Pith. sign in

REVIEW 3 major objections 5 minor 44 references

Towards Collaborative Fairness in Federated Learning Under Imbalanced Covariate Shift

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read In federated learning with imbalanced covariate shift, updating the global model only with each client's correctly predicted high-confidence samples improves collaborative fairness and accuracy, with a proven convergence guarantee.

desk verdict FedAKD is a simple, empirically strong method for collaborative fairness under imbalanced covariate shift, but its theoretical framing has sign and algebra errors that should not be taken at face value. read the letter →

arxiv 2507.08617 v1 pith:3XASDSS4 submitted 2025-07-11 cs.LG cs.AI

classification cs.LGcs.AI
keywords federatedlearningcollaborativefairnesscovariateshiftknowledgedistillationimbalanceddataasynchronouselectronichealthrecordsnon-IID
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

Federated learning assumes clients share similar data, but in practice clients differ in both how much data they hold and how their features are distributed. The paper names this combination imbalanced covariate shift and argues that, contrary to what prior fairness methods assume, the harmful part of the shift is carried by each client's misclassified samples: a PCA and kernel-density analysis on a pancreatic-cancer EHR dataset shows correctly predicted samples track the global feature distribution while wrongly predicted samples deviate strongly. On that basis it proposes FedAKD (Federated Asynchronous Knowledge Distillation), which first distills global knowledge into each client model, then selects each client's correctly predicted high-confidence samples and uses only those to refine the global model, with the server simply averaging the results. The paper proves convergence and reports that this simple filter improves collaborative fairness, maximum accuracy, and average accuracy over ten baselines on FashionMNIST, CIFAR10, and a 50-state EHR dataset, across several heterogeneity settings. The claim, read sympathetically, is that a correctness filter applied during distillation turns distributional imbalance into a fairness advantage without requiring contribution metrics.

What carries the argument

The carrying mechanism is the correctly predicted high-confidence sample set $I_k^t = \{(x,y) \in D_k : \mathrm{Pred}(w_k^t, x) = y\}$, coupled with a two-direction asynchronous distillation loop. In the global-to-local step the global model acts as teacher over the client's full data $D_k$, letting each client absorb global knowledge without abandoning local specialization; in the local-to-global step the roles reverse and the client model teaches the global model using only $I_k^t$, on the premise that these are the shift-free samples. The theoretical scaffolding is a KL-divergence analysis (Theorems 2.1 and 2.2) that expresses the imbalanced covariate shift as a function of the Fisher-information-weighted perturbation $\delta$ and the covariance $\Sigma$, which motivates removing misclassified samples; convergence (Theorem 3.2) is then proved by bounding the local gradient error through $\gamma$-inexact solutions and a subset-gradient dissimilarity assumption.

What would settle it

Construct a synthetic covariate shift that intentionally moves features in a direction the classifier already handles correctly, so that the shift is carried by correctly predicted samples rather than misclassified ones; then run FedAKD against a version that distills from all samples and compare global accuracy and fairness. If the filtered model is biased or worse, the premise that incorrect samples are the primary carrier of imbalanced covariate shift fails.

Watch

Extended reading notes

Core claim

The paper's central discovery is that under imbalanced covariate shift, the perturbation separating a client's feature distribution from the global one is concentrated in the samples that client's model classifies incorrectly. The paper supports this with a PCA/KDE comparison on a real pancreatic-cancer EHR dataset, where the density of correctly classified local samples closely matches the global density while misclassified samples do not. From this it constructs FedAKD, in which the global model is updated only through the set $I_k^t$ of each client's correctly predicted high-confidence samples, using an asynchronous two-stage distillation: global-to-local distillation on the full local data, then local-to-global distillation on $I_k^t$ only, then FedAvg aggregation. The paper presents a convergence theorem for this procedure under smoothness, strong convexity, and bounded-gradient-dissimilarity assumptions, and its experiments report higher collaborative-fairness coefficients, maximum accuracy, and average accuracy than ten baselines on three datasets under multiple non-IID partitions. Stated as the author would state it: the global model should learn from what clients get right, not from what they get wrong, because the wrong answers are where the covariate shift lives.

