Pith. sign in

REVIEW 4 major objections 5 minor 47 references

AutoFed argues that personalized federated traffic prediction can drop manual tuning when shared knowledge is carried by a learned prompt matrix.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 13:16 UTC pith:UXGZYUZY

load-bearing objection AutoFed's prompt-based personalization is a real, useful idea with strong TDP results, but the TFP table imports baselines from other papers and the abstract's 'consistently superior' doesn't survive contact with the data. the 4 major comments →

arxiv 2512.24625 v3 pith:UXGZYUZY submitted 2025-12-31 cs.LG cs.AI

AutoFed: Personalized Federated Traffic Prediction via Adaptive Prompt

Zijian Zhao , Yitong Shang , Sen Li This is my paper
classification cs.LG cs.AI
keywords Federated learningPersonalized federated learningTraffic predictionPrompt learningNon-IID dataSpatiotemporal forecastingCommunication efficiencyAdaptive representation alignment
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to establish that personalized federated learning for traffic prediction does not need dataset-specific manual configuration if the shared knowledge is compressed into a learned prompt matrix rather than hand-designed filters, graphs, or pattern counts. The proposed framework, AutoFed, splits each client into a locally trained predictor and a federated representor; the representor denoises local data with an autoencoder, compresses it with a graph time-series encoder, and aligns the result into a shared prompt through an adapter with shared linear layers and client-specific normalization statistics. That prompt then conditions the decoder of the personalized predictor, so cross-client temporal patterns are injected while local specificity is retained. The practical stake is that the bottleneck for deploying federated traffic models is often per-dataset tuning, and AutoFed claims to remove that bottleneck while also cutting communication to a small prompt-generation module. Evaluations on ride-hailing demand and highway traffic flow show AutoFed matching or beating the comparison methods in most scenarios.

Core claim

AutoFed's central claim is that the reusable knowledge in a federation of traffic predictors can be compressed into a single prompt matrix that is learned rather than specified. The framework consists of a locally trained personalized predictor and a federated representor; the representor denoises the input with an autoencoder, encodes it with a graph time-series network, and then aligns the resulting local representation into a shared space through an adapter whose linear layers are common to all clients while batch-normalization statistics remain client-specific. The aligned representation becomes a prompt that the decoder consumes as its first token, so each client gets cross-client tempo

What carries the argument

The load-bearing object is the prompt matrix produced by the federated representor and fed to the decoder as its first token. It is generated in three steps: an autoencoder denoiser learns stable patterns without choosing a filter; a graph time-series encoder compresses the denoised sequence into a single local vector; and a client-aligned adapter maps that local vector into a global space using shared linear layers while keeping per-client batch-normalization statistics. The adapter's partial sharing is what lets the prompt be global enough to transfer knowledge and local enough to respect non-IID data. An adaptive loss weight balances denoising and prediction automatically, replacing a tun

Load-bearing premise

The framework's benefit rests on the assumption that the client-aligned adapter can project every client's local representation into a shared prompt space in which the aggregated prompt carries genuinely useful temporal knowledge for each client; if that alignment fails for a sufficiently different client, the shared prompt could actively hurt that client's predictions.

What would settle it

Construct a federation with clients whose traffic regimes are deliberately non-overlapping, such as one client on weekday peak hours, another on overnight weekends, and another on a different highway, then compare AutoFed's per-client error to that client's local-training error; if any client is worse with AutoFed than with local training, the claim that the shared prompt transfers useful knowledge is falsified.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Deploying the framework to a new city or operator requires no dataset-specific configuration: the prompt is generated from the client's own input, and the graph is learned adaptively rather than constructed by hand.
  • Because only the prompt-generation modules are exchanged with the server, communication per round stays small; the paper reports a fraction of the parameters of full-model aggregation while retaining the benefit of cross-client knowledge.
  • The same learned adapter can be inspected, for example by clustering as the paper does, to see whether a federation has converged to genuinely shared temporal patterns rather than averaged local ones.
  • AutoFed's performance holds across both travel-demand and traffic-flow tasks, suggesting the mechanism is not tied to one prediction target or one data granularity.

