Heterogeneous Federated Learning Systems for Time-Series Power Consumption Prediction with Multi-Head Embedding Mechanism

Gautam Srivastava; Jerry Chun-Wei Lin; Jia-Hao Syu; Unil Yun

REVIEW 4 major objections 4 minor 33 references

Heterogeneous Federated Learning Systems for Time-Series Power Consumption Prediction with Multi-Head Embedding Mechanism

T0 review · 4 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that a multi-head heterogeneous federated learning scheme, sharing only head-network weights and four-number embedding vectors, reduces power-consumption prediction error by 24.9% to 94.1% against the strongest benchmark.

desk verdict A plausible mechanism for embedding-based heterogeneous federated learning, but the headline error reductions rest on single-run experiments and test-set-tuned hyperparameters. read the letter →

arxiv 2501.12136 v1 pith:MGXL7REL submitted 2025-01-21 cs.LG

Jia-Hao Syu , Jerry Chun-Wei Lin , Gautam Srivastava , Unil Yun This is my paper · ORCID

classification cs.LG

keywords FederatedLearningHeterogeneousMulti-HeadTime-SeriesPredictionPowerConsumptionHeadNetworkEmbeddingKnowledgeTransferAsynchronous

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 for power-consumption prediction usually assumes every client has the same model and data distribution, which breaks when vehicles or factories collect different feature sets. The paper tries to establish that heterogeneous clients can still help each other: give each time-series feature its own small classifier called a head network, embed each head into a four-number force vector derived from its trend-classification behavior, and let each client blend its heads with the heads of the most vector-similar sources in a shared pool. On four vehicle power-consumption datasets, the resulting MHHFL systems are claimed to lower prediction error by 24.9% to 94.1% against the best benchmark, and to keep an edge of 4.9% to 94.5% when predictions are delayed by five time steps. Ablation results are used to argue that the embedding-and-selection mechanism, rather than the multi-head architecture or random transfer, is what drives the gain. If true, the contribution is a recipe for knowledge transfer across domains with different feature counts and meanings without sharing raw data.

What carries the argument

The load-bearing object is the head network embedding $E_j^i$: a four-number vector made of two pulling-pushing force pairs computed from the output layer of a binary trend classifier (Eqs. 7-9 for the gradient-based version, Eq. 10 for the data-based version). It turns an entire trained head into a position in a shared space, so that the selection mechanism can order source heads by Euclidean distance to the target head (Eqs. 11 and 14) and weight the blend with a softmax over those distances (Eqs. 12-13 and 15). The same vector also defines which knowledge enters the source pool and supports asynchronous federated learning by letting clients skip sharing rounds.

What would settle it

Train MHHFL on paired synthetic datasets where the beneficial transfer direction is known by construction; then check whether the heads selected by embedding distance are the ones that reduce a target's test error. If distance-selected blends are no better than random blends over many seeds, or if two datasets known to share no mechanism still have nearby embeddings, the central mechanism is not what drives the reported gains.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is an ordered transfer mechanism for heterogeneous federated learning. For each feature in each client, a tiny head network classifies whether the feature rises or falls; after local training, that head is summarized by the cross-entropy gradient of its softmax output layer, decomposed into pulling and pushing forces for the two output neurons (or by a data-based variant using weights instead of gradients), giving a four-number embedding $E_j^i \in \mathbb{R}^4$. Clients upload only these embeddings plus head weights to a source pool. For a target head, the mechanism picks the source head with the smallest average squared difference between embeddings, chooses either the single best head (MHHFL-S) or the best head per source domain and blends them by softmax weights (MHHFL-M), and mixes weights with scale $\alpha$. The authors report that all four MHHFL variants beat every single-source benchmark on all four test sets, with the data-embedding variants ranking best, and that the gap to the strongest benchmark is a 24.9% to 94.1% MSE reduction. They also report that FedAvg and random selection often hurt, whereas their selection reliably helps.

Load-bearing premise

The claim rests on the assumption that the four-number force vector, computed from the trend classifier's output layer, faithfully captures what a head network has learned, so that near heads in embedding space transfer usefully when blended.

