Home / Curated / Algorithms / FedLC (Federated Learning via Logits Calibration for Label Distribution Skew)

FedLC (Federated Learning via Logits Calibration for Label Distribution Skew)

Algorithms AIMech:FedLcLogitsCalibrationForLabelDistributionSkewInFederatedLearning DRAFT

Aliases: FedLC, Federated Learning via Logits Calibration, FedLC Logits Calibration

A label-skew-calibrated federated learning algorithm (Zhang et al. 2022, Sections 3-4, Definitions 1-3, Theorems 1-3, Equations 1-8) that mitigates minority-class classifier collapse and local feature deviation under non-IID label distribution skew by calibrating each client's predicted logits before computing the local softmax cross-entropy loss using a pairwise label margin Delta(n_y, n_i) proportional to the inverse fourth root of the client's local class sample counts: L_cal(y; f(x)) = -log(exp(f_y(x) - tau * n_y^{-1/4}) / sum_i exp(f_i(x) - tau * n_i^{-1/4})) (Equation 6).

Identity Notes & Concept Scope Distinct from centralized Logit Adjustment (AIMech:LogitAdjustmentForLongTailedClassification, Menon et al. 2021), which shifts logits by log pi_y to achieve Fisher consistency for balanced error on a single centralized long-tailed dataset; FedLC derives a client-specific pairwise margin tau * n_y^{-1/4} from the Rademacher-complexity generalization bound across heterogeneous federated clients and proves via Theorem 3 that it enlarges the majority-to-minority gradient ratio D_{j,r} by exp(tau*(n_r^{-1/4} - n_j^{-1/4})) > 1 to prevent local minority-class weight Updates Delta w_r from having negative inner product with minority representations h^{(r)}.

Claims & Grounded Evidence

4 Claims

Mechanism #r453-minority-gradient-reversal-and-calibrated-softmax-mechanism

Proves in Theorem 1 that under standard local cross-entropy training on skewed client label counts n_j >> n_r, the minority-class linear classifier update Delta w_r satisfies E[Delta w_r^T h^{(r)}] < 0 whenever correlated majority classes dominate the negative gradient push, and proves in Theorem 3 that subtracting tau * n_i^{-1/4} from each class logit i in L_cal (Equation 6) multiplies the gradient suppression ratio D_{j,r} by exp(tau * (n_r^{-1/4} - n_j^{-1/4})) > 1 (Equation 8) to restore E[Delta w_r^T h^{(r)}] >= 0.

supports Source: Federated Learning with Label Distribution Skew via Logits Calibration · Locator: Section 3.3 Theorem 1 & Section 4.1 Equations 6-8, Theorems 2-3, pp. 3-6

Scope #r453-classifier-and-feature-deviation-reduction-scope

Simultaneously reduces both client-to-oracle classifier weight deviation (Definition 1) and feature extractor representation deviation (Definition 2) during local client updates without transmitting any auxiliary parameters, class histograms, or proxy features to the server.

supports Source: Federated Learning with Label Distribution Skew via Logits Calibration · Locator: Section 3.2 Definitions 1-2 & Section 4.2 Discussions, pp. 2-6

Limitation #r453-zero-count-missing-class-and-tau-tuning-limitation

Requires positive local class sample counts n_i > 0 (or Laplace smoothing / clamping when a client has zero samples for a class n_i = 0 so n_i^{-1/4} remains finite) and tuning of the calibration scale hyperparameter tau.

supports Source: Federated Learning with Label Distribution Skew via Logits Calibration · Locator: Section 4.1 Equation 6 & Section 5.3 Sensitivity of Hyperparameter tau, pp. 4-8

Application #r453-label-skew-and-long-tailed-federated-classification-application

Applied to federated image classification under Dirichlet label distribution skew and globally long-tailed federated benchmarks (CIFAR-10/100-LT, ImageNet-LT, iNaturalist).

supports Source: Federated Learning with Label Distribution Skew via Logits Calibration · Locator: Section 5 Experiments (Tables 1-5, Figures 4-6), pp. 6-8

⚠️ Curated Limitations & Negative Results

Requires smoothing or clamping when a client's local sample count n_i for class i is zero so that n_i^{-1/4} is well-defined.
Addresses label distribution skew in classification tasks and does not directly calibrate regression or unlabeled self-supervised federated objectives.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:FederatedLearning` Federated Learning	narrow_match	FedLC is a federated learning method designed specifically to handle client label distribution skew via local logit margin calibration.

Evidence Sources & Primary Literature

3 Sources

Source ID	Publication / Document	Type / Access	Reference & Checksum
`r453-primary-pdf`	Federated Learning with Label Distribution Skew via Logits Calibration Jie Zhang, Zhiqi Li, Bo Li, Jie Xu, Shuang Wu, Shouhong Ding, Chao Wu	primary paper full_text	ID: https://proceedings.mlr.press/v162/zhang22p.html sha256:4fd226449125…
`r453-arxiv-abs`	Federated Learning with Label Distribution Skew via Logits Calibration (arXiv abstract) Jie Zhang, Zhiqi Li, Bo Li, Jie Xu, Shuang Wu, Shouhong Ding, Chao Wu	primary paper abstract_only	ID: https://arxiv.org/abs/2209.00189 sha256:cdc3282809a3…
`r453-metadata-cr`	Federated Learning with Label Distribution Skew via Logits Calibration (Crossref metadata) Jie Zhang, Zhiqi Li, Bo Li, Jie Xu, Shuang Wu, Shouhong Ding, Chao Wu	primary paper abstract_only	ID: https://proceedings.mlr.press/v162/zhang22p.html sha256:3c776274d0f2…

Ontology & Grounding Topology

Open in interactive network graph →

Curation Provenance

Curator: Antigravity

Method: ai_assisted

Model: Gemini

Created: 2026-10-19T08:56:00Z

Updated: 2026-10-19T08:56:00Z

Source File: data/curated/FedLcLogitsCalibrationForLabelDistributionSkewInFederatedLearning.yaml · Schema: curated_concept.yaml

Explore across X-Mech

Related computational & biological Mechs: Embedding Map Network Graph ProteinTraitsMech (ESM-2 pLM) DUFMech (Domain pLM) TraitMech CultureMech All 14 X-Mech Repositories