Where Pith is reading between the lines

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

  • Editorial extension: because the prompt is produced by an input-conditional network rather than a per-client lookup table, AutoFed likely generalizes to clients that never participated in training; a held-out city could be served by generating a prompt from its own data. The paper does not test this.
  • Editorial extension: the adaptive weight is a form of automatic curriculum; if the denoiser converges first, the prediction loss is up-weighted. This may explain the faster validation convergence the paper shows, but the effect is not isolated experimentally.
  • Editorial extension: the alignment adapter's design suggests the framework should transfer to other spatiotemporal forecasting domains, such as energy demand or crowd flow, without architectural change, though only traffic is tested.
  • Editorial extension: the strongest stress test would be a deliberately adversarial federation with clients whose traffic regimes are nearly disjoint, such as different times of day or demand levels. The paper's mixed Uber-plus-Lyft scenario moves in that direction, but the alignment evidence is a single clustering visualization on one scenario.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes AutoFed, a personalized federated learning framework for traffic prediction. It combines a personalized AGCRN-based predictor with a federated representor consisting of an AE denoiser, a graph time-series encoder, and a FedBN-style adapter that generates a globally shared prompt matrix conditioned on local data. The authors claim that AutoFed eliminates manual hyper-parameter tuning and consistently outperforms prior methods, including FedTPS and FedGCN, on ride-hailing demand prediction (TDP) and highway traffic flow prediction (TFP). Experiments include reproductions for TDP, imported results for TFP, ablations, a KMeans visualization of representation alignment, and a training-cost comparison.

Significance. If the empirical claims were fully supported, AutoFed would be a useful contribution: a prompt-based PFL method for traffic prediction that avoids manual graph construction and dataset-specific pattern tuning, with a compact shared model and a public code release. Strengths of the manuscript include the end-to-end learned prompt generator, the ablation of the AE and FedBN components, convergence behavior analysis, and an explicit comparison of computation and communication costs. However, the current evidence does not support the headline claim 'consistently achieves superior performance': the TFP comparison is not controlled, no variance or multiple-seed statistics are reported for TDP, and the communication-cost statement contradicts the authors' own Table 3. The underlying method is plausible and the TDP experiments are mostly reproduced in a common protocol, so the issues are addressable, but the paper needs substantial revision before the claims can be accepted.