Editorial extensions

If this is right

If the embedding distance orders transferability, then federated learning can work across clients whose feature sets have different sizes, meanings, and distributions, without any external or public dataset to align them.
Sharing only head-network weights and four-number vectors shrinks the communication overhead and reduces the surface for gradient-inversion privacy attacks compared with sharing whole models.
The source pool naturally tolerates clients that miss communication rounds, since each round selects from whatever is currently in the pool.
The same selection logic could be used to exclude suspicious heads, because a source whose embedding is far from every target contributes little to any blend.
The recommended blending scale near $\alpha = 0.1$ shows that a small, selective knowledge pull beats both no transfer and heavy averaging.

Reading between the lines

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

None of this is stated in the paper, but the embedding could serve as a cheap domain-similarity metric for other transfer problems: once trained, two heads' distance might predict whether one dataset helps another before running a full training experiment.
A direct testable extension is to run MHHFL on synthetic datasets with known transfer structure; if the selected source heads are not the ones that truly reduce test error, the mechanism is selecting on a correlate rather than the cause.
The paper compares selection against random transfer, but not against other vector-based weighting schemes; a natural follow-up is whether cosine similarity or learned compatibility weights beat Euclidean distance on the same four datasets.
The security claim that selection filters malicious clients is argued rather than measured; poisoning attacks with known bad embeddings would be the easiest way to test it.

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.

Desk Editor's Note

read the letter

The paper's core idea is worth taking seriously. Each head network is embedded as a 4-D vector of pulling-pushing forces computed from output-layer gradients, and a target head selects source heads by L2 distance and blends their weights with a softmax-weighted average. The combination of gradient-based embedding and nearest-neighbor selection is genuinely new, as far as I can tell, and the authors are honest that the multi-head and data-based embedding pieces come from their prior work and FedRS.

What the paper does well: the mechanism is specified concretely enough to reimplement, the ablation isolates the contribution of each piece, and the result that random transfer hurts but selection helps (Ver5 vs Ver4) is a meaningful empirical observation. The task — power-consumption prediction across heterogeneous clients with different feature sets — is a real application for heterogeneous FL.

The soft spots are exactly where the reader put them. All numbers in Table 4 and Figure 3 are single-run MSEs with no error bars, no seeds, no significance tests. The smallest advertised gain, 24.9% on HuskyA (1,477 vs 1,966), is the kind of difference that can invert under re-initialization, so the "significantly outperform" language in the abstract is not supported. The blending scale α was chosen from test-set sensitivity analysis (Section 4.5), which risks inflating the reported results. No code or data are released, so the numbers cannot be checked independently.

There are also two smaller issues. The loss in Eq. (5) is written without a negative sign; as a "loss" it should be negative log-likelihood, and the gradient in Eq. (6) is consistent with the written (positive) objective, which means gradient descent on that L would actually increase likelihood. This is a sign error in the derivation, even if the implementation uses standard cross-entropy. And the abstract says "2-dimensional vectors," but E is in R^4. Minor, but sloppy.

The privacy and security claims (filtering malicious clients, preventing data poisoning) are asserted in the contributions but never tested. The selection mechanism might do this, but there is no experiment.

I don't think the mechanism is bad — it's plausible and the ablation supports it as a direction. But the evidence as presented is not enough to trust the 24.9–94.1% claim. The fix is straightforward: repeat with 5–10 seeds, report mean ± std, run a paired test, and pick α via validation rather than test. That's a revise-and-resubmit, not a reject.

Who is this for? Researchers working on client selection and knowledge transfer in heterogeneous FL, particularly for time-series data. A reader building an embedding-based selection mechanism would find useful building blocks here. It deserves a serious referee; I'd send it out, but ask for the repeated-seed experiments first.