Load-bearing premise

The load-bearing premise is that the samples a client's local model predicts correctly form a covariate-shift-free subset of that client's data, so discarding all locally misclassified samples removes the harmful shift; the paper verifies this premise only through a PCA/KDE analysis of one EHR dataset and does not verify it on the image benchmarks.

Editorial extensions

If this is right

  • Fairness stops requiring explicit contribution metrics: high-quality clients naturally influence the global model more because they contribute more correct samples, while weaker clients still gain from the global-to-local distillation.
  • The method applies directly to hospital- or state-level health data federations, where sample sizes and feature distributions are both unbalanced, since the filter uses only local predictions and no validation set.
  • The convergence proof extends the standard FedAvg-style analysis (smoothness, strong convexity, bounded gradient dissimilarity) to asynchronous bidirectional distillation, providing a template for analysing similar teacher-student FL schemes.
  • The server-side update is unchanged FedAvg, so FedAKD can be combined with existing aggregation, privacy, or communication-saving techniques without protocol redesign.
  • Experiments under classical label-shift partitions (imbalanced classes and Dirichlet splits) show the same filter also improves fairness there, so the benefit is not restricted to feature shift.

Reading between the lines

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

  • The filter is self-referential: the set of 'correctly predicted' samples is defined by the very model being trained, so early rounds may select easy and uninformative examples; an implicit extension is to anneal the confidence threshold or to require prediction agreement across augmentations.
  • The paper's evidence that misclassified samples carry the shift comes from a single EHR analysis; a natural testable extension is to run the same PCA/KDE diagnostic on the image benchmarks or to engineer a shift that lives inside the correctly classified region and see whether the filter still helps.
  • The fairness coefficient used is the Pearson correlation between standalone and post-federation accuracies, which rewards monotone relationships; an extension would be to check worst-client accuracy or rank-based fairness, where the filter's effect could differ.
  • If the premise is correct, the same filtering idea should transfer to federated learning with noisy labels or to federated domain adaptation, where excluding locally uncertain samples could protect the global model from two different kinds of harm at once.
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 / 5 minor

Summary. The paper studies collaborative fairness in federated learning under imbalanced covariate shift, where clients differ in both sample size and feature distribution. It proposes FedAKD, an asynchronous knowledge-distillation method in which each client first learns from the global model on its full local data, then selects samples that its updated local model classifies correctly, and uses only those samples to refine the global model before standard FedAvg aggregation. The paper claims a theoretical characterization of imbalanced covariate shift via KL divergence (Theorems 2.1 and 2.2), a convergence guarantee for FedAKD (Theorem 3.2), and strong empirical gains in accuracy and collaborative fairness over ten baselines on FashionMNIST, CIFAR10, and a pancreatic-cancer EHR dataset with 50 state-level clients.

Significance. If the results hold, the contribution is valuable: FedAKD is a simple, practical client-update rule that appears to improve both accuracy and fairness across heterogeneous settings, with broad comparisons (ten baselines, four heterogeneity settings, three metrics), real-world EHR evaluation, and released source code. The central empirical claim is well supported by consistent improvements in the collaborative-fairness coefficient and average client accuracy. However, the theoretical scaffolding is not currently trustworthy: the proof of Theorem 2.1 contains a sign error in the size-dependent term, and the proof of Theorem 3.2 contains unstated spectral assumptions and an unsatisfiable-looking condition. The motivating premise for the filter in Eq. (4) is also verified only on the same EHR dataset used for evaluation, not on the image benchmarks. The paper is likely correct in its empirical claims, but the theoretical and motivational components need repair before the manuscript is publishable.