major comments (4)
  1. [§4.2, Table 2] The headline claim 'consistently achieves superior performance' is not supported by the TFP experiment. Table 2 explicitly sources all baseline numbers from [45;46;14], while §4.1 claims all methods use the same AGCRN backbone; these statements are incompatible because imported baselines were not run in the authors' harness. The reported numbers themselves contradict the abstract: on PEMS08, AutoFed's MAE/RMSE (15.90/25.15) are worse than FedTPS (15.81/24.91), and on PEMS03, FedGCN's RMSE (23.78) is much lower than AutoFed's (25.76). Section 4.2 also concedes only 'performance similar' to previous SOTA methods. To support the central claim, the authors must rerun all baselines in the same codebase with identical preprocessing, splits, backbone, and seeds, or substantially weaken the claim to 'competitive on TFP and best on TDP.' At minimum, the abstract and conclusion should not assert c
  2. [§4.1, Table 1] The TDP task is the only controlled comparison, yet no variance, standard deviations, or multiple-seed results are reported. Many AutoFed advantages are within small numeric differences that could easily be noise, e.g., Uber S0 MAE 3.35 vs FedGCN 3.33; Lyft S2 MAPE 50.12 vs FedGCN 49.87; Lyft S3 MAPE 50.00 vs the w/o AE variant 49.50. The claim of 'SOTA in almost all scenarios' therefore is not established. The authors should report mean±std over at least three independent runs and, where appropriate, a paired significance test.
  3. [§4.3, Table 3] The text states that 'AutoFed and FedTPS have the lowest communication costs but the highest computation times,' but Table 3 reports 175.0K parameters/round for AutoFed versus 0.1K for FedTPS. AutoFed's communication cost is 1750× larger than FedTPS's, so the categorical claim is false as written. The conclusion's phrase 'reduced communication costs' is also only valid relative to most baselines, not to FedTPS. The computation comparison is similarly mixed: AutoFed is 104.6 s/round versus 76.2 for FedAvg and 106.4 for FedTPS. Please qualify these statements precisely.
  4. [§3.5, Eq. (6)] The adaptive weight α = Lae/Lpre is presented as removing manual tuning, but the total loss becomes Lpre + (Lae)^2 / Lpre. Since α is computed from current batch losses, the gradient may include second-order terms; the paper does not state whether α is treated as a detached constant. If the authors intend the ratio to reweight losses, they should clarify whether it is detached and why this formulation is stable and desirable. More broadly, the 'manual-free' claim is scoped by footnote 1 to exclude learning rate, batch size, communication rounds, and local epochs, but the hidden dimension h and the AE architecture are still manually chosen; the paper should state which framework-level hyper-parameters are truly automatic.
minor comments (5)
  1. [Eq. (1)] The equation for the PFL objective has malformed notation around the |V_i|/sum weighting; please rewrite as a weighted sum over clients with proper parentheses.
  2. [Fig. 2] The KMeans visualization lacks axis labels and a description of the number of clusters and initialization. Please add these details so the reader can interpret the alignment evidence.
  3. [§4.1/Appendix C] For the TFP task, clarify whether all AutoFed runs use the FedTPS repository's preprocessing (column learning, gradient clipping, MultiStepLR) and whether the same settings were applied to the imported baseline numbers.
  4. [Table 3/Fig. 3] Specify the hardware and the number of runs for the timing measurements; currently no variance is reported for wall-clock time or convergence curves.
  5. [§3.4] Minor typos: 'The recovered sequence ˆx be used solely' should read 'is used solely'; consider a careful proofread of equations.

Circularity Check

0 steps flagged

No derivation step reduces to its inputs; comparison and generalization concerns are empirical, not circular.

full rationale

The paper's central mechanism — a Federated Representor that distills local data into a shared prompt pg conditioning a personalized AGCRN-style predictor — is learned end-to-end against external datasets. The inference path in Eq. (4) generates pg from the input x and feeds it to the decoder; no step assumes the ground-truth Y or the evaluation metric, so the reported predictions are not definitionally equal to the training inputs. The adaptive coefficient α = Lae/Lpre in Eq. (6) is a loss-balancing rule computed from the model's own current losses, not a value fitted to benchmark outcomes, so it does not make the claimed improvements tautological. The TFP comparison does import baseline numbers from [45;46;14] (Table 2 note), and the text concedes "our method demonstrates performance similar to the two previous SOTA methods" on TFP; however, importing external baseline numbers is an evidence-quality/controlled-comparison concern, not circularity, because those numbers cannot force AutoFed's own outputs. The self-citations in the paper ([43] for the adaptive α idea, and [32;31;33] for prior federated traffic graph designs) are contextual or motivational; none is invoked as a uniqueness theorem, a fitted-parameter source, or the sole justification of the framework's contribution. No load-bearing derivation step reduces to its own inputs, so the paper is not circular.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical entities and no parameter-free theoretical derivation; its central claim rests on learned modules (AE denoiser, AGCRN encoder/decoder, FedBN adapter) whose effectiveness is taken as a design assumption supported by empirical results. The adaptive weighting α is a function of losses, not a fitted constant. Several architecture and training hyperparameters are set manually, which qualifies the 'manual-free' claim.