Referee Report

4 major / 4 minor

Summary. The paper proposes Multi-Head Heterogeneous Federated Learning (MHHFL) for time-series power consumption prediction. Each client trains multiple head networks (one per input feature) that classify whether each feature increased, plus a prediction network. During federated periods, clients compute a four-dimensional 'pulling-pushing' embedding for each head (gradient-based or data-based), share head weights and embeddings with a central source pool, and then blend each target head with the source heads whose embeddings are closest in L2 distance, using either single or multiple selection. The paper reports experiments on four heterogeneous domains (AIUT, HuskyA, HuskyB, HuskyC) against SMC, NC, ENG, and TASE, claiming 24.9–94.1% test-MSE reduction over the best baseline, plus a robustness evaluation with delayed prediction, ablation studies, and a sensitivity analysis for the blending scale.

Significance. If the claimed improvements are statistically reliable, MHHFL would be a useful contribution to heterogeneous federated learning for time series: the source-pool design is asynchronous, the method shares only head-network weights and compact embeddings, and the computational complexity is argued to match FedAvg. The evaluation setup covers multiple heterogeneous domains and includes a delayed-prediction robustness check. However, the paper currently provides no repeated trials, no error bars, and no significance tests; the main quantitative claims are single-run point estimates, and the key hyperparameter α is chosen using the test sets that later produce the headline numbers. No code or data are released, so the empirical results are not independently checkable. The contribution is therefore plausible but not yet established.

major comments (4)

[Section 4.2, Table 4] The central claim that MHHFL 'significantly outperform[s]' TASE is supported only by a single MSE value per system per dataset; no standard deviations, confidence intervals, repeated seeds, or paired significance tests are reported. The smallest claimed gain (HuskyA: MHHFL-SD 1,477 vs. TASE 1,966, a 24.9% reduction) is within the range of seed-to-seed variability one would expect from a 20% test split and 25 training epochs, so the headline number cannot be evaluated. The authors should report mean ± std over at least five seeds and run paired significance tests (e.g., Wilcoxon signed-rank) across datasets; no code or data are provided to reproduce the reported values.
[Section 4.5, Figure 5] The blending scale α = 0.10 used in all main experiments is selected by inspecting test-set MSE on the same test sets reported in Table 4. This makes the reported improvements optimistic because the test data have effectively been used for model selection; the sensitivity analysis should be based on validation MSE, or the whole procedure should be nested so that the final reported test numbers are not the ones used to choose α.
[Section 4.4, Figure 4] The ablation attributes the gains to the embedding and selection mechanisms by comparing Ver4 (random selection) with Ver5 (proposed selection), but these comparisons are also single-run point estimates. For example, the claimed reductions of 60.0%, 3.8%, 9.0%, and 7.3% from Ver4 to Ver5 have no spread or significance measure, so the causal attribution is not established; repeated-seed runs with statistical testing are needed.
[Section 4.3, Figure 3] The robustness evaluation suffers from the same single-run limitation: the claim that MHHFL has 'almost no overfitting' and 'significantly outperforms' baselines is based on individual MSE bars with no measure of variability. In particular, the smallest gain in this section (HuskyA: 3,101 vs 3,261, 4.9%) is too close to noise to support a claim of superiority without confidence intervals.

minor comments (4)

[Section 4.1] The dataset description says 'HuskyB contains 35 experiments with 73,545 instances; HuskyB contains 17 experiments with 23,342 instances'; the second HuskyB should presumably be HuskyC.
[Section 4.3] The text first says 'predicting power consumption after 5 time steps' but then says 'use Xt to predict yt+4', which is a 4-step-ahead prediction; please make this consistent.
[Equations (7)–(9)] The definitions of pulling and pushing forces omit parentheses around the average term and use T both as a time index and as the batch size; please clarify the notation.
[Section 4.3] The sentence 'The experimental results of robustness evaluation are presented in Figure 3, which obtain similar results in Figure 3' appears self-referential; it should refer to the previous section or to a specific comparison.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims are empirical system comparisons supported by external baselines and ablations, not derivations that reduce to their inputs.