major comments (3)
  1. [Appendix A.1, Eqs. (7)–(14)] The derivation of Theorem 2.1 is internally inconsistent. The text obtains 1/2(-Δ^T I(θ*)Δ) ≈ -R/(2A) and then Eq. (12) concludes D_KL(phat||p_ω+δ) ≈ R/(2A), flipping the sign without any justifying step. This matters because a KL divergence cannot be negative, and the sign of the R/(2A) term is load-bearing for the claimed decomposition in Eq. (13) and for Theorem 2.2. The source of the error is the statement that E_{phat}[∇ℓ(θ*,X)] ≈ E_{pθ*}[∇ℓ(θ*,X)] = 0; retaining the first-order term gives E_{phat}[∇ℓ(θ*)] ≈ I(θ*)Δ, which contributes +Δ^T I(θ*)Δ and leads to the positive R/(2A) term. Please redo this expansion carefully and reconcile it with Eq. (13).
  2. [Appendix C, Eqs. (91)–(97) and Theorem 3.2] The convergence proof has not been established as written. In the transition from Eq. (95) to Eq. (97), the factor ||Ω_2^{-1}/β + Ω_1^{-1}/α|| is replaced by 4/β ||Ω_2|| + 4/α ||Ω_1||; this requires a lower bound on the smallest eigenvalue of Ω_1 and Ω_2 (e.g., ||Ω^{-1}|| ≤ 1/λ_min), which is never stated or justified. Ω1 and Ω2 are empirical second-moment matrices and can be singular. Furthermore, the condition r > 0 in Theorem 3.2 is given as a long implicit expression involving r1 and r2, but no argument shows that α, β, γ can satisfy it under Assumptions 1–4; as a result, the claimed linear convergence rate is not demonstrated. The proof also applies Assumption 4 to the dynamically changing subset I_k^t without arguing that a uniform θ exists over rounds.
  3. [Section 3.2, Eq. (4); Section 1, Figure 2] The core design premise behind Eq. (4) — that correctly predicted samples are approximately covariate-shift-free for every client, so discarding all locally misclassified samples removes the harmful shift — is supported only by the PCA/KDE analysis of one EHR dataset in Figure 2, obtained with a FedAvg-trained model. This premise is not verified on FashionMNIST or CIFAR10 under the ICS settings where the headline improvements are reported, and the figure does not establish that the retained subset is shift-free for each client, only that incorrect samples are more shifted than correct ones. The concern is not purely academic: the filter is model-dependent, so a client whose shifted distribution is systematically misclassified by the distilled local model may have its informative samples removed, biasing rather than de-shifting the global update. I ask for a per-client distribution analysis on the image benchmarks (e.g., KL divergence of correct vs. incorrect subsets, or a controlled experiment varying the filter) to ground this premise independently of the EHR dataset.
minor comments (5)
  1. [Appendix A.2, Eq. (28)] The constant term is written as M(M+4)/(4A) in Eq. (28), but the definition R = M + M(M+1)/2 gives C = M(M+3)/(4A), which is also the value stated in Theorem 2.2, Eq. (1). Please fix the arithmetic typo.
  2. [Section 3.2 and Algorithm 1] The step called 'high-confidence sample selection' actually selects samples with Pred(w_t^k, x) = y, i.e., correctly classified samples, with no confidence threshold. Either use a confidence-based criterion or rename the step to 'correct-sample selection' to avoid confusion.
  3. [Section 4.2 and Appendix E] Algorithm 2 samples data points with replacement according to Gaussian importance weights from a single global approximation; this may not preserve the class-conditional feature structure of the original labels. Please state this as a limitation or justify that the generated covariate shift still represents the intended setting.
  4. [Section 4.3 vs. Appendix D.3] Section 4.3 classifies FedAS and pFedCK as covariate-shift/personalized baselines, but Appendix D.3 states that personalized FL baselines are not included in the traditional non-IID experiments. Please reconcile these statements.
  5. [Appendix C, Eq. (35)] The notation L_KD is used for the distillation loss in Eq. (35) but is not defined; the loss is earlier denoted KD. Please use a single notation consistently.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: FedAKD's design is motivated, not derived, by the EHR observation, and the theory and convergence proofs are conditional and independent of the reported metrics.

