pith. sign in

arxiv: 2607.01474 · v1 · pith:EMBLKS2Enew · submitted 2026-07-01 · 💻 cs.LG

Class-Grouped Normalized Momentum and Faster Hyperparameter Exploration to Tackle Class Imbalance in Federated Learning

Haemin Park , Diego Klabjan , Martin W. Braun , Xiuqi Li , Balakrishnan Ananthanarayanan This is my paper

Pith reviewed 2026-07-03 21:03 UTC · model grok-4.3

classification 💻 cs.LG
keywords federated learningclass imbalancemomentum optimizationclass groupingnormalized momentumresampling rateslong-tailed dataconvergence analysis
0
0 comments X

The pith

Class grouping by variance plus per-group momentum normalization equalizes gradients across imbalanced classes in federated learning.

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

Class imbalance in federated learning leaves rare classes with weak performance because data cannot be centralized for rebalancing. The paper shows that partitioning classes into a few groups chosen for minimum within-group variance, then keeping separate momentum for each group, normalizing every group momentum to unit length, and adding the normalized vectors produces the update direction. This step equalizes the effective size of gradients from common and rare classes while damping the high variance that comes with scarce minority samples. A convergence guarantee is given that holds when resampling rates vary over time. FedHOO, an X-armed-bandit procedure, lets clients test pairs of candidate rates in parallel at linear cost, yielding further gains on long-tailed image sets and a chip-defect collection.

Core claim

FedCGNM partitions classes into a small number of groups based on minimum within-group variance, maintains a momentum per group, normalizes each group momentum to unit length, and uses the summation of the normalized group momentums as an update direction. This design both equalizes gradient magnitude across majority and minority groups and mitigates the noise inherent in rare-class gradients. The approach supplies a convergence analysis that accounts for time-varying resampling rates and introduces FedHOO to search those rates efficiently via federated parallelism.

What carries the argument

Class-grouped normalized momentum: classes are partitioned by minimum within-group variance, momentum is tracked and normalized to unit length per group, and the normalized vectors are summed to form the update direction.

If this is right

  • Gradient magnitudes become comparable between majority and minority groups.
  • Noise from rare-class gradients is reduced by the per-group normalization.
  • Convergence holds when resampling rates change over training.
  • FedHOO evaluates pairs of resampling rates at linear cost by exploiting client parallelism.
  • The method improves accuracy on four public long-tailed datasets and one proprietary defect-detection set.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same grouping-plus-normalization step could be tested in centralized training on imbalanced data to see whether the privacy constraint is necessary for the benefit.
  • Dynamic adjustment of the number of groups during training might handle shifting class distributions without extra tuning.
  • The bandit search for rates could be combined with other client optimizers to explore additional hyperparameters at low extra cost.
  • Extreme heterogeneity across clients might require the variance-based grouping to be recomputed periodically rather than once.

Load-bearing premise

That partitioning classes by minimum within-group variance and then normalizing per-group momentums will reliably equalize effective gradient magnitudes and reduce noise for minority classes under the heterogeneity and privacy constraints of federated learning.

What would settle it

Running FedCGNM on a standard long-tailed federated benchmark and finding that minority-class accuracy stays no better than with ordinary momentum would show the equalization step does not work as claimed.

Figures

Figures reproduced from arXiv: 2607.01474 by Balakrishnan Ananthanarayanan, Diego Klabjan, Haemin Park, Martin W. Braun, Xiuqi Li.

