Pith. sign in

REVIEW 3 major objections 5 minor 40 references

HSTMixer, an all-MLP model with hierarchical spatiotemporal mixing, claims state-of-the-art results on four large-scale traffic datasets while keeping computational cost linear in the number of sensors.

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 20:05 UTC pith:PN6YFDFG

load-bearing objection Solid architectural extension of the MLP-mixer family with a plausible linear-complexity SOTA claim, but the evidence as shipped — no code, no variance, no baseline tuning details — is not yet strong enough to take the exact margins on faith. the 3 major comments →

arxiv 2512.07854 v2 pith:PN6YFDFG submitted 2025-11-26 cs.LG

HieraMix: A Hierarchical MLP-Mixer for Large-Scale Traffic Forecasting

Yongyao Wang , Xie Yu , Jingyuan Wang , Jiahao Ji , Chao Li This is my paper
classification cs.LG
keywords traffic forecastingMLP-Mixerlarge-scale spatiotemporal datahierarchical feature pyramidall-MLP architecturelinear complexityadaptive region mixingmulti-resolution feature learning
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 a purely linear architecture — no attention, no graph convolution — can beat both quadratic-complexity transformers and existing linear models on large-scale traffic forecasting, provided it respects the hierarchical structure of traffic data. The proposed model, HSTMixer, builds feature pyramids by aggregating nearby sensors into regions of growing size, then propagates regional context back down to individual sensors. The authors report consistent state-of-the-art results across four large datasets with 2–4% lower average error than the best previous baselines, at a cost that scales linearly with the number of sensors rather than quadratically. If correct, it shows that scale itself is not a reason to reach for heavier architectures; the right inductive bias can make a simple MLP the practical choice for city-wide forecasting.

Core claim

The central claim is that spatiotemporal traffic data possesses a hierarchy — fine-scale local similarity and coarse-scale regional semantics — and that a model which explicitly extracts features at each resolution, then merges them top-down, captures the signal that global mixing dilutes in large networks. Concretely, HSTMixer processes time windows of increasing size across stacked blocks, and in space progressively aggregates nodes into region levels via learned mixers; an adaptive region mixer generates each region's transformation weights from a shared parameter pool, with an orthogonality penalty that keeps distinct regions distinct. On four large-scale datasets (716 to 8,600 sensors,

What carries the argument

The key machinery is the spatiotemporal mixing block, which combines two feature pyramids. Bottom-up, a temporal aggregation mixer folds short windows into progressively coarser time steps, and a spatial cascade mixer aggregates nodes through K region scales. Top-down, spatial and temporal propagation paths re-inject coarse context into finer levels, so the final representation carries both local and regional information. The adaptive region mixer is the distinctive component: it generates per-region linear transformations from a small pool of base weights selected by a softmax over region semantics, and a parameter orthogonal loss pushes different regions' weights apart. This is what lets a

Load-bearing premise

The load-bearing premise is that the 20 baselines were all evaluated under exactly the same data preprocessing (5-minute readings aggregated to 15 minutes, seasonal 6:2:2 splits) and that the grid search for hyperparameters was confined to validation data; the paper reports no code, no per-run variance, and no explicit statement limiting hyperparameter selection to the validation split.

What would settle it

Rerun any strong baseline (for example, a random-projection mixer or a lightweight spatiotemporal network) on the same 15-minute aggregated data with the same seasonal splits, tuning hyperparameters purely on the validation set, and check whether the reported 2–4% gaps reproduce. If the gaps shrink to noise or reverse, the state-of-the-art claim fails; if baseline numbers came from different preprocessing, the comparison is invalid. Also, computing the standard deviation across HSTMixer's five runs would reveal whether the reported margins exceed run-to-run noise, since the paper omits those v

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

If this is right

  • Large-scale traffic forecasting does not require quadratic-complexity attention or graph convolutions; a linear model with hierarchical mixing is sufficient and outperforms both.
  • The reported performance gap between HSTMixer and the second-best baseline widens on the largest dataset (CA), suggesting the hierarchy advantage becomes more pronounced as the network scales.
  • Training time drops by an order of magnitude compared to graph-based models: on the 2,352-node dataset, HSTMixer trains in about 4.5 hours versus roughly 59 hours for a diffusion GNN baseline.
  • The full training process fits on a single 48GB GPU, implying the reported accuracy can be reproduced without distributed infrastructure.