free parameters (3)
  • hidden dimension h (AGCRN, AE, MLP adapter) = not reported
    Chosen by hand; architecture capacity affects results; not swept or disclosed.
  • AE denoiser architecture (layer sizes, latent size) = not reported
    No details given beyond 'AE encoder/decoder'; capacity is a free hyperparameter.
  • learning rate, batch size, communication rounds, local epochs = 1e-3, 128, 50/200, 1
    Manually set training hyperparameters, contradicting 'manual-free' framing; though not framework-structure hyperparameters.
axioms (5)
  • domain assumption AGCRN's adaptive adjacency from EE^T captures spatial correlations without graph engineering (eq. 2)
    Borrowed from [4]; underlies the 'no manual graph construction' claim.
  • domain assumption Autoencoder denoising separates stable traffic patterns from noise without a frequency threshold
    Section 3.4, AE Denoiser; the core replacement for low-pass filters in FedTPS. If false, prompt quality degrades.
  • domain assumption FedBN-style shared linear layers with client-specific batch norm align local features to a common prompt space
    Section 3.4, Client-Aligned Adapter; assumed to handle non-IID while sharing knowledge.
  • domain assumption A prompt token as decoder prefix effectively guides autoregressive prediction
    Section 3.2/3.3; design assumption from prompt learning literature.
  • standard math PFL objective in Eq. 1 is well-posed with weighted average loss
    Standard PFL objective; no proof of convergence given.

pith-pipeline@v1.3.0-alltime-deepseek · 13865 in / 11344 out tokens · 111336 ms · 2026-08-03T13:16:57.336649+00:00 · methodology

0 comments
read the original abstract

Accurate traffic prediction is essential for Intelligent Transportation Systems, including ride-hailing, urban road planning, and vehicle fleet management. However, due to significant privacy concerns surrounding traffic data, most existing methods rely on local training, resulting in data silos and limited knowledge sharing. Federated Learning (FL) offers an efficient solution through privacy-preserving collaborative training; however, standard FL struggles with the non-independent and identically distributed (non-IID) problem among clients. This challenge has led to the emergence of Personalized Federated Learning (PFL) as a promising paradigm. Nevertheless, current PFL frameworks require further adaptation for traffic prediction tasks, such as specialized graph feature engineering, data processing, and network architecture design. A notable limitation of many prior studies is their reliance on hyper-parameter optimization across datasets-information that is often unavailable in real-world scenarios-thus impeding practical deployment. To address this challenge, we propose AutoFed, a novel PFL framework for traffic prediction that eliminates the need for manual hyper-parameter tuning. Inspired by prompt learning, AutoFed introduces a federated representor that employs a client-aligned adapter to distill local data into a compact, globally shared prompt matrix. This prompt then conditions a personalized predictor, allowing each client to benefit from cross-client knowledge while maintaining local specificity. Extensive experiments on real-world datasets demonstrate that AutoFed consistently achieves superior performance across diverse scenarios. The code of this paper is provided at https://github.com/RS2002/AutoFed .

Figures

Figures reproduced from arXiv: 2512.24625 by Sen Li, Yitong Shang, Zijian Zhao.

Figure 1
Figure 1. Figure 1: Network Architecture: The network consists of the PP and FR. PP utilizes a graph time series network with an encoder-decoder structure. FR employs an AE-based denoiser for robust feature extraction, a graph time series encoder for feature compression, and an client-aligned adapter for transferring local representations to global representations, providing a guided prompt matrix for the decoder in PP. In th… view at source ↗
Figure 2
Figure 2. Figure 2: Comparison Between Local Representation and Global Representation the spatial relationships among districts, resulting in mu￾tual interference due to the limited influence among them. In contrast, scenarios S1-S3 treat the districts as separate clients and leverage FL to aggregate their common temporal relationships, leading to improved performance. (ii) Second, in scenario S3, most methods experience a de… view at source ↗
Figure 3
Figure 3. Figure 3: Training Process: This figure shows the change of different metrics in valid set during training. Regarding the TFP task, our method demonstrates perfor￾mance similar to the two previous SOTA methods for traffic prediction tasks, namely FedTPS and FedGCN. However, it is important to note that both of these methods are highly dependent on hyper-parameters (such as the pattern amount and low-pass filter meth… view at source ↗