Figure 1
Figure 1. Figure 1: Test accuracy on CIFAR-100-LT (ξ = 20, K = 5) with respect to the number of classes assigned to the minority group. The red line marks the threshold selected by our grouping rule. Taking expectation, we obtain E∥∆h∥ 2 =  S 2 h+ Sh(1−Sh) B  ∥wˆ h−uh∥ 2+ Sh B  1− X c∈Gh w 2 c|h  . The optimization problem for grouping is minGh E∥∆h∥ 2 . We solve this problem heuristically by finding the best threshold th… view at source ↗
Figure 2
Figure 2. Figure 2: shows that the standard FedAvg baseline stalls at around 73, while other FL baselines provide only marginal improvements. FedCGN pushes performance into 85, and FedCGNM adds another improvement. Using FedHOO as resampling strategy consistently improves training. This demonstrates that variance-aware grouping, per-group mo￾mentum, and efficient rate exploration are not only effective on public benchmarks bu… view at source ↗
Figure 3
Figure 3. Figure 3: Training and validation loss on CIFAR-100-LT (ξ = 20, K = 20) for FedCGNM with different group counts H [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Distribution of sample counts by factory and defect code. Each row represents one factory; the left panel displays defect counts and the right panel shows non-defect counts, both on a logarithmic y-axis. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Instability of PCN in a multi-class setting in centralized learning setting. PCN shows large loss spikes and intermittent drops in validation accuracy, supporting that per-class normalization can yield unstable optimization and poor generalization when class-wise directions are noisy or poorly estimated. D.2. Validation of the Variance-Based Grouping Threshold (a) CIFAR-10 - Imbalance rate=20 (b) CIFAR-10 … view at source ↗
Figure 6
Figure 6. Figure 6: Test accuracy versus the number of rarest classes assigned to the minority group on CIFAR-10-LT. The dashed red line at t = 6 marks the threshold chosen by our variance-based grouping rule, which aligns with the peak test accuracy in both imbalance scenarios. To confirm that our grouping rule reliably identifies the optimal split, we sweep the threshold t ∈ {1, . . . , 9}, i.e. the number of rarest classes… view at source ↗
Figure 7
Figure 7. Figure 7: Training and validation loss for different numbers of groups H ∈ {2, 3, 4}. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Per-client resampling-rate trajectories selected by FedHOO across datasets and imbalance settings. E.2. Resampling-rates in FedHOO To better understand and illustrate the benefit of FedHOO, we plot and compare the trajectories of sampling-rate selection under FedHOO and standard HOO when each search method is run throughout training. In FedHOO, the selected sampling￾rates correspond to the candidates yield… view at source ↗
Figure 9
Figure 9. Figure 9: Per-client resampling-rate trajectories selected by standard HOO on CIFAR-10 iid setting with imbalance rate ξ = 20. Under standard HOO, the algorithm explores only one branch of the search tree per round. As a result, the sampling-rate assigned to each client fluctuates heavily throughout training, and fail to converge in reasonable training time. The per-client trajectories remain noisy even after hundre… view at source ↗
Figure 10
Figure 10. Figure 10: Alignment comparison over rounds. The top panel plots similarity values for FedCGNM-induced updates versus stochastic￾gradient updates, and the bottom panel shows their difference (FedCGNM minus stochastic gradient). Positive differences dominate, indicating improved alignment under FedCGNM. To empirically assess this effect, we track an alignment metric over communication rounds and compare it against a … view at source ↗
Figure 11
Figure 11. Figure 11: Distribution of per-class test accuracies on CIFAR-100-LT (imbalance rate=100, K=5). From [PITH_FULL_IMAGE:figures/full_fig_p025_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: shows the results on AdultIncome and CIFAR-100-LT. On AdultIncome, the cosine similarity is often high, but the magnitude gap remains consistently positive and non-negligible. Thus, the magnitude-gap branch of the assumption is supported. On CIFAR-100-LT, the cosine similarity remains far below perfect alignment throughout training, and the magnitude gap is also positive. Thus, the low-alignment branch is… view at source ↗
read the original abstract

