acceptodds
Withdrawn from ICLR 2027

Calibration Is Not Enough: Support Restoration and Residual Correctability in Federated Long-Tailed Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Federated Long-Tailed learning (Fed-LT) couples global class imbalance with fragmented client label support. Exact local-prior calibration removes locally missing classes from the client's calibrated softmax competition, even when those classes are observed elsewhere in the federation. Restoring class competition, however, does not eliminate client-update heterogeneity, and the resulting residuals alone do not determine how the server should correct the aggregate update. Therefore, we propose Federated Prior Recalibration and Residual Stabilization (FedPRS). On clients, Federation-Anchored Calibration (FAC) mixes local and federation priors to restore competition over federation-supported classes while retaining local-prior information. On the server, Label-Geometric Regression (LGR) fits composition-associated update residuals and applies a bounded affine correction around FedAvg. Our analysis separates support restoration from predictive utility and shows that server correction depends jointly on affine correctability, directional alignment, and correction scale, while the proposed LGR explicitly controls deviation from the FedAvg update. FedPRS achieves the highest mean Macro-F1 among the reported baselines across the evaluated CIFAR-10/100-LT settings.

Primary area:
learning theory
Keywords:
Federated Long-Tailed Learning, Federated Learning, Logit Calibration, Support Restoration, Label-Aware Aggregation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_AevDOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FedPRS for federated long-tailed learning. It combines Federation-Anchored Calibration (FAC), which mixes local and federation-level class priors to restore class competition, with Label-Geometric Regression (LGR), which applies a bounded server-side correction. The method is evaluated on CIFAR-10-LT and CIFAR-100-LT under different imbalance, heterogeneity, participation, and backbone settings.

Critical strengths
  1. The paper addresses a relevant issue in federated long-tailed learning: locally absent classes can be excluded from class competition during prior calibration.

  2. The component ablations help identify the sources of improvement, showing that most of the reported gain comes from FAC rather than LGR.

Critical weaknesses
  1. Both components substantially overlap with existing methods. FedYoYo already mixes local and global class priors, although FAC places more weight on the local prior. The paper also acknowledges the relationship between LGR and FedPALS, and Appendix A.3.2 connects LGR's unscaled coefficients to regularized label matching. These similarities substantially narrow the independent methodological contribution.

  2. The experiments do not adequately validate the proposed mechanisms. FAC's gains are not separated from generic prior correction: the optimizer-matched baselines lack a simple global-prior logit-adjustment control, and Appendix B.2 acknowledges that the experiments do not causally isolate support fragmentation. For LGR, Section 3.2 states that the correction bound controls magnitude but not direction, while Theorem 2 requires favorable alignment and scale to reduce reference error. However, the paper reports no measured alignment A or reference-error reduction RSE. LGR adds only 0.045 mean Macro-F1 over FAC, with one seed declining. In the headline CIFAR-10-LT setting (α=0.1, IF=100), the 0.084 margin over CReFF is comparable to the reported seed standard deviations (approximately 0.07), while the methods use different training protocols.

Questions for authors

Can the authors report LGR's measured alignment A and reference-error reduction RSE against the matched full-participation reference, together with a FAC mixing-weight sweep including λ=0 and λ=1? These tests would clarify whether LGR provides a beneficially directed correction and whether FAC's performance gains arise from its specific local–global prior mixture rather than generic prior adjustment.

Reviewer_3TvFOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies two problems in federated long-tailed learning: missing classes on local clients may be excluded from classification, and client model updates can still differ even after restoring class competition, making FedAvg insufficient.

The authors propose FedPRS, which has two components: FAC mixes local and global class priors to restore competition among globally supported classes, while LGR uses the relationship between client label distributions and update residuals to make bounded corrections to FedAvg.

Critical strengths

Clear Method Design: FAC restores class competition on the client side, while LGR corrects model updates on the server side.