Review history (2 revisions) →

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

47 extracted references · 5 linked inside Pith

  1. [1]

    A., Piamrat, K., and Ji, Y

    Aouedi, O., Le, V . A., Piamrat, K., and Ji, Y . Deep learning on network traffic prediction: Recent ad- vances, analysis, and future directions.ACM Com- puting Surveys, 57(6):1–37, 2025

    2025

  2. [2]

    G., Aggarwal, V ., Singh, A

    Arivazhagan, M. G., Aggarwal, V ., Singh, A. K., and Choudhary, S. Federated learning with personalization layers.arXiv preprint arXiv:1912.00818, 2019

    Pith/arXiv arXiv 1912

  3. [3]

    T., Dauwels, J., Goh, C

    Asif, M. T., Dauwels, J., Goh, C. Y ., Oran, A., Fathi, E., Xu, M., Dhanya, M. M., Mitrovic, N., and Jail- let, P. Spatiotemporal patterns in large-scale traffic speed prediction.IEEE Transactions on Intelligent Transportation Systems, 15(2):794–804, 2013

    2013

  4. [4]

    Adap- tive graph convolutional recurrent network for traffic forecasting.Advances in neural information process- ing systems, 33:17804–17815, 2020

    Bai, L., Yao, L., Li, C., Wang, X., and Wang, C. Adap- tive graph convolutional recurrent network for traffic forecasting.Advances in neural information process- ing systems, 33:17804–17815, 2020

    2020

  5. [5]

    Feedsign: robust full- parameter federated fine-tuning of large models with extremely low communication overhead of one bit

    Cai, Z., Chen, H., and Zhu, G. Feedsign: robust full- parameter federated fine-tuning of large models with extremely low communication overhead of one bit. arXiv preprint arXiv:2501.17610, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    O., Pfister, T., Zheng, Y ., Ye, W., and Liu, Y

    Cao, D., Jia, F., Arik, S. O., Pfister, T., Zheng, Y ., Ye, W., and Liu, Y . Tempo: Prompt-based genera- tive pre-trained transformer for time series forecasting. InThe Twelfth International Conference on Learning Representations, 2024

    2024

  7. [7]

    Dual prompt personalized federated learning in foun- dation models.Scientific Reports, 15(1):28026, 2025

    Chang, Y ., Shi, X., Zhao, X., Chen, Z., and Ma, D. Dual prompt personalized federated learning in foun- dation models.Scientific Reports, 15(1):28026, 2025

    2025

  8. [8]

    University of California, Berkeley, 2002

    Chen, C.Freeway performance measurement system (PeMS). University of California, Berkeley, 2002

    2002

  9. [9]

    An overview of domain- specific foundation model: key technologies, appli- cations and challenges.Science China Information Sciences, 69(1):111301, 2026

    Chen, H., Chen, H., Zhao, Z., Han, K., Zhu, G., Zhao, Y ., Du, Y ., Xu, W., and Shi, Q. An overview of domain- specific foundation model: key technologies, appli- cations and challenges.Science China Information Sciences, 69(1):111301, 2026

    2026

  10. [10]

    Deri, J. A. and Moura, J. M. Taxi data in new york city: A network perspective. In2015 49th asilomar con- ference on signals, systems and computers, pp. 1829–

  11. [11]

    Person- alized federated learning: A meta-learning approach

    Fallah, A., Mokhtari, A., and Ozdaglar, A. Person- alized federated learning: A meta-learning approach. arXiv preprint arXiv:2002.07948, 2020