Class imbalance poses a critical challenge in federated learning (FL), where underrepresented classes suffer from poor predictive performance yet cannot be addressed by standard centralized techniques due to privacy and heterogeneity constraints. We propose FedCGNM (Federated Class-Grouped Normalized Momentum), a client-side optimizer in FL that partitions classes into a small number of groups based on minimum within-group variance, maintains a momentum per group, normalizes each group momentum to unit length, and uses the summation of the normalized group momentums as an update direction. This design both equalizes gradient magnitude across majority and minority groups and mitigates the noise inherent in rare-class gradients. We further provide a theoretical convergence analysis explicitly accounting for time-varying resampling-rates. Additionally, to efficiently optimize these rates in small-client regimes, we introduce FedHOO, an X-armed-bandit (XAB) based algorithm that exploits federated parallelism that evaluates many combinations of two candidate rates per client at linear cost. Empirical evaluation on four public long-tailed benchmarks and a proprietary chip-defect dataset demonstrates that FedCGNM consistently outperforms baselines, with FedHOO yielding further gains in small-scale federations.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces FedCGNM, a client-side optimizer for federated learning under class imbalance. It partitions classes into a small number of groups by minimizing within-group variance, maintains a momentum vector per group, normalizes each to unit length, and sums the normalized vectors as the update direction. This is claimed to equalize effective gradient magnitudes across majority and minority classes while mitigating noise from rare classes. The manuscript provides a convergence analysis that explicitly incorporates time-varying resampling rates, proposes FedHOO (an X-armed-bandit algorithm that evaluates candidate resampling-rate pairs at linear cost by exploiting federated parallelism), and reports empirical gains on four public long-tailed benchmarks plus a proprietary chip-defect dataset.

Significance. If the central claims hold, the work supplies a practical client-local mechanism for class-imbalance mitigation that respects FL privacy and heterogeneity constraints, together with an efficient hyper-parameter search procedure that scales with the number of clients. The explicit accounting for time-varying resampling rates in the convergence statement and the linear-cost parallel evaluation in FedHOO are concrete strengths that distinguish the contribution from purely heuristic re-weighting approaches.

major comments (2)
  1. [§4, Theorem 2] §4 (Convergence Analysis), Theorem 2 and the surrounding discussion of the normalized update: the stated bound accounts for the time-varying resampling rates but supplies no explicit control on the directional error that arises when a high-variance momentum estimate (from a minority-class group) is normalized to unit length and injected at full magnitude into the summed direction. Under client heterogeneity this term is not averaged away, so the noise-mitigation claim central to FedCGNM is not yet supported by the analysis.
  2. [§3.2] §3.2 (Group Construction): the minimum-within-group-variance partitioning is performed once on the global class statistics; the manuscript does not analyze how sensitive the subsequent normalized-momentum update is to errors in this partitioning when only local client data are available, which directly affects whether the magnitude-equalization property holds in practice.
minor comments (2)
  1. [Eq. (3)] Notation for the per-group momentum vectors is introduced without an explicit index over groups in Eq. (3); adding a group index would improve readability.
  2. [Table 2] Table 2 caption does not state whether the reported accuracies are averaged over the same number of random seeds as the main tables; consistency of experimental protocol should be clarified.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. The comments identify important aspects of the convergence analysis and group construction that merit clarification and strengthening. We address each point below and indicate the revisions we will make.

read point-by-point responses

  1. Referee: [§4, Theorem 2] §4 (Convergence Analysis), Theorem 2 and the surrounding discussion of the normalized update: the stated bound accounts for the time-varying resampling rates but supplies no explicit control on the directional error that arises when a high-variance momentum estimate (from a minority-class group) is normalized to unit length and injected at full magnitude into the summed direction. Under client heterogeneity this term is not averaged away, so the noise-mitigation claim central to FedCGNM is not yet supported by the analysis.

    Authors: We appreciate the referee's careful reading of the analysis. Theorem 2 already incorporates the time-varying resampling rates into the convergence rate. The unit-length normalization is intended to equalize magnitudes, and the momentum update itself damps variance over iterations. However, we agree that an explicit bound on the directional deviation induced by noisy minority-group normalization (which does not average out under heterogeneity) is not derived. In the revision we will extend the proof of Theorem 2 to include a term that controls the expected cosine deviation of each normalized group momentum, using standard concentration arguments on the momentum estimator. This will make the noise-mitigation property explicit in the bound. revision: yes

  2. Referee: [§3.2] §3.2 (Group Construction): the minimum-within-group-variance partitioning is performed once on the global class statistics; the manuscript does not analyze how sensitive the subsequent normalized-momentum update is to errors in this partitioning when only local client data are available, which directly affects whether the magnitude-equalization property holds in practice.