Where Pith is reading between the lines

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

  • If the hierarchy hypothesis holds, the same bottom-up/top-down mixing should transfer to other spatiotemporal forecasting tasks with natural coarse-graining, such as energy demand or crowd flow, where regional semantics are similarly stable.
  • The adaptive region mixer's learned parameter pool could be used as an interpretability tool: the softmax assignments define functional regions in a purely data-driven way, without geographic priors, which may be useful for transportation planning.
  • Because the paper aggregates 5-minute readings to 15 minutes before evaluation, the benefit of the temporal pyramid at finer granularities is untested; a natural stress test is to run the same architecture on the native 5-minute data and check whether the hierarchy still dominates.
  • The smallest dataset (SD) shows a smaller improvement margin, so the practical advantage may concentrate on larger networks; agencies with mid-size networks should re-run the comparison rather than assume the average gain applies.

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

3 major / 5 minor

Summary. The manuscript presents HSTMixer (also called HieraMix in the title/abstract), an all-MLP spatio-temporal hierarchy for large-scale traffic forecasting. It stacks spatiotemporal mixing blocks, each containing a temporal aggregation mixer and a spatial cascade mixer; the latter produces multi-resolution region features and uses adaptive region mixers whose weights are generated from a key-value pool, plus an orthogonal loss over region weights. A top-down propagation path merges the multi-scale features. The empirical section reports MAE/RMSE/MAPE on four LargeST datasets (SD, GBA, GLA, CA) and claims average improvements of 4.41%, 3.15%, and 2.03% over the best baselines, together with linear complexity in the number of nodes. The paper also reports ablations, sensitivity, training time, and t-SNE visualizations.

Significance. The architecture is a reasonable extension of the MLP-Mixer line: the hierarchical aggregation and the adaptive region mixer are simple, plausible mechanisms for large graphs, and the complexity derivation in §5.4 indicates linear scaling with N. If the empirical claims are reproducible, the method would be a useful new state of the art for large-scale traffic forecasting. I verified that the reported average improvements are internally consistent with the values in Table 2. However, the paper provides no code, no per-seed variance or statistical tests, no numeric ablation values, and no complete baseline-tuning documentation; these are essential to support the 'consistently outperforms' claim. The contribution is therefore promising but not yet fully evidenced.

major comments (3)
  1. [§5.2, Table 2] The manuscript states that all models were run five times and mean results are reported, but only the means appear. Several wins over the second-best baseline are small (e.g., GLA MAPE 9.53 vs 9.65; GBA MAPE 12.71 vs 12.99). Without standard deviations, per-run values, or significance tests, the claim of consistent outperformance is not established. Please provide variance measures and a paired test across seeds, or at least report all five runs.
  2. [§5.1, Appendix A.1] Baseline fairness is the load-bearing measurement premise. The text says baseline details are in Appendix A.1, but the appendix lists only model names; no per-baseline hyperparameters, no statement that the same 15-minute aggregation and seasonal 6:2:2 splits were used for every model, and no confirmation that grid search was restricted to validation data. Since the SOTA claim depends on this parity, please provide the full evaluation protocol, baseline configurations, and validation/test separation, or release code and logs.
  3. [§5.3, Figure 5] Figure 5 is referenced as the ablation study but contains no quantitative information; the surrounding text is a placeholder. The five ablations (w/o AM, w/o TH, w/o SH, w/o TP, w/o SP) are central to the claim that each component contributes. Please replace the placeholder with a table or figure reporting metric values for each variant on each dataset, with variances.
minor comments (5)
  1. [Title/Abstract] The title and abstract use 'HieraMix' while the body and methodology use 'HSTMixer'. Please unify the nomenclature.
  2. [§5.1, Appendix A.1] The text says 20 baselines are compared, but Table 2 and Appendix A.1 list 19. Please correct the count and ensure the category totals match.
  3. [Eq. (7)] The tensor multiplication H_{l,k} * K_{l,k}^T is not fully specified; the axes are unclear. Define the einsum or dimension alignment. Also, K is used both for the number of region scales and for the key matrix, which is confusing.
  4. [§4.2, Spatial Cascade Mixer] The spatial aggregation path is described as 'K FC layers', but the equations for how H_{l,k} is derived from H_l are not given. Please provide the aggregation formula.
  5. [Table 5] The first column header appears to be batch size but is not labeled. Also, the split between 'Training' and 'Infer' columns is not visually explicit; consider making the headers clearer.