full rationale

The paper's derivation chain is an empirical system design rather than a formal derivation, and no step reduces by construction to its own inputs. The head-network embedding in Section 3.3.1 is adapted from FedRS [32] and from the authors' prior work [31], but the selection and blending mechanisms in Section 3.3.2 are evaluated against random selection (Ver4) and FedAvg (Ver3) in the Section 4.4 ablation, so the claimed benefit of embedding-based selection is not forced by definition. The headline 24.9-94.1% reduction is computed against TASE [19], which is the authors' prior work, but the comparison uses reported MSE numbers and does not define MHHFL in terms of TASE; this is a baseline-selection choice, not a circular step. The sensitivity analysis on alpha and the fixed hyperparameters are transparent, and no fitted parameter is renamed as a prediction. The absence of repeated-seed error bars is a statistical robustness concern, not a circularity concern. Overall, the paper is self-contained against external benchmarks and its own ablations, so no significant circularity is present.

Assumptions & free parameters 6 free parameters · 5 assumptions · 2 invented entities

The central claims depend on hyperparameters (alpha, DR, W, architecture, batch size) and on unverified modeling assumptions about the embedding and selection. The most important is that the four-dimensional gradient-based vector orders cross-domain model compatibility; the paper provides ablation evidence but no formal justification.

free parameters (6)

blending scale alpha = 0.10
Chosen from test-set sensitivity analysis in Section 4.5; final results use 0.10.
dropout probability DR = 0.5
Probability a domain skips sharing in a federated period; hand-set in Section 4.1 without sensitivity analysis.
window size W = 5
Number of past time steps used per feature; fixed in Section 4.1.
head network layer widths = 8-64-32-8 per Table 3
Architecture sizes chosen by hand; no architecture search is reported.
training epochs = 25 per domain for MHHFL, 100 for baselines
Chosen so that MHHFL total epochs across four domains equals 100; affects the fairness of comparison.
batch size = 600
Used for training and defines the federated period (every ten batches); hand-set in Section 4.1.

assumptions (5)

domain assumption The four-dimensional pulling-pushing force vector E is a faithful representation of a head network's transferable knowledge.
Adopted in Section 3.3.1 and used for selection in Section 3.3.2; no theoretical or independent empirical proof is provided.
domain assumption The binary pre-classification C_i,t (increase vs non-increase) is a meaningful common signal across heterogeneous domains.
Defined in Eq. (2) and used as the head-network supervision and embedding basis; if the trend semantics differ across domains, the sharing may be harmful.
domain assumption L2 distance in embedding space orders the usefulness of source head networks for a target head network.
Eqs. (11)-(14) select nearest neighbors by Avg((E)^2); this proximity heuristic is not justified beyond intuition.
domain assumption The benchmark systems [16,17,18,19] are reimplemented fairly with comparable parameter counts.
Section 4.1 states neurons were adjusted to similar parameter budgets, but the adjustment procedure and training details are not fully described.
standard math The softmax cross-entropy gradient decomposition into class-wise forces is a valid basis for model compatibility comparison.
Gradient formulas in Eqs. (4)-(9) use standard calculus, but the 'loss' sign in Eq. (5) conflicts with the derived gradient direction.

invented entities (2)

Pulling-pushing force embedding vector E independent evidence
purpose: Four-number descriptor of each head network used to select and blend compatible source heads (Section 3.3.1).
The ablation comparing selection to random transfer provides indirect evidence that E carries usable information, but the interpretation of E as a physical force is metaphorical.
Centralized source pool
purpose: Repository that stores shared head-network weights and embeddings to support asynchronous federated learning (Section 3.3).
A design artifact; its asynchronous operation is described but not tested against a synchronous baseline.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Heterogeneous Federated Learning Systems for Time-Series Power Consumption Prediction with Multi-Head Embedding Mechanism." pith.science (2026). https://pith.science/paper/MGXL7REL