    Authors: The partitioning step uses global class-frequency statistics, which can be obtained in a privacy-preserving manner via secure aggregation of per-client class counts. We acknowledge that the manuscript provides no formal sensitivity analysis of how local estimation errors or client-specific class distributions affect the resulting groups and therefore the equalization property. In the revised version we will add a short theoretical sensitivity result in §3.2 (showing that small perturbations in the frequency vector induce bounded changes in the normalized direction) together with an empirical study that perturbs the global statistics and measures the resulting performance degradation on the long-tailed benchmarks. revision: yes

Circularity Check

0 steps flagged

No circularity detected; derivation self-contained

full rationale

The provided abstract and description introduce FedCGNM (class partitioning by minimum within-group variance, per-group momentum, unit normalization, summation as update) and FedHOO (X-armed bandit for rate optimization) as novel client-side methods, with a convergence analysis that explicitly accounts for time-varying resampling rates. No equations, fitted parameters presented as predictions, self-citations, or uniqueness theorems are quoted that would reduce any central claim to its inputs by construction. The design choices are motivated directly from the problem of class imbalance under FL constraints and validated empirically on benchmarks, making the derivation independent rather than tautological. This matches the default expectation of no significant circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented entities; all such elements would require the full manuscript.

pith-pipeline@v0.9.1-grok · 5749 in / 1094 out tokens · 22252 ms · 2026-07-03T21:03:21.775991+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

69 extracted references · 69 canonical work pages · 1 internal anchor

  1. [1]

    Artificial Intelligence Review , volume=

    A survey on imbalanced learning: latest research, applications and future directions , author=. Artificial Intelligence Review , volume=. 2024 , publisher=

    work page 2024

  2. [2]

    Journal of Big Data , volume=

    Survey on deep learning with class imbalance , author=. Journal of Big Data , volume=. 2019 , publisher=

    work page 2019

  3. [3]

    Neural Processing Letters , volume=

    Learning from imbalanced data sets with weighted cross-entropy function , author=. Neural Processing Letters , volume=. 2019 , publisher=

    work page 2019

  4. [4]

    Proceedings of the IEEE International Conference on Computer Vision , pages=

    Focal loss for dense object detection , author=. Proceedings of the IEEE International Conference on Computer Vision , pages=

    work page

  5. [5]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Class-balanced loss based on effective number of samples , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    work page

  6. [6]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Addressing class imbalance in federated learning , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    work page

  7. [7]

    International Conference on Machine Learning , pages=

    A theoretical analysis of the learning dynamics under class imbalance , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    work page 2023

  8. [8]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Gradient reweighting: Towards imbalanced class-incremental learning , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    work page

  9. [9]

    International Conference on Database Systems for Advanced Applications , pages=

    FedGR: Federated learning with gravitation regulation for double imbalance distribution , author=. International Conference on Database Systems for Advanced Applications , pages=. 2023 , organization=

    work page 2023

  10. [11]

    Frontiers in Digital Health , volume=

    A review on over-sampling techniques in classification of multi-class imbalanced datasets: Insights for medical problems , author=. Frontiers in Digital Health , volume=. 2024 , publisher=

    work page 2024

  11. [12]

    Journal of Artificial Intelligence Research , volume=

    SMOTE: synthetic minority over-sampling technique , author=. Journal of Artificial Intelligence Research , volume=

    work page

  12. [13]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Generative adversarial minority oversampling , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    work page

  13. [14]

    Engineering Applications of Artificial Intelligence , volume=

    An improved generative adversarial network to oversample imbalanced datasets , author=. Engineering Applications of Artificial Intelligence , volume=. 2024 , publisher=

    work page 2024

  14. [15]

    IEEE Transactions on Parallel and Distributed Systems , volume=

    Fast: enhancing federated learning through adaptive data sampling and local training , author=. IEEE Transactions on Parallel and Distributed Systems , volume=. 2023 , publisher=

    work page 2023

  15. [16]

    Journal of Big Data , volume=

    Resampling approaches to handle class imbalance: a review from a data perspective , author=. Journal of Big Data , volume=. 2025 , publisher=

    work page 2025

  16. [17]

    Neural Networks , volume=

    A systematic study of the class imbalance problem in convolutional neural networks , author=. Neural Networks , volume=. 2018 , publisher=

    work page 2018

  17. [18]

