Pith. sign in

REVIEW 5 major objections 6 minor 30 references

Towards Group Fairness with Multiple Sensitive Attributes in Federated Foundation Models

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A sensitive attribute's causal effect on the ground-truth label determines how hard its bias is to remove in federated foundation models, and debiasing one attribute can worsen another.

desk verdict A sensible multi-attribute fairness extension with an overclaimed causal law: the causal effects are never identified as written. read the letter →

arxiv 2506.18732 v1 pith:VUIXGROB submitted 2025-06-23 cs.LG

classification cs.LG
keywords federatedlearningfoundationmodelsgroupfairnessmultiplesensitiveattributescausalinferencediscoverydemographicparityequalizedodds
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that in federated foundation models, group fairness across several sensitive attributes at once is governed by causal structure: a sensitive attribute's total causal effect on the ground-truth label predicts how stubbornly its bias resists correction, and debiasing one attribute can make another worse. To test this, the authors extend an existing fairness-aware federated fine-tuning method into a multi-attribute version that feeds demographic-only prompts to the model in parallel and combines per-attribute fairness regularizers with tunable weights. They then compute causal effect estimates for each sensitive attribute on CelebA and FairFace face-recognition tasks and compare those estimates with the observed fairness improvements. The paper's conclusion is that the larger the causal effect of an attribute on the label, the harder it is to ensure group fairness for that attribute.

What carries the argument

The load-bearing object is the total causal effect $TE_{A_k\to Y}=E[Y|\mathrm{do}(A_k=0)]-E[Y|\mathrm{do}(A_k=1)]$, computed for each sensitive attribute $A_k$ against the ground-truth label $Y$. Around it, the paper builds the multi-attribute fairness module: demographic-only prompts per sensitive attribute, a local regularizer $L_{lf}=\sum_k\alpha_k\,\mathrm{KL}(Pr(A_k)\,\|\,U(1,|A_k|))$, and a global regularizer $L_{gf}=\sum_k\beta_k\Phi_k$ with $\Phi_k$ being demographic parity or equalized odds. The claimed link is the proportionality $|TE_{A_k\to Y}|\propto 1/|\Delta\Phi_k|$, so the causal estimate determines how much fairness improvement an attribute can deliver and why attributes compete.

What would settle it

Measure $TE_{A_k\to Y}$ on a labeled dataset, then apply the same debiasing effort to each sensitive attribute; if an attribute with smaller $|TE|$ still requires larger fairness improvement $\Delta\Phi_k$ to reach parity, the claimed inverse proportionality fails. Another concrete check: add an unobserved confounder to the causal graph and recompute $TE$; if the estimate changes substantially, the ranking of attributes is not stable and cannot guide fairness interventions.

Watch

Extended reading notes

Core claim

The paper's central claim is that sensitive attributes with strong causal effects on labels are more likely to confound group unfairness, and conversely that eliminating the causal pathway from an attribute to the label is what group-fairness metrics such as demographic parity ultimately demand. The authors propose EFF-DVP, an extension of FF-DVP, in which each sensitive attribute is converted into demographic-only text prompts that enter the vision-language model in parallel with images; a local fairness regularizer enforces uniform relevance across groups, and a global regularizer combines demographic parity and equalized odds across attributes. Causal discovery is used to draw a directed acyclic graph among attributes, inputs, and labels, and causal inference computes the total effect $TE_{A_k\to Y}$ via the do-operator. On the three tasks they test, attributes with larger estimated total effects (age on attractiveness, for example) are also the ones whose bias is the most difficult to reduce, and alleviating one attribute's bias can increase another's. From this they generalize that the larger the causal effect, the harder it is to ensure group fairness.

Load-bearing premise

The argument rests on the premise that the estimated total causal effects of sensitive attributes on the ground-truth label are true effects — requiring a correctly specified causal graph with no unobserved confounding — and that these label-level effects carry over to how hard the model's prediction bias is to remove.