@misc{pith2026250112136,
  author       = {Pith},
  title        = {Pith review of: Heterogeneous Federated Learning Systems for Time-Series Power Consumption Prediction with Multi-Head Embedding Mechanism},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MGXL7REL}},
  note         = {Machine review of arXiv:2501.12136}
}

read the original abstract

Time-series prediction is increasingly popular in a variety of applications, such as smart factories and smart transportation. Researchers have used various techniques to predict power consumption, but existing models lack discussion of collaborative learning and privacy issues among multiple clients. To address these issues, we propose Multi-Head Heterogeneous Federated Learning (MHHFL) systems that consist of multiple head networks, which independently act as carriers for federated learning. In the federated period, each head network is embedded into 2-dimensional vectors and shared with the centralized source pool. MHHFL then selects appropriate source networks and blends the head networks as knowledge transfer in federated learning. The experimental results show that the proposed MHHFL systems significantly outperform the benchmark and state-of-the-art systems and reduce the prediction error by 24.9% to 94.1%. The ablation studies demonstrate the effectiveness of the proposed mechanisms in the MHHFL (head network embedding and selection mechanisms), which significantly outperforms traditional federated average and random transfer.

Figures

Figures reproduced from arXiv: 2501.12136 by the authors.

**Figure 2.** For each federated period (detailed in Section 3.3.3), each source [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

**Figure 2.** Proposed heterogeneous federated learning [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗

Figures from the paper (3 more)

**Figure 3.** Figure 3: Robustness Evaluation 21 [PITH_FULL_IMAGE:figures/full_fig_p021_3.png]

**Figure 4.** Figure 4: Ablation studies on four datasets 4.4. Ablation Studies In this section, we perform ablation studies to evaluate the effectiveness of each proposed mechanism, and we compare the performance of eight versions of MHHFL (Ver1 to Ver8). Ver5, Ver6, Ver7, and Ver8 refer to …

**Figure 5.** Figure 5: Sensitivity analysis on blending scale α MHHFL, we perform the sensitivity analysis on the proposed MHHFL-MG, as shown in [PITH_FULL_IMAGE:figures/full_fig_p024_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 27 canonical work pages

[1]

Zhang, Y

C. Zhang, Y. Xie, H. Bai, B. Yu, W. Li, Y. Gao, A survey on federated learning, Knowledge-Based Systems 216 (2021) 106775

2021
[2]

Koneˇ cn` y, H

J. Koneˇ cn` y, H. B. McMahan, F. X. Yu, P. Richt´ arik, A. T. Suresh, D. Bacon, Federated learning: Strategies for improving communication efficiency, arXiv preprint arXiv:1610.05492 (2016)

arXiv 2016
[3]

McMahan, E

B. McMahan, E. Moore, D. Ramage, S. Hampson, B. A. y Arcas, Communication-efficient learning of deep networks from decentralized data, in: Artificial intelligence and statistics, PMLR, 2017, pp. 1273– 1282

2017
[4]

Y. Cui, K. Cao, G. Cao, M. Qiu, T. Wei, Client scheduling and resource management for efficient training in heterogeneous iot-edge federated learning, IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems 41 (8) (2021) 2407–2420

work page 2021
[5]

Zhang, X

J. Zhang, X. Cheng, C. Wang, Y. Wang, Z. Shi, J. Jin, A. Song, W. Zhao, L. Wen, T. Zhang, Fedada: Fast-convergent adaptive federated learning in heterogeneous mobile edge computing environment, World Wide Web 25 (5) (2022) 1971–1998

work page 2022
[6]

Z. Wang, P. Yu, H. Zhang, Privacy-preserving regulation capacity eval- uation for hvac systems in heterogeneous buildings based on federated learning and transfer learning, IEEE Transactions on Smart Grid (2022)

work page 2022
[7]

K. Guo, Z. Chen, H. H. Yang, T. Q. Quek, Dynamic scheduling for het- erogeneous federated learning in private 5g edge networks, IEEE Journal of Selected Topics in Signal Processing 16 (1) (2021) 26–40. 26

work page 2021
[8]

Huang, M

W. Huang, M. Ye, B. Du, Learn from others and be yourself in heteroge- neous federated learning, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 10143–10153

work page 2022

Show all 33 references

[9]

J. H. Syu, J. C. W. Lin, M. Fojcik, R. Cupek, Htps: Heterogeneous transferring prediction system for healthcare datasets, in: 31st Euromi- cro International Conference on Parallel, Distributed, and Network- Based Processing, 2023