    International Journal of Data Science and Analytics , pages=

    Leveraging local data sampling strategies to improve federated learning , author=. International Journal of Data Science and Analytics , pages=. 2024 , publisher=

    work page 2024

  18. [19]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Dynamic curriculum learning for imbalanced data classification , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    work page

  19. [20]

    , author=

    Tackling class-imbalanced learning issues based on local neighborhood information and generative adversarial networks. , author=. Sensors & Materials , volume=

    work page

  20. [21]

    World Journal of Methodology , volume=

    Challenges and limitations of synthetic minority oversampling techniques in machine learning , author=. World Journal of Methodology , volume=

    work page

  21. [22]

    , author=

    X-Armed Bandits. , author=. Journal of Machine Learning Research , volume=

    work page

  22. [23]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Large-scale long-tailed recognition in an open world , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    work page

  23. [24]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Implicit gradient alignment in distributed and federated learning , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    work page

  24. [25]

    UCI Machine Learning Repository , volume=

    Adult , author=. UCI Machine Learning Repository , volume=

    work page

  25. [26]

    2015 Military Communications and Information Systems Conference (MilCIS) , pages=

    UNSW-NB15: a comprehensive data set for network intrusion detection systems (UNSW-NB15 network data set) , author=. 2015 Military Communications and Information Systems Conference (MilCIS) , pages=. 2015 , organization=

    work page 2015

  26. [27]

    2024 21st International Conference on Electrical Engineering, Computing Science and Automatic Control (CCE) , pages=

    Enhanced network intrusion detection using a hybrid CNN-LSTM approach on the UNSW-NB15 dataset , author=. 2024 21st International Conference on Electrical Engineering, Computing Science and Automatic Control (CCE) , pages=. 2024 , organization=

    work page 2024

  27. [28]

    Machine Learning , volume=

    Finite-time analysis of the multiarmed bandit problem , author=. Machine Learning , volume=. 2002 , publisher=

    work page 2002

  28. [29]

    Artificial Intelligence and Statistics , pages=

    Communication-efficient learning of deep networks from decentralized data , author=. Artificial Intelligence and Statistics , pages=. 2017 , organization=

    work page 2017

  29. [30]

    2009 , publisher=

    Learning multiple layers of features from tiny images , author=. 2009 , publisher=

    work page 2009

  30. [31]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages=

    Deep residual learning for image recognition , author=. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages=

    work page

  31. [32]

    International Conference on Learning Representations , year=

    An agnostic approach to federated learning with class imbalance , author=. International Conference on Learning Representations , year=

    work page

  32. [33]

    Advances in Neural Information Processing Systems , volume=

    Fed-grab: Federated long-tailed learning with self-adjusting gradient balancer , author=. Advances in Neural Information Processing Systems , volume=

    work page

  33. [36]

    Proceedings of Machine learning and systems , volume=

    Federated optimization in heterogeneous networks , author=. Proceedings of Machine learning and systems , volume=

    work page

  34. [38]

    International Conference on Machine Learning , pages=

    The non-iid data quagmire of decentralized machine learning , author=. International Conference on Machine Learning , pages=. 2020 , organization=

    work page 2020

  35. [39]

    M., Albalkhi, I., and Naswhan, A

    Alkhawaldeh, I. M., Albalkhi, I., and Naswhan, A. J. Challenges and limitations of synthetic minority oversampling techniques in machine learning. World Journal of Methodology, 13 0 (5): 0 373, 2023

    work page 2023

  36. [40]

    Finite-time analysis of the multiarmed bandit problem

    Auer, P., Cesa-Bianchi, N., and Fischer, P. Finite-time analysis of the multiarmed bandit problem. Machine Learning, 47 0 (2): 0 235--256, 2002

    work page 2002

  37. [41]

    S., De Almeida, G

    Aurelio, Y. S., De Almeida, G. M., de Castro, C. L., and Braga, A. P. Learning from imbalanced data sets with weighted cross-entropy function. Neural Processing Letters, 50: 0 1937--1949, 2019

    work page 1937

  38. [42]

    and Kohavi, R

    Becker, B. and Kohavi, R. Adult. UCI Machine Learning Repository, 10: 0 C5XW20, 1996

    work page 1996

  39. [43]