full rationale

The paper's derivation chain is not circular. Theorems 2.1 and 2.2 provide standard asymptotic KL-divergence approximations in terms of the Fisher information and Gaussian perturbations; they do not mention correctly or incorrectly predicted samples, so the filtering rule is not derived from them. The filtering rule in Eq. (4) is an inductive design hypothesis motivated by the empirical observation in Figure 2, not a quantity fitted to the evaluation metric. The convergence proof of Theorem 3.2 is conditional on explicit assumptions, including Assumption 4 (bounded gradient dissimilarity on any subset), which is a hypothesis about the data and not the target conclusion; the theorem would be a valid conditional statement regardless of whether the filter is effective. The experimental claims are tested against ten baselines on FashionMNIST, CIFAR10, and the EHR data under the same protocol, so the accuracy and fairness gains are measured outcomes rather than consequences of the motivating observation. The synthetic ICS benchmarks are generated from the same Gaussian-perturbation model used in the theory, but this is a controlled simulation design and does not equate the evaluation metric to the theory's input. The only self-citation ([14], a GRU+attention architecture for EHR) is a standard architectural reference and is not load-bearing. The concern that the design motivation and the EHR evaluation share the same dataset is a generalization and selection issue, not a circular derivation, so it does not raise the circularity score.

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

The central claims rest on hand-set hyperparameters (alpha, beta, tau, eta), a chosen simulation shift magnitude C, and several unverified structural assumptions about distribution shape, convergence, and the fairness metric. None of these are fitted to the outcome metric, but several are load-bearing for the theoretical analysis.

free parameters (6)
  • Distillation coefficient alpha = 1.0
    Weight of global-to-local KD loss in Eq. (2); set by hand; the convergence theorem only requires existence of alpha satisfying r>0 without giving a concrete range.
  • Distillation coefficient beta = 1.0
    Weight of local-to-global KD loss in Eq. (5); set by hand.
  • Distillation temperature tau = 1.0
    KD temperature is fixed to 1 in experiments and in the convergence analysis.
  • Learning rate eta = 0.001 (FashionMNIST), 0.005 (CIFAR10), 1e-5 (EHR)
    Local SGD step size; chosen per dataset, not derived from theory.
  • Mahalanobis shift magnitude C in data generation = 5 (BCS); 2, 5, 10 (ICS)
    Controls simulated covariate shift strength in Algorithm 2; chosen by hand to create easy and hard regimes, not fitted to FedAKD's performance.
  • Inexactness parameters gamma_1, gamma_2 = not identified
    Abstract quantities in Definition 3.1; no concrete values or verification are given, and the theorem's r>0 condition depends on them.
assumptions (6)
  • standard math Asymptotic regularity conditions for a smooth parametric family and MLE asymptotics (Theorem 2.1)
    Required for the Fisher-information expansion of KL divergence; assumed, not verified on the data.
  • domain assumption Client data distributions are multivariate Gaussian with small perturbations (Theorem 2.2)
    Section 2 and Appendix B approximate latent features as Gaussian via a VAE; on the real EHR data this is an approximation, not a proven property.
  • ad hoc to paper The KL divergence between empirical client distribution and global distribution decomposes additively into p_{omega+delta} vs p_omega and bp vs p_{omega+delta} with negligible cross term e
    The cross term e in Eq. (13) is asserted to be approximately zero because bp is close to p_{omega+delta}; this is circular under the large-sample limit it is supposed to justify.
  • domain assumption Assumptions 1-4: L-smoothness, mu-strong convexity, bounded gradient dissimilarity, and bounded subset dissimilarity
    Appendix C; standard in FL theory, but mu-strong convexity does not hold for the CNN and GRU models used in the experiments, and the subset dissimilarity bound is unverified.
  • ad hoc to paper Correctly predicted samples have similar feature distributions across clients, while incorrectly predicted samples account for covariate shift (Figure 2)
    Section 3.2; derived from one EHR dataset and assumed to generalize to other federated settings; this is the basis for Algorithm 1 Step 2.
  • domain assumption Collaborative fairness is correctly measured by CF = 100 * Pearson(standalone accuracy, post-federation accuracy)
    Metric adopted from prior CF literature; it is not a formal fairness notion and trivially equals 100 if no collaboration occurs, which is not reported as a baseline.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Collaborative Fairness in Federated Learning Under Imbalanced Covariate Shift." pith.science (2026). https://pith.science/paper/3XASDSS4

@misc{pith2026250708617,
  author       = {Pith},
  title        = {Pith review of: Towards Collaborative Fairness in Federated Learning Under Imbalanced Covariate Shift},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3XASDSS4}},
  note         = {Machine review of arXiv:2507.08617}
}
read the original abstract