    Pith/arXiv arXiv 2002

  12. [12]

    When spatio-temporal meet wavelets: Disentangled traffic forecasting via efficient spectral graph attention networks

    Fang, Y ., Qin, Y ., Luo, H., Zhao, F., Xu, B., Zeng, L., and Wang, C. When spatio-temporal meet wavelets: Disentangled traffic forecasting via efficient spectral graph attention networks. In2023 IEEE 39th Interna- tional Conference on Data Engineering (ICDE), pp. 517–529. IEEE, 2023

    2023

  13. [13]

    O., and Yoon, J

    Han, S., Lee, S., Cha, M., Arik, S. O., and Yoon, J. Retrieval augmented time series forecasting. In Forty-second International Conference on Machine Learning, 2025

    2025

  14. [14]

    Hu, N., Liang, W., Zhang, D., Xie, K., Li, K., and Zomaya, A. Y . Fedgcn: A federated graph convolu- tional network for privacy-preserving traffic prediction. IEEE Transactions on Sustainable Computing, 9(6): 925–935, 2024

    2024

  15. [15]

    Timedp: Learning to generate multi-domain time series with domain prompts

    Huang, Y .-H., Xu, C., Wu, Y ., Li, W.-J., and Bian, J. Timedp: Learning to generate multi-domain time series with domain prompts. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pp. 17520–17527, 2025

    2025

  16. [16]

    Time-llm: Time series forecasting by reprogramming large language models

    Jin, M., Wang, S., Ma, L., Chu, Z., Zhang, J., Shi, X., Chen, P.-Y ., Liang, Y ., Li, Y .-f., Pan, S., et al. Time-llm: Time series forecasting by reprogramming large language models. InInternational Conference on Learning Representations, 2024

    2024

  17. [17]

    Ke, J., Zheng, H., Yang, H., and Chen, X. M. Short-term forecasting of passenger demand under on- demand ride services: A spatio-temporal deep learning approach.Transportation research part C: Emerging technologies, 85:591–608, 2017

    2017

  18. [18]

    Fast-convergent and communication-alleviated heterogeneous hierarchical federated learning in autonomous driving.IEEE Trans- actions on Intelligent Transportation Systems, 2025

    Kou, W.-B., Lin, Q., Tang, M., Ye, R., Wang, S., Zhu, G., and Wu, Y .-C. Fast-convergent and communication-alleviated heterogeneous hierarchical federated learning in autonomous driving.IEEE Trans- actions on Intelligent Transportation Systems, 2025

    2025

  19. [19]

    Lee, W.-H., Ozger, M., Challita, U., and Sung, K. W. Noise learning-based denoising autoencoder.IEEE Communications Letters, 25(9):2983–2987, 2021

    2021

  20. [20]

    Li, J., Sun, X., Li, Y ., Li, Z., Cheng, H., and Yu, J. X. Graph intelligence with large language models and prompt learning. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp. 6545–6554, 2024. 9 AutoFed: Manual-Free Federated Traffic Prediction via Personalized Prompt

    2024

  21. [21]

    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 hetero- geneous networks.Proceedings of Machine learning and systems, 2:429–450, 2020

    2020

  22. [22]

    Fedbn: Federated learning on non-iid features via local batch normalization.arXiv preprint arXiv:2102.07623, 2021

    Li, X., Jiang, M., Zhang, X., Kamp, M., and Dou, Q. Fedbn: Federated learning on non-iid features via local batch normalization.arXiv preprint arXiv:2102.07623, 2021

    Pith/arXiv arXiv 2021

  23. [23]

    A survey on prompt tuning.arXiv preprint arXiv:2507.06085, 2025

    Li, Z., Su, Y ., and Collier, N. A survey on prompt tuning.arXiv preprint arXiv:2507.06085, 2025

    Pith/arXiv arXiv 2025

  24. [24]