    X-armed bandits

    Bubeck, S., Munos, R., Stoltz, G., and Szepesv \'a ri, C. X-armed bandits. Journal of Machine Learning Research, 12 0 (5), 2011

    work page 2011

  40. [44]

    Buda, M., Maki, A., and Mazurowski, M. A. A systematic study of the class imbalance problem in convolutional neural networks. Neural Networks, 106: 0 249--259, 2018

    work page 2018

  41. [45]

    J., and Br \'a s, S

    Carvalho, M., Pinho, A. J., and Br \'a s, S. Resampling approaches to handle class imbalance: a review from a data perspective. Journal of Big Data, 12 0 (1): 0 71, 2025

    work page 2025

  42. [46]

    V., Bowyer, K

    Chawla, N. V., Bowyer, K. W., Hall, L. O., and Kegelmeyer, W. P. Smote: synthetic minority over-sampling technique. Journal of Artificial Intelligence Research, 16: 0 321--357, 2002

    work page 2002

  43. [47]

    Tackling class-imbalanced learning issues based on local neighborhood information and generative adversarial networks

    Chen, C.-C., Lin, Y.-S., and Chen, H.-Y. Tackling class-imbalanced learning issues based on local neighborhood information and generative adversarial networks. Sensors & Materials, 36, 2024 a

    work page 2024

  44. [48]

    Chen, W., Yang, K., Yu, Z., Shi, Y., and Chen, C. P. A survey on imbalanced learning: latest research, applications and future directions. Artificial Intelligence Review, 57 0 (6): 0 137, 2024 b

    work page 2024

  45. [49]

    Class-balanced loss based on effective number of samples

    Cui, Y., Jia, M., Lin, T.-Y., Song, Y., and Belongie, S. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 9268--9277, 2019

    work page 2019

  46. [50]

    Implicit gradient alignment in distributed and federated learning

    Dandi, Y., Barba, L., and Jaggi, M. Implicit gradient alignment in distributed and federated learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pp.\ 6454--6462, 2022

    work page 2022

  47. [51]

    Leveraging local data sampling strategies to improve federated learning

    D \"u sing, C., Cimiano, P., and Paa en, B. Leveraging local data sampling strategies to improve federated learning. International Journal of Data Science and Analytics, pp.\ 1--20, 2024

    work page 2024

  48. [52]

    A theoretical analysis of the learning dynamics under class imbalance

    Francazi, E., Baity-Jesi, M., and Lucchi, A. A theoretical analysis of the learning dynamics under class imbalance. In International Conference on Machine Learning, pp.\ 10285--10322. PMLR, 2023

    work page 2023

  49. [53]

    Gradient reweighting: Towards imbalanced class-incremental learning

    He, J. Gradient reweighting: Towards imbalanced class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 16668--16677, 2024

    work page 2024

  50. [54]

    Deep residual learning for image recognition

    He, K., Zhang, X., Ren, S., and Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp.\ 770--778, 2016

    work page 2016

  51. [55]

    Measuring the Effects of Non-Identical Data Distribution for Federated Visual Classification

    Hsu, T.-M. H., Qi, H., and Brown, M. Measuring the effects of non-identical data distribution for federated visual classification. arXiv preprint arXiv:1909.06335, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1909

  52. [56]

    Johnson, J. M. and Khoshgoftaar, T. M. Survey on deep learning with class imbalance. Journal of Big Data, 6 0 (1): 0 1--54, 2019

    work page 2019

  53. [57]

    and Hinton, G

    Krizhevsky, A. and Hinton, G. Learning multiple layers of features from tiny images. 2009

    work page 2009

  54. [58]

    K., Zaheer, M., Sanjabi, M., Talwalkar, A., and Smith, V

    Li, T., Sahu, A. K., Zaheer, M., Sanjabi, M., Talwalkar, A., and Smith, V. Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems, 2: 0 429--450, 2020

    work page 2020

  55. [59]

    Focal loss for dense object detection

    Lin, T.-Y., Goyal, P., Girshick, R., He, K., and Doll \'a r, P. Focal loss for dense object detection. In Proceedings of the IEEE International Conference on Computer Vision, pp.\ 2980--2988, 2017

    work page 2017

  56. [60]