Collaborative fairness is a crucial challenge in federated learning. However, existing approaches often overlook a practical yet complex form of heterogeneity: imbalanced covariate shift. We provide a theoretical analysis of this setting, which motivates the design of FedAKD (Federated Asynchronous Knowledge Distillation)- simple yet effective approach that balances accurate prediction with collaborative fairness. FedAKD consists of client and server updates. In the client update, we introduce a novel asynchronous knowledge distillation strategy based on our preliminary analysis, which reveals that while correctly predicted samples exhibit similar feature distributions across clients, incorrectly predicted samples show significant variability. This suggests that imbalanced covariate shift primarily arises from misclassified samples. Leveraging this insight, our approach first applies traditional knowledge distillation to update client models while keeping the global model fixed. Next, we select correctly predicted high-confidence samples and update the global model using these samples while keeping client models fixed. The server update simply aggregates all client models. We further provide a theoretical proof of FedAKD's convergence. Experimental results on public datasets (FashionMNIST and CIFAR10) and a real-world Electronic Health Records (EHR) dataset demonstrate that FedAKD significantly improves collaborative fairness, enhances predictive accuracy, and fosters client participation even under highly heterogeneous data distributions.

Figures

Figures reproduced from arXiv: 2507.08617 by the authors.

Figure 1
Figure 1. (a) KL divergence vs. sample size for each client’s local data, revealing both data imbalance and feature covariate [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. (a) Distribution of locally versus globally correct samples. (b) Distribution of locally versus globally incorrect samples. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Efficiency comparison of FedAKD and baselines. 5.2 Performance Analysis [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Comparison of KL divergences on real data (red), [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 36 canonical work pages

  1. [1]

    Liang Gao, Hongchao Fu, Lili Li, Yanyan Chen, Min Xu, and Cheng-Zhong Xu

  2. [2]

    Ozgu Goksu and Nicolas Pugeault. 2024. Robust Federated Learning in the Face of Covariate Shift: A Magnitude Pruning with Hybrid Regularization Framework for Enhanced Model Aggregation. arXiv preprint arXiv:2412.15010 (2024). https: //arxiv.org/abs/2412.15010

  3. [3]

    Wei Guo, Wei Ge, Longbo Cui, Hua Li, and Li Kong. 2019. An interpretable disease onset predictive model using crossover attention mechanism from electronic health records. IEEE Access 7 (2019), 134236–134244

  4. [4]

    Geoffrey Hinton. 2015. Distilling the Knowledge in a Neural Network. arXiv preprint arXiv:1503.02531 (2015)

  5. [5]

    Stich, and Ananda Theertha Suresh

    Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sanjiv Reddi, Sebastian U. Stich, and Ananda Theertha Suresh. 2020. Scaffold: Stochastic Controlled Aver- aging for Federated Learning. In Proceedings of the 37th International Conference on Machine Learning (ICML) . PMLR, 5132–5143

  6. [6]

    Jakub Konecný. 2016. Federated Learning: Strategies for Improving Communica- tion Efficiency. arXiv preprint arXiv:1610.05492 (2016). arXiv:1610.05492 [cs.LG]

  7. [7]

    Alex Krizhevsky and Geoffrey Hinton. 2009. Learning Multiple Layers of Features from Tiny Images. Technical Report. University of Toronto. Technical Report

  8. [8]

    Wonbin Kweon, SeongKu Kang, and Hwanjo Yu. 2021. Bidirectional distillation for top-K recommender system. In Proceedings of the Web Conference 2021 . 3861– 3871

Show all 44 references
  1. [9]

    Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. 2020. Federated learning: Challenges, methods, and future directions. IEEE Signal Processing Magazine 37, 3 (2020), 50–60

  2. [10]

    Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith

    Tian Li, Anit K. Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. 2020. Federated optimization in heterogeneous networks. In Proceedings of Machine Learning and Systems , Vol. 2. 429–450

  3. [11]

    Xiang Li, Kaixuan Huang, Wenhao Yang, Shusen Wang, and Zhihua Zhang. 2019. On the Convergence of FedAvg on Non-IID Data. arXiv:1907.02189 [cs.LG] arXiv preprint arXiv:1907.02189

