

    InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics, Dan Jurafsky, Joyce Chai, Natalie Schluter, and Joel Tetreault (Eds.)

    Null It Out: Guarding Protected Attributes by Iterative Nullspace Projection. InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics, Dan Jurafsky, Joyce Chai, Natalie Schluter, and Joel Tetreault (Eds.). Association for Computational Linguis...

    2020 doi

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.