Comprehensive Experiments: The authors evaluate two datasets under different class imbalance levels and Dirichlet distributions, and test different backbones, client participation rates, and local training epochs.

Critical weaknesses

Both FAC and LGR rely heavily on class-distribution statistics. However, using class frequencies has already been explored in prior work, such as FedLC (ICML 2022).

LGR mainly uses class distribution differences to estimate client update discrepancies, but other factors, such as data heterogeneity and local optimization, may also matter. Considering these factors could improve LGR’s effectiveness and robustness.

The ablation study shows that FAC contributes most of the performance gain, while LGR provides additional improvements. However, it remains unclear whether mixing local and global priors is necessary, or whether using the global prior alone would achieve similar results. A comparison with global-prior-only calibration would help isolate the benefit of the proposed prior mixing strategy.

Questions for authors

Please see the weaknesses discussed above.

Reviewer_ThPjOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Federated Prior Recalibration and Residual Stabilization (FedPRS) to address federated long-tailed learning under global class imbalance and heterogeneous client label distributions.

The method consists of two components. First, Federation-Anchored Calibration (FAC) combines local and federation-level class priors to preserve softmax competition over globally supported classes, including those absent from individual clients. Second, Label-Geometric Regression (LGR) models the relationship between client update residuals and label-distribution differences, using the resulting structure to construct a bounded affine correction to FedAvg aggregation.

The paper provides theoretical analyses of prior calibration, support restoration, affine correctability, and correction magnitude. Experiments on CIFAR-10-LT and CIFAR-100-LT demonstrate that FedPRS achieves the highest mean Macro-F1 among the reported baselines across twelve evaluated imbalance and heterogeneity settings. Additional experiments investigate component contributions, client participation rates, model architectures, and local optimization budgets.

Critical strengths

S1. Well-motivated problem formulation and theoretical analysis. The paper identifies a meaningful limitation of local-prior calibration under fragmented label support. Proposition 1 characterizes support truncation and Hessian-rank reduction, motivating FAC for support restoration and LGR for residual correction. The theoretical analysis appropriately distinguishes structural guarantees from predictive improvements.

S2. Consistent empirical improvements. FedPRS achieves the highest mean Macro-F1 among the reported baselines across all twelve CIFAR-10/100-LT settings. Under IF = 100 and α = 0.1, it achieves 0.521/0.318 Macro-F1, compared with 0.437/0.247 for the strongest baselines. Ablation and sensitivity experiments further support its effectiveness across different configurations.

Critical weaknesses

W1. Insufficient evidence for the novelty and individual contributions of FAC and LGR.

FAC relies on a straightforward mixture of local and global priors, but comparisons with simpler alternatives, such as prior smoothing and global-prior-only calibration, are missing. It is therefore unclear whether its improvements arise from the proposed formulation or merely from restoring nonzero class probabilities.

Similarly, Appendix A.3.2 connects LGR to regularized label matching, while Eq. (9) determines aggregation coefficients from label distributions rather than directly from update residuals. Comparisons with simpler label-aware aggregation methods under the same FAC objective are needed.

Moreover, Table 3 shows that FAC alone achieves 0.476 Macro-F1 versus 0.521 for FedPRS, suggesting that most improvements originate from FAC. LGR also reduces performance for one seed. Direct measurements of correction alignment and reference-update error reduction would better validate the proposed mechanism.

W2. Potentially confounded comparisons and limited evaluation scope.

Appendix B.1 reveals differences in momentum, weight decay, learning-rate scheduling, and normalization across baselines, making it difficult to attribute performance gains solely to the proposed method. More controlled comparisons or systematic hyperparameter tuning are needed.

Additionally, experiments are limited to CIFAR-10/100-LT with 20 clients and three seeds. The missing CReFF results on CIFAR-100-LT further weaken comparison completeness. Evaluation on more diverse, naturally heterogeneous federated datasets would strengthen the generalizability claims.

Questions for authors

Same as weaknesses

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.