    Liu, Z., Miao, Z., Zhan, X., Wang, J., Gong, B., and Yu, S. X. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 2537--2546, 2019

    work page 2019

  57. [61]

    McMahan, B., Moore, E., Ramage, D., Hampson, S., and y Arcas, B. A. Communication-efficient learning of deep networks from decentralized data. In Artificial Intelligence and Statistics, pp.\ 1273--1282. PMLR, 2017

    work page 2017

  58. [62]

    and Slay, J

    Moustafa, N. and Slay, J. Unsw-nb15: a comprehensive data set for network intrusion detection systems (unsw-nb15 network data set). In 2015 Military Communications and Information Systems Conference (MilCIS), pp.\ 1--6. IEEE, 2015

    work page 2015

  59. [63]

    S., Datta, S., and Das, S

    Mullick, S. S., Datta, S., and Das, S. Generative adversarial minority oversampling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 1695--1704, 2019

    work page 2019

  60. [64]

    An improved generative adversarial network to oversample imbalanced datasets

    Pan, T., Pedrycz, W., Yang, J., and Wang, J. An improved generative adversarial network to oversample imbalanced datasets. Engineering Applications of Artificial Intelligence, 132: 0 107934, 2024

    work page 2024

  61. [65]

    Pear, Z. T. and Kibria, H. B. Enhanced network intrusion detection using a hybrid cnn-lstm approach on the unsw-nb15 dataset. In 2024 21st International Conference on Electrical Engineering, Computing Science and Automatic Control (CCE), pp.\ 1--6. IEEE, 2024

    work page 2024

  62. [66]

    An agnostic approach to federated learning with class imbalance

    Shen, Z., Cervino, J., Hassani, H., and Ribeiro, A. An agnostic approach to federated learning with class imbalance. In International Conference on Learning Representations, 2022. URL https://openreview.net/forum?id=Xo0lbDt975

    work page 2022

  63. [67]

    Addressing class imbalance in federated learning

    Wang, L., Xu, S., Wang, X., and Zhu, Q. Addressing class imbalance in federated learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pp.\ 10165--10173, 2021

    work page 2021

  64. [68]

    Dynamic curriculum learning for imbalanced data classification

    Wang, Y., Gan, W., Yang, J., Wu, W., and Yan, J. Dynamic curriculum learning for imbalanced data classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 5017--5026, 2019

    work page 2019

  65. [69]

    Fast: enhancing federated learning through adaptive data sampling and local training

    Wang, Z., Xu, H., Xu, Y., Jiang, Z., Liu, J., and Chen, S. Fast: enhancing federated learning through adaptive data sampling and local training. IEEE Transactions on Parallel and Distributed Systems, 35 0 (2): 0 221--236, 2023

    work page 2023

  66. [70]

    Fednoro: Towards noise-robust federated learning by addressing class imbalance and label noise heterogeneity

    Wu, N., Yu, L., Jiang, X., Cheng, K.-T., and Yan, Z. Fednoro: Towards noise-robust federated learning by addressing class imbalance and label noise heterogeneity. arXiv preprint arXiv:2305.05230, 2023

    work page arXiv 2023

  67. [71]

    T., Wu, J., Yang, H., and Liu, Z

    Xiao, Z., Chen, Z., Liu, S., Wang, H., Feng, Y., Hao, J., Zhou, J. T., Wu, J., Yang, H., and Liu, Z. Fed-grab: Federated long-tailed learning with self-adjusting gradient balancer. Advances in Neural Information Processing Systems, 36: 0 77745--77757, 2023

    work page 2023

  68. [72]

    Adaptive gradient normalization and independent sampling for (stochastic) generalized-smooth optimization

    Yang, Y., Tripp, E., Sun, Y., Zou, S., and Zhou, Y. Adaptive gradient normalization and independent sampling for (stochastic) generalized-smooth optimization. arXiv preprint arXiv:2410.14054, 2024

    work page arXiv 2024

  69. [73]

    A survey on class imbalance in federated learning

    Zhang, J., Li, C., Qi, J., and He, J. A survey on class imbalance in federated learning. arXiv preprint arXiv:2303.11673, 2023

    work page arXiv 2023