Circularity Check

0 steps flagged

No circularity found: the SOTA claim is an empirical benchmarking result on external LargeST data, and the architecture equations do not reduce to their own outputs.

full rationale

The paper's central claim is an empirical SOTA result on the external LargeST benchmark (SD/GBA/GLA/CA). No equation derives a prediction from a fitted input of that same prediction: the hierarchical mixer (Eqs. 5-13) maps input series to forecasts via standard MLP/attention-like mixing, with learned keys/values in Eq. 7 constituting a self-conditioning mechanism rather than a fitted parameter renamed as a prediction. The benchmark numbers in Table 2 are external comparisons, not quantities computed from the model's own fitted values; the reported 4.41%/3.15%/2.03% improvements are arithmetic summaries of Table 2, not circular. Ablations, the O(LKMdhNT) complexity statement, and sensitivity analyses are self-contained analyses of the proposed components. While the paper cites several works by its own authors (e.g., LibCity [Wang et al., 2021]), these citations support tooling and related work and are not load-bearing for the architecture's validity; no uniqueness theorem or ansatz is imported from the authors' prior work. Benchmarking concerns such as unverified baseline tuning or lack of released code are correctness/reproducibility risks, not circularity. Therefore no circular step can be quoted, and the score is 0.

Axiom & Free-Parameter Ledger

7 free parameters · 4 axioms · 0 invented entities

No novel physical entities are introduced. The model has many tunable hyperparameters, several tuned separately per dataset; alpha/beta values are never reported. The main domain axioms are the hierarchy assumption and the use of Node2Vec embeddings as regional semantics.

free parameters (7)
  • temporal window size p = 2
    Searched over {2,3,4}; sensitivity analysis shows best at p=2; controls temporal hierarchy depth.
  • number of ST mixing blocks L = 4
    Searched over {2,3,4}; limited by input length 12 and p=2.
  • number of region scales K = 1 (SD), 2 (GBA/GLA/CA)
    Tuned per dataset from {1,2,3}; larger K degrades on smaller SD.
  • region counts S_k = [128] SD; [256,32] GBA/CA; [128,16] GLA
    Searched from {16,32,64,128,256} per scale; dataset-specific.
  • base weight sizes M_k = [32] SD; [32,4] GBA/CA; [32,2] GLA
    Searched from {2,4,8,16,32}; M1=32 peak, M1=64 degrades.
  • embedding dimensions d and h = 64 and 128
    Fixed after grid search over {32,64,128} and {64,128,256}.
  • loss weights alpha and beta = not reported
    Equation (14) defines L = alpha*L_REG + beta*L_POL, but the paper never reports alpha/beta values or their tuning.
axioms (4)
  • domain assumption Traffic data at different resolutions is dominated by different patterns: local similarity at micro level, periodicity/trend and regional semantics at macro level.
    Section 1 and Figure 2 motivate the hierarchy; if false, bottom-up aggregation may lose rather than filter noise.
  • domain assumption Node2Vec-derived static node embeddings capture the semantic regional structure needed by the adaptive region mixer.
    Section 4.1 sets E_sp = E_static + E_dynamic; no evidence that topology-based static vectors correspond to traffic-semantic regions.
  • domain assumption MLP linear mixing along spatial/temporal/feature dimensions is sufficient to model spatiotemporal dependencies after hierarchical aggregation.
    Borrowed from MLP-Mixer/TSMixer/RPMixer; the paper provides no theoretical guarantee.
  • domain assumption Parameter orthogonal loss improves semantic distinctiveness and therefore prediction accuracy.
    Equations (9)-(10); ablation w/o AM supports it empirically for this architecture, but the mechanism is not separately measured.