  4. [12]

    Tsung-Yi Lin, Priyal Goyal, Ross Girshick, Kaiming He, Piotr Dollár, and Serge Belongie. 2017. Focal Loss for Dense Object Detection. arXiv preprint arXiv:1708.02002 (2017)

  5. [13]

    Lingjuan Lyu, Xinyang Xu, Qiang Wang, Han Yu, et al . 2020. Collaborative Fairness in Federated Learning. In Federated Learning: Privacy and Incentive . 189–204

  6. [14]

    Fenglong Ma, Radha Chitta, Jing Zhou, Quanzeng You, Tong Sun, and Jing Gao

  7. [15]

    Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y Arcas. 2017. Communication-Efficient Learning of Deep Networks from Decentralized Data. In Proceedings of the 20th International Conference on Artificial Intelligence and Statistics (AISTATS) (Proce...

  8. [16]

    Xuanming Ni, Xinyuan Shen, and Huimin Zhao. 2022. Federated optimization via knowledge codistillation. Expert Systems with Applications 191 (2022), 116310. https://doi.org/10.1016/j.eswa.2021.116310

  9. [17]

    Ertong Shang, Hui Liu, Zhuo Yang, Junzhao Du, and Yiming Ge. 2023. FedBiKD: Federated Bidirectional Knowledge Distillation for Distracted Driving Detection. IEEE Internet of Things Journal (2023)

  10. [18]

    Qingxiong Tan, Min Ye, Bin Yang, S. Liu, A. J. Ma, T. C. F. Yip, Y. Zhao, S. C. Hui, T. M. F. Chan, F. K. Chan, J. J. Y. Sung, E. C. Cheung, and P. Yuen. 2020. Data-GRU: Dual-Attention Time-Aware Gated Recurrent Unit for Irregular Multivariate Time Series. In Proceedings of th...

  11. [19]

    Zihui Wang, Zhe Peng, Xinyu Fan, Zheng Wang, Siyang Wu, Rui Yu, ..., and Chunyan Wang. 2024. FedAVE: Adaptive data value evaluation framework for collaborative fairness in federated learning. Neurocomputing 574 (2024), 127227

  12. [20]

    Zihui Wang, Zheng Wang, Lingjuan Lyu, Zhigang Peng, Zhiquan Yang, Chuan Wen, and Xiaohui Fan. 2024. FedSAC: Dynamic Submodel Allocation for Collab- orative Fairness in Federated Learning. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining ....

  13. [21]

    Wickramaratne and Md Shaad Mahmud

    Sajila D. Wickramaratne and Md Shaad Mahmud. 2020. Bi-directional gated recurrent unit based ensemble model for the early detection of sepsis. In 2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC). IEEE, 70–73

  14. [22]

    Han Xiao, Kashif Rasul, and Roland Vollgraf. 2017. Fashion-MNIST: A Novel Image Dataset for Benchmarking Machine Learning Algorithms. arXiv preprint arXiv:1708.07747 (2017)

  15. [23]

    Xinyi Xu, Lingjuan Lyu, Xiaofeng Ma, Chunyan Miao, Chee Seng Foo, and Bo An Kiat Huat Low. 2021. Gradient driven rewards to guarantee fairness in collaborative machine learning. In Advances in Neural Information Processing Systems, Vol. 34. 16104–16117

  16. [24]

    Gang Yan, Haiyan Wang, Xue Yuan, and Jia Li. 2023. Criticalfl: A critical learning periods augmented client selection framework for efficient federated learning. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. ACM, 2898–2907

  17. [25]

    Xiyuan Yang, Wenke Huang, and Mang Ye. 2024. FedAS: Bridging Inconsistency in Personalized Federated Learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . IEEE, 11986–11995. https: //doi.org/10.1109/CVPR52733.2024.01139

  18. [26]

    Yang Yang, Xiangwei Zheng, and Cun Ji. 2019. Disease prediction model based on bilstm and attention mechanism. In 2019 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) . IEEE, 1141–1148

  19. [27]