2023
[10]

Alzoubi, Machine learning for intelligent energy consumption in smart homes, International Journal of Computations, Information and Manufacturing (IJCIM) 2 (1) (2022)

A. Alzoubi, Machine learning for intelligent energy consumption in smart homes, International Journal of Computations, Information and Manufacturing (IJCIM) 2 (1) (2022)

2022
[11]

Kim, D.-H

E. Kim, D.-H. Huh, S. Kim, Knowledge-based power monitoring and fault prediction system for smart factories, Personal and Ubiquitous Computing (2022) 1–12

2022
[12]

J. H. Syu, J. C. W. Lin, G. Srivastava, Distributed learning mechanisms for anomaly detection in privacy-aware energy grid management system (2023)

2023
[13]

Zheng, Y

Q. Zheng, Y. Hou, H. Yang, P. Tan, H. Shi, Z. Xu, Z. Ye, N. Chen, X. Qu, X. Han, et al., Towards a sustainable monitoring: A self-powered smart transportation infrastructure skin, Nano Energy 98 (2022) 107245

2022
[14]

Islam, A

S. Islam, A. Iqbal, M. Marzband, I. Khan, A. M. Al-Wahedi, State-of- the-art vehicle-to-everything mode of operation of electric vehicles and its future perspectives, Renewable and Sustainable Energy Reviews 166 (2022) 112574

2022
[15]

H. Wei, L. Fan, Q. Ai, W. Zhao, T. Huang, Y. Zhang, Optimal energy allocation strategy for electric vehicles based on the real-time model predictive control technology, Sustainable Energy Technologies and As- sessments 50 (2022) 101797

2022
[16]

Kim, S.-B

T.-Y. Kim, S.-B. Cho, Predicting residential energy consumption using cnn-lstm neural networks, Energy 182 (2019) 72–81

2019
[17]

Benecki, D

P. Benecki, D. Kostrzewa, P. Grzesik, B. Shubyn, D. Mrozek, Forecast- ing of energy consumption for anomaly detection in automated guided vehicles: Models and feature selection, in: 2022 IEEE International 27 Conference on Systems, Man, and Cybernetics (SMC), IEEE, 2022, pp. 2073–2079

2022
[18]

Y. Li, I. Y. Zeng, Z. Niu, J. Shi, Z. Wang, Z. Guan, Predicting vehicle fuel consumption based on multi-view deep neural network, Neurocom- puting 502 (2022) 140–147

2022
[19]

J. H. Syu, J. C. W. Lin, P. S. Yu, Distributed multi-head learning sys- tems for power consumption prediction (2023)

2023
[20]

Geiping, H

J. Geiping, H. Bauermeister, H. Dr¨ oge, M. Moeller, Inverting gradients- how easy is it to break privacy in federated learning?, Advances in Neural Information Processing Systems 33 (2020) 16937–16947

2020
[21]

S. Li, Y. Cheng, W. Wang, Y. Liu, T. Chen, Learning to detect malicious clients for robust federated learning, arXiv preprint arXiv:2002.00211 (2020)

2020 arXiv
[22]

Tolpegin, S

V. Tolpegin, S. Truex, M. E. Gursoy, L. Liu, Data poisoning attacks against federated learning systems, in: Computer Security–ESORICS 2020: 25th European Symposium on Research in Computer Security, ESORICS 2020, Guildford, UK, September 14–18, 2020, Proceedings, Part I 25, Spr...

2020
[23]

G. Long, Y. Tan, J. Jiang, C. Zhang, Federated learning for open bank- ing, in: Federated Learning: Privacy and Incentive, Springer, 2020, pp. 240–254

2020
[24]

L. U. Khan, W. Saad, Z. Han, E. Hossain, C. S. Hong, Federated learning for internet of things: Recent advances, taxonomy, and open challenges, IEEE Communications Surveys & Tutorials 23 (3) (2021) 1759–1799

2021
[25]

F. Yu, W. Zhang, Z. Qin, Z. Xu, D. Wang, C. Liu, Z. Tian, X. Chen, Het- erogeneous federated learning, arXiv preprint arXiv:2008.06767 (2020)

