Pith. sign in

REVIEW 1 major objections 2 minor 29 references

MoE Enhanced Federated Learning for Spatiotemporal Prediction

T0 review · 1 major / 2 minor · reviewed 2026-06-27 · grok-4.3

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

Pith's one-line read A mixture-of-experts federated model improves traffic predictions for cities with scarce data by dynamically fusing source-city experts.

desk verdict MoE-FedTP layers a standard MoE personalization trick onto federated spatiotemporal models and reports gains on four traffic datasets, but the support for the gains is thin. read the letter →

arxiv 2606.10499 v1 pith:7FSPFADV submitted 2026-06-09 cs.LG cs.AI

Zhehao Dai , Xiao Han , Zhaolin Deng , Zijian Zhang , Xiangyu Zhao , Guojiang Shen , Xiangjie Kong This is my paper · ORCID
classification cs.LGcs.AI
keywords trafficpredictionfederatedlearningmixtureofexpertscross-citytransferspatiotemporalmodelingprivacypreservationurbancomputing
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

The paper establishes that a federated framework can transfer traffic knowledge across cities while respecting privacy by deriving expert networks from multiple source cities through partial parameter sharing and combining them with a learned gating network. This structure lets the target-city model selectively draw on diverse spatiotemporal patterns instead of averaging all sources equally. The resulting predictions are more accurate than those from prior federated or cross-city baselines on real traffic datasets. If the claim holds, data-scarce cities gain usable forecasts without needing to collect or expose large amounts of local sensor data.

What carries the argument

Mixture-of-experts architecture in which each expert is a partially shared network from one source city and the gating network learns to route or combine their outputs for a given target city.

What would settle it

MoE-FedTP would be falsified if it stopped outperforming the baselines on a fresh collection of cities whose traffic patterns differ in structure from the four original datasets.

Watch

Extended reading notes

Core claim

MoE-FedTP first runs spatiotemporal neural networks on source and target city data to extract features, then creates a set of expert networks from different source cities via partial parameter sharing and trains a gating mechanism that dynamically weights and fuses the experts to model city-specific traffic dynamics.

Load-bearing premise

Partial parameter sharing from a fixed collection of source-city experts plus a gating network will be enough to represent the range of traffic heterogeneities in new target cities.

Editorial extensions

If this is right

  • Target cities obtain higher prediction accuracy than with standard federated averaging or single-source transfer.
  • Only model updates are exchanged, so raw traffic observations remain local.
  • The gating network produces city-specific weightings that reflect distinct spatiotemporal regimes.
  • Communication cost stays lower than full-model sharing because only partial parameters move between cities.

Reading between the lines

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

  • The same partial-sharing-plus-gating pattern could be tested on other privacy-sensitive spatiotemporal tasks such as energy load or air-quality forecasting.
  • Increasing the number of source experts might extend coverage to cities whose patterns lie outside the current set.
  • If the gating weights turn out to cluster cities by similarity, the framework could also support rapid expert selection for entirely new cities.
Share X Bluesky LinkedIn Reddit HN

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

1 major / 2 minor

Summary. The manuscript proposes MoE-FedTP, a personalized federated cross-city spatiotemporal prediction framework that extracts features via spatiotemporal neural networks, derives expert networks from source cities via partial parameter sharing, and uses a learned gating mechanism to dynamically fuse experts for modeling urban heterogeneity while preserving privacy. The central empirical claim is that MoE-FedTP consistently outperforms state-of-the-art cross-city and federated learning baselines on four real-world traffic datasets.

Significance. If the outperformance claim holds with rigorous validation, the work would offer a practical advance in privacy-preserving knowledge transfer for data-scarce cities in intelligent transportation systems, extending standard MoE personalization patterns to spatiotemporal federated settings.

major comments (1)
  1. [Abstract] Abstract: the claim of consistent outperformance on four datasets is presented without any quantitative metrics, error values, statistical tests, ablation results, or error analysis; this absence directly undermines assessment of the central empirical claim and its support for the design premise of partial sharing plus gating.
minor comments (2)
  1. Notation for the gating network and expert fusion is introduced without an explicit equation or diagram reference, making the adaptation mechanism harder to follow.
  2. The description of 'lightweight' MoE networks lacks a parameter count comparison to baselines, which would clarify the efficiency claim.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the detailed feedback. We address the comment on the abstract below and will incorporate the suggested changes in the revised manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim of consistent outperformance on four datasets is presented without any quantitative metrics, error values, statistical tests, ablation results, or error analysis; this absence directly undermines assessment of the central empirical claim and its support for the design premise of partial sharing plus gating.