Editorial extensions

If this is right

  • In a federated foundation model with multiple sensitive attributes, fairness interventions should be weighted by the causal effect of each attribute on the label; attributes with larger effects need more aggressive regularizers.
  • Fixing one sensitive attribute can inflate bias on another, so single-attribute fairness methods are insufficient when several protected attributes matter.
  • Causal refutation by adding random confounders can be used as a stability check before trusting a causal-effect-guided fairness decision.
  • Accuracy and group fairness trade off under multi-attribute debiasing, so a deployment must choose how much accuracy to spend for a target fairness level.

Reading between the lines

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

  • If the claimed proportionality holds beyond the three reported tasks, one could pre-screen a new dataset: estimate $TE_{A_k\to Y}$ on a labeled subset and use the ranking to allocate debiasing effort before training; the paper itself does not propose this adaptive budget strategy.
  • A direct validation would be to measure the rank correlation between $TE_{A_k\to Y}$ and the achieved $\Delta\Phi_k$ across many datasets and sensitive attributes; Tables I-II cover only two datasets and three tasks, so the general claim remains open.
  • The causal effect is measured on ground-truth labels while fairness is measured on predictions, so the argument assumes that label-level causality carries over to model-level bias; a model that decorrelates predictions from an attribute could break the link, which would require extending the theory.
  • In federated settings, causal structure may vary across clients; the paper reports different per-client $TE$ values, and a global estimate could conceal client-specific fairness failures, pointing toward client-aware causal fairness monitoring.
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 / 6 minor

Summary. The paper proposes EFF-DVP, an extension of the FF-DVP federated vision-language fine-tuning method that enforces group fairness for multiple sensitive attributes simultaneously. The local model uses demographic-only prompts for each sensitive attribute and combines per-attribute fairness regularizers with tunable weights, while causal discovery and inference are used to quantify total effects TE_k of each sensitive attribute on the ground-truth label. The central claim is that the larger the causal effect of a sensitive attribute on the label, the harder it is to debias that attribute in the federated foundation model, with supporting experiments on CelebA and FairFace.

Significance. If the central claim held, the paper would offer a practical principle: data-level causal estimates could rank sensitive attributes by debiasing difficulty and explain trade-offs in multi-attribute fairness. The paper is the first in this direction for federated foundation models, and the proposed multi-prompt architecture is a reasonable extension of FF-DVP. The evaluation uses real datasets and standard fairness metrics, which is a strength. However, the causal identification is unspecified and the key proportionality in Eq. (12) is not derived, so the paper currently provides an interesting empirical correlation rather than a validated causal theory. No code is shipped, so the causal estimates cannot be reproduced from the manuscript as written.