    Discovering spatio-temporal causal interactions in traffic data streams

    Liu, W., Zheng, Y ., Chawla, S., Yuan, J., and Xing, X. Discovering spatio-temporal causal interactions in traffic data streams. InProceedings of the 17th ACM SIGKDD international conference on Knowledge discovery and data mining, pp. 1010–1018, 2011

    2011

  25. [25]

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

    2017

  26. [26]

    H., Nguyen, V .-D., Luu, Q.-T., Gian, T

    Nguyen, X. H., Nguyen, V .-D., Luu, Q.-T., Gian, T. D., and Shin, O.-S. Robust wifi sensing-based human pose estimation using denoising autoencoder and cnn with dynamic subcarrier attention.IEEE Internet of Things Journal, 2025

    2025

  27. [27]

    A review of federated learning methods in heterogeneous sce- narios.IEEE Transactions on Consumer Electronics, 70(3):5983–5999, 2024

    Pei, J., Liu, W., Li, J., Wang, L., and Liu, C. A review of federated learning methods in heterogeneous sce- narios.IEEE Transactions on Consumer Electronics, 70(3):5983–5999, 2024

    2024

  28. [28]

    Federated learning in a medical context: a systematic literature review.ACM Transactions on Internet Technology (TOIT), 21(2):1–31, 2021

    Pfitzner, B., Steckhan, N., and Arnrich, B. Federated learning in a medical context: a systematic literature review.ACM Transactions on Internet Technology (TOIT), 21(2):1–31, 2021

    2021

  29. [29]

    Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y ., Li, W., and Liu, P. J. Ex- ploring the limits of transfer learning with a unified text-to-text transformer.Journal of machine learning research, 21(140):1–67, 2020

    2020

  30. [30]

    Scott, J., Zakerinia, H., and Lampert, C. H. Pefll: Personalized federated learning by learning to learn. InThe Twelfth International Conference on Learning Representations, 2024

    2024

  31. [31]

    and Li, S

    Shang, Y . and Li, S. Security-enhanced spatiotemporal ride-hailing demand prediction—part ii: Vertical feder- ated learning.IEEE Transactions on Intelligent Trans- portation Systems, 26(9):13363–13377, 2025. doi: 10.1109/TITS.2025.3564629

    arXiv 2025

  32. [32]

    and Li, S

    Shang, Y . and Li, S. Security-enhanced spatiotemporal ride-hailing demand prediction—part i: Horizontal federated learning.IEEE Transactions on Intelligent Transportation Systems, 2025

    2025

  33. [33]

    Explainable spa- tiotemporal multi-task learning for electric vehicle charging demand prediction.Applied Energy, 384: 125460, 2025

    Shang, Y ., Li, D., Li, Y ., and Li, S. Explainable spa- tiotemporal multi-task learning for electric vehicle charging demand prediction.Applied Energy, 384: 125460, 2025

    2025

  34. [34]

    Personalized federated learning with moreau envelopes.Advances in neural information processing systems, 33:21394– 21405, 2020

    T Dinh, C., Tran, N., and Nguyen, J. Personalized federated learning with moreau envelopes.Advances in neural information processing systems, 33:21394– 21405, 2020

    2020

  35. [35]

    Z., Yu, H., Cui, L., and Yang, Q

    Tan, A. Z., Yu, H., Cui, L., and Yang, Q. Towards personalized federated learning.IEEE transactions on neural networks and learning systems, 34(12):9587– 9603, 2022

    2022

  36. [36]

    A., Bao, Z., Zheng, B., Choudhury, F

    Tedjopurnomo, D. A., Bao, Z., Zheng, B., Choudhury, F. M., and Qin, A. K. A survey on modern deep neural network for traffic prediction: Trends, methods and challenges.IEEE Transactions on Knowledge and Data Engineering, 34(4):1544–1561, 2020