    Authors: We agree that the abstract would be strengthened by the inclusion of quantitative metrics to support the outperformance claim. In the revised manuscript we will add concise performance highlights (e.g., average relative improvements in MAE and RMSE across the four datasets) while keeping the abstract within length limits. The full quantitative results, statistical significance tests, ablation studies, and error analyses are already reported in Sections 4 and 5; the abstract revision will simply provide readers with an immediate quantitative anchor for the central claim. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity; empirical framework evaluated on external real-world datasets

full rationale

The paper proposes MoE-FedTP as a personalized federated learning architecture using spatiotemporal networks, source-city experts via partial parameter sharing, and a learned gating mechanism. Its central claim is empirical outperformance on four real-world traffic datasets against baselines. No derivation chain, equations, or predictions are presented that reduce by construction to fitted parameters or self-citations within the paper; the architecture is a standard MoE personalization pattern whose effectiveness is precisely what the external-dataset experiments are designed to test. The work is self-contained against external benchmarks with no load-bearing internal reductions.

Assumptions & free parameters 0 free parameters · 0 assumptions · 0 invented entities

Abstract-only review; no explicit free parameters, axioms, or invented entities are stated or derivable from the provided text.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MoE Enhanced Federated Learning for Spatiotemporal Prediction." pith.science (2026). https://pith.science/paper/7FSPFADV

@misc{pith2026260610499,
  author       = {Pith},
  title        = {Pith review of: MoE Enhanced Federated Learning for Spatiotemporal Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7FSPFADV}},
  note         = {Machine review of arXiv:2606.10499}
}
read the original abstract

Traffic prediction is fundamental to intelligent transportation systems and urban computing, yet many cities continue to suffer from traffic data scarcity due to limited sensor deployment and uneven urban development. Cross-city knowledge transfer has thus attracted increasing attention, enabling data-rich cities to assist data-scarce ones. However, centralized approaches raise privacy concerns, while existing federated methods struggle with pronounced spatiotemporal heterogeneity across cities. To address these challenges, we propose MoE-FedTP, a personalized federated cross-city spatiotemporal prediction framework based on lightweight Mixture-of-Experts (MoE) networks. MoE-FedTP first employs spatiotemporal neural networks to extract features from both source and target cities, then introduces a set of expert networks derived from different source cities through partial parameter sharing. A gating mechanism dynamically fuses the experts to capture diverse traffic dynamics, achieving fine-grained modeling of urban heterogeneity while preserving privacy. Experiments on four real-world traffic datasets show that MoE-FedTP consistently outperforms state-of-the-art cross-city and federated learning baselines, demonstrating its effectiveness in enhancing prediction accuracy for data-scarce cities.

Figures

Figures reproduced from arXiv: 2606.10499 by the authors.

Figure 1
Figure 1. The framework overview of MoE-FedsT. It leverages a Spatialtemporal Net to process local spatiotemporal data and [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Ablation study of MoE-FedST. performance in both long-term and short-term traffic prediction tasks. Specifically, on the METR-LA dataset, the MAE and RMSE metrics improved by 6.23% and 5.20%, respectively; while on the DiDi-Shenzhen dataset, an average performance gain of 4.68% was achieved. These results further validate the strong generalization ca￾pability of our method, making it effectively applicable to differ… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 2 canonical work pages

  1. [1]

    Lingxiao Cao, Bin Wang, Guiyuan Jiang, Yanwei Yu, and Junyu Dong. 2025. Spatiotemporal-aware Trend-Seasonality Decomposition Network for Traffic Flow Forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 39. 11463–11471

    2025

  2. [2]

    Junyoung Chung, Caglar Gulcehre, Kyunghyun Cho, and Yoshua Bengio. 2014. Empirical evaluation of gated recurrent neural networks on sequence modeling. In NIPS 2014 Workshop on Deep Learning, December 2014

    2014

  3. [3]

    Xiao Han, Guojiang Shen, Xi Yang, and Xiangjie Kong. 2020. Congestion recog- nition for hybrid urban road systems via digraph convolutional network. Trans- portation Research Part C: Emerging Technologies 121 (2020), 102877

    2020

  4. [4]

    Xiao Han, Zijian Zhang, Xiangyu Zhao, Yuanshao Zhu, Guojiang Shen, Xiangjie Kong, Xuetao Wei, Liqiang Nie, and Jieping Ye. 2025. Garlic: Gpt-augmented rein- forcement learning with intelligent control for vehicle dispatching. InProceedings of the AAAI Conference on Artificial Intelligence , Vol. 39. 255–263

    2025

  5. [5]

    Xiao Han, Xiangyu Zhao, Liang Zhang, and Wanyu Wang. 2023. Mitigating action hysteresis in traffic signal control with traffic predictive reinforcement learning. In Proceedings of the 29th ACM SIGKDD conference on knowledge discovery and data mining. 673–684

    2023

  6. [6]

    Xiao Han, Ding-Xuan Zhou, Guojiang Shen, Xiangjie Kong, and Yulong Zhao

  7. [7]

    IEEE Transactions on Vehicular Technology 73, 11 (2024), 16051–16062

    Deep trajectory recovery approach of offline vehicles in the internet of vehicles. IEEE Transactions on Vehicular Technology 73, 11 (2024), 16051–16062

    2024

  8. [8]

    Junfeng Hu, Xu Liu, Zhencheng Fan, Yifang Yin, Shili Xiang, Savitha Ramasamy, and Roger Zimmermann. 2024. Prompt-based spatio-temporal graph transfer learning. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management. 890–899