major comments (5)
  1. [IV-C, Eq. (12)] The central claim that |TE_{A_k→Y}| ∝ 1/|ΔΦ_k| is asserted without derivation. The text says this can be observed from Eqs. (5)–(11), but TE is defined on the ground-truth label Y under do(A_k), while ΔΦ_k is the change in a prediction-based fairness metric under a specific debiasing procedure. No mechanism connects P(Y|do(A_k)) to P(Ŷ|A_k), and the debiasing operation is not formally defined. Please provide a derivation with explicit assumptions, or restate the relation as an empirical observation rather than a theoretical proportionality.
  2. [IV-A/B, Table II] The causal estimates TE_k are not identifiable from the manuscript. Eq. (9) uses a do-operator, but the paper never specifies the causal graph used, the discovery algorithm actually executed (PC, GES, and LiNGAM are only mentioned as examples), the variables in V, the identification assumptions (e.g., no unmeasured confounding), the adjustment set, or the estimator. The 'refute' experiment only perturbs the estimate with one synthetic random common cause; high p-values show insensitivity to that perturbation, not that the estimate identifies a true causal effect. Please supply these details or label the reported quantities as associational.
  3. [V-D/V-E, Tables I and II] The experimental support consists of three task/dataset combinations with no standard deviations, confidence intervals, or multiple seeds. Several reported ΔΦ values are very small (e.g., 0.013 and 0.007 in Table I), and the ordering that supports the central claim could easily change with random variation. Please add repeated runs with variance reporting and a statistical comparison (e.g., paired tests) for the ΔΦ ordering.
  4. [IV-C] The sentence claiming that causal analysis in FL 'integrates heterogeneous data distributions from different clients via parameter aggregation' is unsupported. Table II reports per-client TE values and an average, but the paper never explains whether causal discovery is run locally per client, whether the resulting graphs are aggregated, or how aggregation affects identification. Please clarify the federated causal estimation protocol.
  5. [VI] The conclusion 'the larger the causal effect, the harder it is to ensure group fairness' is stronger than the evidence. The paper shows a qualitative pattern in three tasks, but with an unspecified causal estimator and no derivation of Eq. (12), this remains a hypothesis. Please temper the conclusion or supply the missing support.
minor comments (6)
  1. [IV-B, Eqs. (10)–(11)] The expression for NDE is nonstandard: E[Y|do(A_k=0), M=m] − E[Y|do(A_k=1), M=m] is a controlled direct effect, not a natural direct effect, which requires M set to its value under a reference intervention. Similarly, Eq. (11) does not match the standard NIE definition. If these quantities are not used in the paper, please remove or correct them.
  2. [IV-B] The decomposition 'TE = NDE + DIE' should read 'TE = NDE + NIE'; 'DIE' is undefined.
  3. [V-C, Table II] The p-value in the refute experiment is not defined: it is not stated what null hypothesis is tested, what test statistic is used, or how the synthetic confounder is generated.
  4. [V-A] The dataset description is terse: for CelebA the text says 'We choose gender and age as sensitive attributes to assist images for predicting labels attractive or smiling,' but the experiments report separate Attractive Detection and Smiling Detection tasks; please clarify whether two separate models are trained per task.
  5. [Table I] The table formatting is corrupted (e.g., the check marks and some cells are run together), and the text contains the typo 'tabel I' instead of 'Table I'. Please re-typeset.
  6. [Eq. (14)] The accuracy parity formula appears to have unmatched brackets and an unclear double summation; please rewrite it so the averaging over groups is unambiguous.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: TE and ΔΦ are measured from independent quantities; Eq. (12) is an asserted correlation, not an identity.

full rationale

The paper's central claim connects TE_{A_k→Y}, computed from ground-truth labels via the do-operator in Eq. (9), with ΔΦ_k, the change in model-level demographic parity or equalized odds from Eqs. (5)-(6) after debiasing. These are not the same quantity by construction: the causal effect is on Y, the fairness metrics are on Ŷ, and debiasing is performed by fairness regularizers (Eqs. (3)-(7)) that never use TE_{A_k→Y}. Thus the observed ordering in Tables I and II is an empirical association rather than a mathematically forced identity. Eq. (12) is asserted without derivation, but assertion is not circularity. The self-citations [3], [4], [7], and [8] are background references and are not load-bearing for the fairness-causality claim. The absence of a specified causal graph, adjustment set, or estimator for Eq. (9) is a causal identification gap that would affect correctness, but it does not make the derivation circular.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new entities. Its central load-bearing causal claims depend on unverified causal identification assumptions and unspecified weighting parameters.

free parameters (2)
  • alpha_k (per-attribute local fairness weights) = not specified
    Weights in Eq. (4) that trade off local-level fairness regularizers across sensitive attributes; no selection rule or values given, so results may depend on hand-tuning.
  • beta_k (per-attribute global fairness weights) = not specified
    Weights in Eq. (7) for global-level fairness regularizers; similar lack of specification.
assumptions (3)
  • domain assumption No unmeasured confounding and correct causal graph for the sensitive attributes and labels
    Section IV.B uses do-calculus formulas for TE/NDE/NIE that require the causal graph G to be correct and all confounders observed; the paper does not actually run causal discovery, so this is assumed.
  • standard math The do-operator and counterfactual consistency hold
    Standard causal inference assumptions invoked implicitly by the TE/NDE/NIE definitions in Section IV.B.
  • domain assumption Conditional independence tests can identify the causal structure from federated local data
    Section IV.A describes PC/GES/LiNGAM at a high level but does not implement them or discuss how data heterogeneity affects graph identification across clients.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Group Fairness with Multiple Sensitive Attributes in Federated Foundation Models." pith.science (2026). https://pith.science/paper/VUIXGROB