2020 arXiv
[26]

B. Luo, W. Xiao, S. Wang, J. Huang, L. Tassiulas, Tackling system and statistical heterogeneity for federated learning with adaptive client sam- pling, in: IEEE INFOCOM 2022-IEEE conference on computer commu- nications, IEEE, 2022, pp. 1739–1748. 28

2022
[27]

X. Fang, M. Ye, Robust federated learning with noisy and heterogeneous clients, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 10072–10081

2022
[28]

Z. Tang, Y. Zhang, S. Shi, X. He, B. Han, X. Chu, Virtual homogeneity learning: Defending against data heterogeneity in federated learning, in: International Conference on Machine Learning, PMLR, 2022, pp. 21111–21132

2022
[29]

J. H. Syu, J. C. W. Lin, G. Srivastava, Call auction-based energy man- agement system with adaptive subsidy and dynamic operating reserve, Sustainable Computing: Informatics and Systems (2022) 100786

2022
[30]

J. H. Syu, G. Srivastava, M. Fojcik, R. Cupek, J. C. W. Lin, Energy grid management system with anomaly detection and q-learning decision modules, Computers and Electrical Engineering 107 (2023) 108639

2023
[31]

J. H. Syu, J. C. W. Lin, G. Srivastava, Mhts: Multi-head heterogenous transfer system of power consumption prediction for electric vehicles (2022)

2022
[32]

Li, D.-C

X.-C. Li, D.-C. Zhan, Fedrs: Federated learning with restricted softmax for label distribution non-iid data, in: Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, 2021, pp. 995–1005

2021
[33]

Choudhry, S

A. Choudhry, S. Lau, J. Patrikar, B. Moon, T. A. Rodrigues, A. Gahlaut, Energy consumption data for package delivery with an uncrewed ground vehicle (2021). URL https://doi.org/10.1184/R1/13623944.v1 29

2021 doi

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

Zhang, Y

C. Zhang, Y. Xie, H. Bai, B. Yu, W. Li, Y. Gao, A survey on federated learning, Knowledge-Based Systems 216 (2021) 106775

2021

[2] [2]

Koneˇ cn` y, H

J. Koneˇ cn` y, H. B. McMahan, F. X. Yu, P. Richt´ arik, A. T. Suresh, D. Bacon, Federated learning: Strategies for improving communication efficiency, arXiv preprint arXiv:1610.05492 (2016)

arXiv 2016

[3] [3]

McMahan, E

B. McMahan, E. Moore, D. Ramage, S. Hampson, B. A. y Arcas, Communication-efficient learning of deep networks from decentralized data, in: Artificial intelligence and statistics, PMLR, 2017, pp. 1273– 1282

2017

[4] [4]

Y. Cui, K. Cao, G. Cao, M. Qiu, T. Wei, Client scheduling and resource management for efficient training in heterogeneous iot-edge federated learning, IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems 41 (8) (2021) 2407–2420

work page 2021

[5] [5]

Zhang, X

J. Zhang, X. Cheng, C. Wang, Y. Wang, Z. Shi, J. Jin, A. Song, W. Zhao, L. Wen, T. Zhang, Fedada: Fast-convergent adaptive federated learning in heterogeneous mobile edge computing environment, World Wide Web 25 (5) (2022) 1971–1998

work page 2022

[6] [6]

Z. Wang, P. Yu, H. Zhang, Privacy-preserving regulation capacity eval- uation for hvac systems in heterogeneous buildings based on federated learning and transfer learning, IEEE Transactions on Smart Grid (2022)

work page 2022

[7] [7]

K. Guo, Z. Chen, H. H. Yang, T. Q. Quek, Dynamic scheduling for het- erogeneous federated learning in private 5g edge networks, IEEE Journal of Selected Topics in Signal Processing 16 (1) (2021) 26–40. 26

work page 2021

[8] [8]

Huang, M

W. Huang, M. Ye, B. Du, Learn from others and be yourself in heteroge- neous federated learning, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 10143–10153

work page 2022