    2020

  37. [37]

    and Salim, F

    Xue, H. and Salim, F. D. Promptcast: A new prompt- based learning paradigm for time series forecasting. IEEE Transactions on Knowledge and Data Engineer- ing, 36(11):6851–6864, 2023

    2023

  38. [38]

    Traffic flow prediction using lstm with feature enhancement

    Yang, B., Sun, S., Li, J., Lin, X., and Tian, Y . Traffic flow prediction using lstm with feature enhancement. Neurocomputing, 332:320–327, 2019

    2019

  39. [39]

    Yuan, X., Chen, J., Yang, J., Zhang, N., Yang, T., Han, T., and Taherkordi, A. Fedstn: Graph representation driven federated learning for edge computing enabled urban traffic flow prediction.IEEE Transactions on Intelligent Transportation Systems, 24(8):8738–8748, 2022

    2022

  40. [40]

    A survey on federated learning.Knowledge-Based Systems, 216:106775, 2021

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

    2021

  41. [41]

    An improved k-nearest neighbor model for short-term traffic flow prediction.Procedia-Social and Behav- ioral Sciences, 96:653–662, 2013

    Zhang, L., Liu, Q., Yang, W., Wei, N., and Dong, D. An improved k-nearest neighbor model for short-term traffic flow prediction.Procedia-Social and Behav- ioral Sciences, 96:653–662, 2013

    2013

  42. [42]

    A survey on federated learning in intelli- gent transportation systems.IEEE Transactions on Intelligent Vehicles, 2024

    Zhang, R., Mao, J., Wang, H., Li, B., Cheng, X., and Yang, L. A survey on federated learning in intelli- gent transportation systems.IEEE Transactions on Intelligent Vehicles, 2024

    2024

  43. [43]

    Zhao, Z. Let network decide what to learn: Sym- bolic music understanding model based on large-scale 10 AutoFed: Manual-Free Federated Traffic Prediction via Personalized Prompt adversarial pre-training. InProceedings of the 2025 International Conference on Multimedia Retrieval, pp. 2128–2132, 2025

    2025

  44. [44]

    Csi-bert2: A bert-inspired framework for efficient csi prediction and classification in wireless commu- nication and sensing.IEEE Transactions on Mobile Computing, pp

    Zhao, Z., Meng, F., Lyu, Z., Li, H., Li, X., and Zhu, G. Csi-bert2: A bert-inspired framework for efficient csi prediction and classification in wireless commu- nication and sensing.IEEE Transactions on Mobile Computing, pp. 1–17, 2025. doi: 10.1109/TMC.2025. 3640420

    doi:10.1109/tmc.2025 2025

  45. [45]

    Traffic pattern sharing for federated traffic flow prediction with personalization

    Zhou, H., Yu, W., Wan, S., Tong, Y ., Gu, T., and Gong, C. Traffic pattern sharing for federated traffic flow prediction with personalization. In2024 IEEE International Conference on Data Mining (ICDM), pp. 639–648. IEEE, 2024

    2024

  46. [46]

    Fedtps: traffic pattern sharing for personal- ized federated traffic flow prediction.Knowledge and Information Systems, pp

    Zhou, H., Yu, W., Wan, S., Tong, Y ., Gu, T., and Gong, C. Fedtps: traffic pattern sharing for personal- ized federated traffic flow prediction.Knowledge and Information Systems, pp. 1–27, 2025

    2025

  47. [47]

    A comprehensive survey on transfer learning.Proceedings of the IEEE, 109(1): 43–76, 2020

    Zhuang, F., Qi, Z., Duan, K., Xi, D., Zhu, Y ., Zhu, H., Xiong, H., and He, Q. A comprehensive survey on transfer learning.Proceedings of the IEEE, 109(1): 43–76, 2020. Algorithm 1Training Process of AutoFed Require: Number of clients n, local datasets {Di}n i=1, shared model parameters Θ0, personal model parame- ters {θ0 i }n i=1, communication rounds M,...

    2020