@misc{pith2026250618732,
  author       = {Pith},
  title        = {Pith review of: Towards Group Fairness with Multiple Sensitive Attributes in Federated Foundation Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VUIXGROB}},
  note         = {Machine review of arXiv:2506.18732}
}
read the original abstract

The deep integration of foundation models (FM) with federated learning (FL) enhances personalization and scalability for diverse downstream tasks, making it crucial in sensitive domains like healthcare. Achieving group fairness has become an increasingly prominent issue in the era of federated foundation models (FFMs), since biases in sensitive attributes might lead to inequitable treatment for under-represented demographic groups. Existing studies mostly focus on achieving fairness with respect to a single sensitive attribute. This renders them unable to provide clear interpretability of dependencies among multiple sensitive attributes which is required to achieve group fairness. Our paper takes the first attempt towards a causal analysis of the relationship between group fairness across various sensitive attributes in the FFM. We extend the FFM structure to trade off multiple sensitive attributes simultaneously and quantify the causal effect behind the group fairness through causal discovery and inference. Extensive experiments validate its effectiveness, offering insights into interpretability towards building trustworthy and fair FFM systems.

Figures

Figures reproduced from arXiv: 2506.18732 by the authors.

Figure 1
Figure 1. An example of the biased FL system. Gender: male (1) and female [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The local model overview of EFF-DVP. Orange modules with the fire logo can be updated and shared with the server, while blue modules with the snow logo are fixed. Multiple sensitive attributes are converted to demographic-only prompts for text encoding and processed in parallel to trade off fairness. In addition, the causal analysis is observed simultaneously to reveal the dependency of sensitive attributes with gro… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 25 canonical work pages

  1. [1]

    Communication-efficient learning of deep networks from decentralized data,

    Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas, “Communication-efficient learning of deep networks from decentralized data,” inArtificial intelligence and statistics. PMLR, 2017, pp. 1273–1282

  2. [2]

    Advances and open problems in federated learning,

    Peter Kairouz, H Brendan McMahan, Brendan Avent, Aur ´elien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al., “Advances and open problems in federated learning,”Foundations and Trends® in Machine Learning, vol. 14, no. 1–2, pp. 1–210, 2021

  3. [3]

    Fedmbp: Multi-branch prototype federated learning on heterogeneous data,

    Tianrun Gao, Xiaohong Liu, Yuning Yang, and Guangyu Wang, “Fedmbp: Multi-branch prototype federated learning on heterogeneous data,” in2023 IEEE International Conference on Image Processing (ICIP). IEEE, 2023, pp. 2180–2184

  4. [4]

    Dense contrastive-based federated learning for dense prediction tasks on medical images,

    Yuning Yang, Xiaohong Liu, Tianrun Gao, Xiaodong Xu, Ping Zhang, and Guangyu Wang, “Dense contrastive-based federated learning for dense prediction tasks on medical images,”IEEE Journal of Biomedical and Health Informatics, 2024

  5. [5]

    Rethinking architecture design for tackling data heterogeneity in federated learning,

    Liangqiong Qu, Yuyin Zhou, Paul Pu Liang, Yingda Xia, Feifei Wang, Ehsan Adeli, Li Fei-Fei, and Daniel Rubin, “Rethinking architecture design for tackling data heterogeneity in federated learning,” inPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 10061–10071

  6. [6]

    Buffalo: Biomedical vision-language understanding with cross-modal prototype and federated foundation model collaboration,

    Bingjie Yan, Qian Chen, Yiqiang Chen, Xinlong Jiang, Wuliang Huang, Bingyu Wang, Zhirui Wang, Chenlong Gao, and Teng Zhang, “Buffalo: Biomedical vision-language understanding with cross-modal prototype and federated foundation model collaboration,” inProceedings of the 33rd ACM International Conference on Information and Knowledge Management, 2024, pp. 2775–2785

  7. [7]

    Advances and open challenges in federated learning with foundation models,

    Chao Ren, Han Yu, Hongyi Peng, Xiaoli Tang, Anran Li, Yulan Gao, Alysa Ziying Tan, Bo Zhao, Xiaoxiao Li, Zengxiang Li, et al., “Advances and open challenges in federated learning with foundation models,”arXiv preprint arXiv:2404.15381, 2024