    Xiangyang Ye, Q. T. Zeng, Julio C. Facelli, Diana I. Brixner, Mike Conway, and Bradley E. Bray. 2020. Predicting optimal hypertension treatment pathways using recurrent neural networks. International Journal of Medical Informatics 139 (2020), 104122

  20. [28]

    Mikhail Yurochkin, Mayank Agarwal, Soumya Ghosh, Kristjan Greenewald, Natesh Hoang, and Yasaman Khazaeni. 2019. Bayesian Nonparametric Federated Learning of Neural Networks. In International Conference on Machine Learning (Proceedings of Machine Learning Research) . PMLR, 7252–7261

  21. [29]

    Jinghe Zhang, Kamran Kowsari, James H Harrison, Jason M Lobo, and Laura E Barnes. 2018. Patient2vec: A personalized interpretable deep representation of the longitudinal electronic health record. IEEE Access 6 (2018), 65333–65346

  22. [30]

    Jianfei Zhang and Yongqiang Shi. 2024. A Personalized Federated Learning Method Based on Clustering and Knowledge Distillation. Electronics 13, 5 (2024),

  23. [31]

    Yue Zhao, Meng Li, Liangzhen Lai, Naveen Suda, Dave Civin, and Vikas Chandra

  24. [36]

    THETA_U (grad upload)

  25. [37]

    CLIP_NORM (clip thr.)

  26. [38]

    update) (1) 0.5 (2) 5.0 (3) 0.05 (4) 1.0 CGSV

    ALPHA (rep. update) (1) 0.5 (2) 5.0 (3) 0.05 (4) 1.0 CGSV

  27. [39]

    ALTRUISM (1) 0.9 (2) 2.0 (3) True (4) 1.0 FedA VE

  28. [40]

    ALPHA / BETA (1) 0.5 (2) 0.3 (3) 0.9 / 1.0 FedAvg — — FedDC 1) ALPHA (penalty)

  29. [41]

    coeff.) 0.01 FedSAC 1) BETA (c_i mapping)

    drift_vars (1) 1.0 (2) init=0 FedMPR PRUNE_PERCENT 0.1 FedProx MU (prox. coeff.) 0.01 FedSAC 1) BETA (c_i mapping)

  30. [42]

    mid_round (1) 2.0 (2) 15 SCAFFOLD 1)𝜂𝑔 (global LR) 2)𝜂𝑙 (local LR)

  31. [43]

    c_global, c_local (1) 0.005 (2) 0.1 (3) 1 (4) init=0 FedAKD

  32. [44]

    Temp𝑇 (1) 1.0 (2) 1.0 (3) 1.0 Table 7: State Data Statistics State Total Pos. Neg. State Total Pos. Neg. AK 558 196 362 MT 636 221 415 AL 3,410 1,292 2,118 NC 7,263 2,222 5,041 AR 2,341 842 1,499 ND 605 179 426 AZ 5,521 2,347 3,174 NE 1,429 424 1,005 CA 20,040 7,116 12,924 NH ...

  33. [857]

    https://doi.org/10.3390/electronics13050857

  34. [2017]

    In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining

    Dipole: Diagnosis prediction in healthcare via attention-based bidirectional recurrent neural networks. In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining . 1903–1911

  35. [2018]

    Real KL” means the empirical KL divergence between each client’s approxi- mated distribution and the global fitted distribution. “Approx

    Federated learning with non-iid data. arXiv preprint arXiv:1806.00582 (2018). A IMBALANCED COV ARIATE SHIFT PROOF A.1 Proof of Theorem 2.1 We first aim to expand 𝐷KL 𝑝𝜔+𝛿 𝑝𝜔 := E𝑋∼𝑝𝜔+𝛿 h log𝑝𝜔+𝛿(𝑋) −log𝑝𝜔(𝑋) i . Observe that log𝑝𝜔(𝑥) = log𝑝(𝜔+𝛿)−𝛿(𝑥) = ℓ(𝜔+𝛿)− 𝛿, 𝑥 . We make a...

  36. [2022]

    In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    FedDC: Federated Learning with Non-iid Data via Local Drift Decoupling and Correction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 10112–10121. Towards Collaborative Fairness in Federated Learning Under Imbalanced Covariate Shift KDD ’...

Pith tools

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