pith-pipeline@v1.3.0-alltime-deepseek · 17134 in / 14998 out tokens · 159428 ms · 2026-08-03T20:05:04.453998+00:00 · methodology

0 comments
read the original abstract

Traffic forecasting task is significant to modern urban management. Recently, there is growing attention on large-scale forecasting, as it better reflects the complexity of real-world traffic networks. However, existing models often exhibit quadratic computational complexity, making them impractical for large-scale real-world scenarios. In this paper, we propose a novel framework, Spatio-Temporal Hierarchical Mixer (HieraMix), which leverages an all-MLP architecture for efficient and effective large-scale traffic forecasting. HieraMix employs a hierarchical spatiotemporal mixing block to extract multi-resolution features through bottom-up aggregation and top-down propagation. Furthermore, an adaptive region mixer generates transformation matrices based on regional semantics, enabling our model to dynamically capture evolving spatiotemporal patterns for different regions. Extensive experiments conducted on four large-scale real-world datasets demonstrate that the proposed method not only achieves state-of-the-art performance but also exhibits competitive computational efficiency.

Figures

Figures reproduced from arXiv: 2512.07854 by Chao Li, Jiahao Ji, Jingyuan Wang, Xie Yu, Yongyao Wang.

Figure 1
Figure 1. Figure 1: Attention performance on graphs of different scales. In [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Spatiotemporal Hierarchy. At the macro-level, temporal [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The overall framework of HSTMixer. node n, x h n denotes traffic data at the h-th time slice. For a traffic network with N nodes, Xh ∈ R N denotes over￾all traffic data at the h-th time slice. Further, X t−T +1 : t denotes a traffic data series from t − T + 1 to t, where X t−T +1 : t = [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Structure of the Adaptive Region Mixer, where [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Results of ablation study. Verified the effectiveness of each [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: MAE and the total training time (Unit: hours) of each [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: t-SNE visualization of embedding OL,1 of HSTMixer and w/o AM on GBA. weakening the model’s ability to capture hierarchical tempo￾ral dependencies, leading to degraded performance. Further￾more, we conducted sensitivity experiments on the depth of ST mixing blocks L and the number of region scales K, with the detailed results presented in Appendix. A.2. 5.6 Visualization To evaluate HSTMixer’s ability to ca… view at source ↗
Figure 9
Figure 9. Figure 9: Sensitivity experiment results for L and K. Parameter Sensitivity. We further analyze the impact of hierarchy depth, as shown in [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗

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

40 extracted references · 5 linked inside Pith

  1. [1]

    Adaptive graph convolutional recurrent network for traffic forecasting.Advances in neural infor- mation processing systems, 33:17804–17815,

    [Baiet al., 2020 ] Lei Bai, Lina Yao, Can Li, Xianzhi Wang, and Can Wang. Adaptive graph convolutional recurrent network for traffic forecasting.Advances in neural infor- mation processing systems, 33:17804–17815,

    2020

  2. [5]

    Against other scalable models, HSTMixer’s time consumption is highly competitive, posi- tioning it in the top-tier of efficiency

    Compared to traditional GNN-based and transformer-based models, our method achieves a dramatic reduction in both training time and inference latency, which validates its excellent scalabil- ity for large-scale scenarios. Against other scalable models, HSTMixer’s time consumption is highly competitive, posi- tioning it in the top-tier of efficiency. While ...