  8. [8]

    The prospect of enhancing large-scale heterogeneous federated learning with foundation models,

    Yulan Gao, Zhaoxiang Hou, Chengyi Yang, Zengxiang Li, Han Yu, and Xiaoxiao Li, “The prospect of enhancing large-scale heterogeneous federated learning with foundation models,” in2024 IEEE International Conference on Multimedia and Expo (ICME). IEEE, 2024, pp. 1–6

Show all 30 references
  1. [9]

    Trustworthy federated learning: A survey,

    Asadullah Tariq, Mohamed Adel Serhani, Farag Sallabi, Tariq Qayyum, Ezedin S Barka, and Khaled A Shuaib, “Trustworthy federated learning: A survey,”arXiv preprint arXiv:2305.11537, 2023

  2. [10]

    Towards fairness-aware feder- ated learning,

    Yuxin Shi, Han Yu, and Cyril Leung, “Towards fairness-aware feder- ated learning,”IEEE Transactions on Neural Networks and Learning Systems, 2023

  3. [11]

    Proportionally fair hospital collaborations in federated learning of histopathology images,

    S Maryam Hosseini, Milad Sikaroudi, Morteza Babaie, and Hamid R Tizhoosh, “Proportionally fair hospital collaborations in federated learning of histopathology images,”IEEE transactions on medical imaging, vol. 42, no. 7, pp. 1982–1995, 2023

  4. [12]

    Unified fair federated learning for digital healthcare,

    Fengda Zhang, Zitao Shuai, Kun Kuang, Fei Wu, Yueting Zhuang, and Jun Xiao, “Unified fair federated learning for digital healthcare,” Patterns, vol. 5, no. 1, 2024

  5. [13]

    Al- gorithmic fairness in artificial intelligence for medicine and healthcare,

    Richard J Chen, Judy J Wang, Drew FK Williamson, Tiffany Y Chen, Jana Lipkova, Ming Y Lu, Sharifa Sahai, and Faisal Mahmood, “Al- gorithmic fairness in artificial intelligence for medicine and healthcare,” Nature biomedical engineering, vol. 7, no. 6, pp. 719–742, 2023

  6. [14]

    Fair federated learning for heterogeneous data,

    Samhita Kanaparthy, Manisha Padala, Sankarshan Damle, and Sujit Gu- jar, “Fair federated learning for heterogeneous data,” inProceedings of the 5th Joint International Conference on Data Science & Management of Data (9th ACM IKDD CODS and 27th COMAD), 2022, pp. 298–299

  7. [15]

    Fair federated learning with biased vision-language models,

    Huimin Zeng, Zhenrui Yue, Yang Zhang, Lanyu Shang, and Dong Wang, “Fair federated learning with biased vision-language models,” inFindings of the Association for Computational Linguistics ACL 2024, 2024, pp. 10002–10017

  8. [16]

    Fairness- aware agnostic federated learning,

    Wei Du, Depeng Xu, Xintao Wu, and Hanghang Tong, “Fairness- aware agnostic federated learning,” inProceedings of the 2021 SIAM International Conference on Data Mining (SDM). SIAM, 2021, pp. 181– 189

  9. [17]

    Fairfed: Enabling group fairness in federated learning,

    Yahya H Ezzeldin, Shen Yan, Chaoyang He, Emilio Ferrara, and A Salman Avestimehr, “Fairfed: Enabling group fairness in federated learning,” inProceedings of the AAAI conference on artificial intelli- gence, 2023, vol. 37, pp. 7494–7502

  10. [18]

    Fairfl: A fair federated learning approach to reducing demographic bias in privacy-sensitive classification models,

    Daniel Yue Zhang, Ziyi Kou, and Dong Wang, “Fairfl: A fair federated learning approach to reducing demographic bias in privacy-sensitive classification models,” in2020 IEEE International Conference on Big Data (Big Data). IEEE, 2020, pp. 1051–1060

  11. [19]

    Fair-fate: Fair federated learning with momentum,

    Teresa Salazar, Miguel Fernandes, Helder Ara ´ujo, and Pedro Henriques Abreu, “Fair-fate: Fair federated learning with momentum,” inIn- ternational Conference on Computational Science. Springer, 2023, pp. 524–538