    2024

Show all 29 references
  1. [9]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. 2017. Semi-Supervised Classification with Graph Convolutional Networks. In Proc. of ICLR

    2017

  2. [10]

    Zhenyu Lei, Yushun Dong, Jundong Li, and Chen Chen. 2025. ST-FiT: Inductive Spatial-Temporal Forecasting with Limited Training Data. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 12031–12039

    2025

  3. [11]

    Yaguang Li, Rose Yu, Cyrus Shahabi, and Yan Liu. 2018. Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting. In International Conference on Learning Representations

    2018

  4. [12]

    Duanyang Liu, Longfeng Tang, Guojiang Shen, and Xiao Han. 2019. Traffic speed prediction: An attention-based method. Sensors 19, 18 (2019), 3836

    2019

  5. [13]

    Xu Liu, Yutong Xia, Yuxuan Liang, Junfeng Hu, Yiwei Wang, Lei Bai, Chao Huang, Zhenguang Liu, Bryan Hooi, and Roger Zimmermann. 2023. Largest: A benchmark dataset for large-scale traffic forecasting. Advances in Neural Information Processing Systems 36 (2023), 75354–75371

    2023

  6. [14]

    Bin Lu, Xiaoying Gan, Weinan Zhang, Huaxiu Yao, Luoyi Fu, and Xinbing Wang

  7. [15]

    In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining

    Spatio-temporal graph few-shot learning with cross-city knowledge trans- fer. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 1162–1172

  8. [16]

    Chuizheng Meng, Sirisha Rambhatla, and Yan Liu. 2021. Cross-node federated graph neural network for spatio-temporal data modeling. In Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining . 1202–1211

    2021

  9. [17]

    Veličković Petar, Cucurull Guillem, Casanova Arantxa, Romero Adriana, Lio Pietro, and B Yoshua. 2018. Graph attention networks. In Proc. of ICLR

    2018

  10. [18]

    Weilin Ruan, Wenzhuo Wang, Siru Zhong, Wei Chen, Li Liu, and Yuxuan Liang

  11. [19]

    arXiv preprint arXiv:2411.09251 (2024)

    Cross space and time: A spatio-temporal unitized model for traffic flow forecasting. arXiv preprint arXiv:2411.09251 (2024)

    2024

  12. [20]

    Guojiang Shen, Xiao Han, KwaiSang Chin, and Xiangjie Kong. 2021. An attention- based digraph convolution network enabled framework for congestion recog- nition in three-dimensional road networks. IEEE Transactions on Intelligent Transportation Systems 23, 9 (2021), 14413–14426

    2021

  13. [21]

    Yihong Tang, Ao Qu, Andy HF Chow, William HK Lam, Sze Chun Wong, and Wei Ma. 2022. Domain adversarial spatial-temporal network: A transferable framework for short-term traffic forecasting across cities. In Proceedings of the 31st ACM international conference on information & k...

    2022

  14. [22]

    Guangyu Wang, Yujie Chen, Ming Gao, Zhiqiao Wu, Jiafu Tang, and Jiabi Zhao

  15. [23]

    arXiv preprint arXiv:2409.17440 (2024)

    A time series is worth five experts: Heterogeneous mixture of experts for traffic flow prediction. arXiv preprint arXiv:2409.17440 (2024)

    2024

  16. [24]

    Leye Wang, Xu Geng, Xiaojuan Ma, Feng Liu, and Qiang Yang. 2019. Cross-City Transfer Learning for Deep Spatio-Temporal Prediction. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence . Interna- tional Joint Conferences on Artificial In...

    2019

  17. [25]

    Bing Yu, Haoteng Yin, and Zhanxing Zhu. 2018. Spatio-temporal graph convolu- tional networks: a deep learning framework for traffic forecasting. InProceedings of the 27th International Joint Conference on Artificial Intelligence . 3634–3640

    2018

  18. [26]

    Yu Zhang, Hua Lu, Ning Liu, Yonghui Xu, Qingzhong Li, and Lizhen Cui. 2024. Personalized federated learning for cross-city traffic prediction. In 33rd Interna- tional Joint Conference on Artificial Intelligence, IJCAI . 5526–5534

    2024

  19. [27]

    Yan Zhang, Xiaoye Miao, Bin Li, Yangyang Wu, and Yongheng Shang. 2025. Proxy- Validated Importance-Aware Federated Sample Selection with Meta Learning. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 3855–3866

    2025

  20. [28]

    Yan Zhang, Guojiang Shen, Xiao Han, Wei Wang, and Xiangjie Kong. 2022. Spatio-temporal digraph convolutional network-based taxi pickup location rec- ommendation. IEEE Transactions on Industrial Informatics 19, 1 (2022), 394–403

    2022

  21. [29]

    Yudong Zhang, Xu Wang, Pengkun Wang, Binwu Wang, Zhengyang Zhou, and Yang Wang. 2024. Modeling spatio-temporal mobility across data silos via per- sonalized federated learning. IEEE Transactions on Mobile Computing (2024)

    2024

Pith tools

Reviewed June 27, 2026 · model on record in the stance chip above.