  12. [20]

    Glo- calfair: Jointly improving global and local group fairness in federated learning,

    Syed Irfan Ali Meerza, Luyang Liu, Jiaxin Zhang, and Jian Liu, “Glo- calfair: Jointly improving global and local group fairness in federated learning,”arXiv preprint arXiv:2401.03562, 2024

  13. [21]

    Bias mitigation in federated learning for edge computing,

    Yasmine Djebrouni, Nawel Benarba, Ousmane Touat, Pasquale De Rosa, Sara Bouchenak, Angela Bonifati, Pascal Felber, Vania Marangozova, and Valerio Schiavoni, “Bias mitigation in federated learning for edge computing,”Proceedings of the ACM on Interactive, Mobile, Wearable and U...

  14. [22]

    Fedcsl: A scalable and accurate approach to federated causal structure learning,

    Xianjie Guo, Kui Yu, Lin Liu, and Jiuyong Li, “Fedcsl: A scalable and accurate approach to federated causal structure learning,” inProceedings of the AAAI Conference on Artificial Intelligence, 2024, vol. 38, pp. 12235–12243

  15. [23]

    Causal representation learning via counterfactual intervention,

    Xiutian Li, Siqi Sun, and Rui Feng, “Causal representation learning via counterfactual intervention,” inProceedings of the AAAI Conference on Artificial Intelligence, 2024, vol. 38, pp. 3234–3242

  16. [24]

    When foundation model meets federated learning: Motivations, challenges, and future directions,

    Weiming Zhuang, Chen Chen, and Lingjuan Lyu, “When foundation model meets federated learning: Motivations, challenges, and future directions,”arXiv preprint arXiv:2306.15546, 2023

  17. [25]

    A survey on group fairness in federated learning: Challenges, taxonomy of solutions and directions for future research,

    Teresa Salazar, Helder Ara ´ujo, Alberto Cano, and Pedro Henriques Abreu, “A survey on group fairness in federated learning: Challenges, taxonomy of solutions and directions for future research,”arXiv preprint arXiv:2410.03855, 2024

  18. [26]

    Feature selection under fairness and performance constraints,

    Ginel Dorleon, Imen Megdiche, Nathalie Bricon-Souf, and Olivier Teste, “Feature selection under fairness and performance constraints,” inInternational Conference on Big Data Analytics and Knowledge Discovery. Springer, 2022, pp. 125–130

  19. [27]

    Improving fairness in ai models on electronic health records: The case for federated learning methods,

    Raphael Poulain, Mirza Farhan Bin Tarek, and Rahmatollah Beheshti, “Improving fairness in ai models on electronic health records: The case for federated learning methods,” inProceedings of the 2023 ACM conference on fairness, accountability, and transparency, 2023, pp. 1599–1608

  20. [28]

    Analyzing the impact of personalization on fairness in federated learning for healthcare,

    Tongnian Wang, Kai Zhang, Jiannan Cai, Yanmin Gong, Kim- Kwang Raymond Choo, and Yuanxiong Guo, “Analyzing the impact of personalization on fairness in federated learning for healthcare,”Journal of Healthcare Informatics Research, vol. 8, no. 2, pp. 181–205, 2024

  21. [29]

    Causal machine learning for predicting treatment outcomes,

    Stefan Feuerriegel, Dennis Frauen, Valentyn Melnychuk, Jonas Schweisthal, Konstantin Hess, Alicia Curth, Stefan Bauer, Niki Kilber- tus, Isaac S Kohane, and Mihaela van der Schaar, “Causal machine learning for predicting treatment outcomes,”Nature Medicine, vol. 30, no. 4, pp....

  22. [30]

    Causal discovery for fairness,

    R ¯uta Binkyt ˙e, Karima Makhlouf, Carlos Pinz ´on, Sami Zhioua, and Catuscia Palamidessi, “Causal discovery for fairness,” inWorkshop on Algorithmic Fairness through the Lens of Causality and Privacy. PMLR, 2023, pp. 7–22

Pith tools

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