    2066

  3. [9]

    [Fuscoet al., 2023 ] Francesco Fusco, Damian Pascual, Peter W. J. Staar, and Diego Antognini. pnlp-mixer: an efficient all-mlp architecture for language. InACL (industry), pages 53–60. Association for Computational Linguistics,

    2023

  4. [10]

    node2vec: Scalable feature learning for networks

    [Grover and Leskovec, 2016] Aditya Grover and Jure Leskovec. node2vec: Scalable feature learning for networks. InKDD, pages 855–864. ACM,

    2016

  5. [14]

    Interpretable spatiotemporal deep learning model for traffic flow prediction based on poten- tial energy fields

    [Jiet al., 2020 ] Jiahao Ji, Jingyuan Wang, Zhe Jiang, Jing- tian Ma, and Hu Zhang. Interpretable spatiotemporal deep learning model for traffic flow prediction based on poten- tial energy fields. InICDM,

    2020

  6. [15]

    Stden: Towards physics-guided neu- ral networks for traffic flow prediction

    [Jiet al., 2022 ] Jiahao Ji, Jingyuan Wang, Zhe Jiang, Jiawei Jiang, and Hu Zhang. Stden: Towards physics-guided neu- ral networks for traffic flow prediction. InAAAI,

    2022

  7. [16]

    Multi-factor spatio-temporal prediction based on graph decomposition learning.arXiv preprint arXiv:2310.10374,

    [Jiet al., 2023b ] Jiahao Ji, Jingyuan Wang, Yu Mou, and Cheng Long. Multi-factor spatio-temporal prediction based on graph decomposition learning.arXiv preprint arXiv:2310.10374,

    Pith/arXiv arXiv

  8. [17]

    Seeing the unseen: Learning basis con- founder representations for robust traffic prediction.arXiv Preprint,

    [Jiet al., 2025 ] Jiahao Ji, Wentao Zhang, Jingyuan Wang, and Chao Huang. Seeing the unseen: Learning basis con- founder representations for robust traffic prediction.arXiv Preprint,

    2025

  9. [18]

    Graphsparsenet: a novel method for large scale traffic flow prediction.arXiv preprint arXiv:2502.19823,

    [Konget al., 2025 ] Weiyang Kong, Kaiqi Wu, Sen Zhang, and Yubao Liu. Graphsparsenet: a novel method for large scale traffic flow prediction.arXiv preprint arXiv:2502.19823,

    Pith/arXiv arXiv 2025

  10. [19]

    Short-term traffic flow prediction using seasonal arima model with limited input data

    [Kumar and Vanajakshi, 2015] S Vasantha Kumar and Lelitha Vanajakshi. Short-term traffic flow prediction using seasonal arima model with limited input data. European Transport Research Review, 7:1–9,

    2015

  11. [22]

    Diffusion convolutional recurrent neural net- work: Data-driven traffic forecasting

    [Liet al., 2018 ] Yaguang Li, Rose Yu, Cyrus Shahabi, and Yan Liu. Diffusion convolutional recurrent neural net- work: Data-driven traffic forecasting. InICLR (Poster). OpenReview.net,

    2018

  12. [24]

    Airformer: Predicting nationwide air quality in china with transformers

    [Lianget al., 2023 ] Yuxuan Liang, Yutong Xia, Songyu Ke, Yiwei Wang, Qingsong Wen, Junbo Zhang, Yu Zheng, and Roger Zimmermann. Airformer: Predicting nationwide air quality in china with transformers. InAAAI,

    2023

  13. [25]

    Rethinking sensors modeling: Hierarchical information enhanced traffic forecasting

    [Maet al., 2023 ] Qian Ma, Zijian Zhang, Xiangyu Zhao, Haoliang Li, Hongwei Zhao, Yiqi Wang, Zitao Liu, and Wanyu Wang. Rethinking sensors modeling: Hierarchical information enhanced traffic forecasting. InCIKM,

    2023

  14. [26]

    Towards efficient large scale spatial-temporal time series forecasting via improved inverted transformers.arXiv preprint arXiv:2503.10858,

    [Sunet al., 2025 ] Jiarui Sun, Chin-Chia Michael Yeh, Yu- jie Fan, Xin Dai, Xiran Fan, Zhimeng Jiang, Uday Singh Saini, Vivian Lai, Junpeng Wang, Huiyuan Chen, et al. Towards efficient large scale spatial-temporal time series forecasting via improved inverted transformers.arXiv preprint arXiv:2503.10858,

    Pith/arXiv arXiv 2025

  15. [27]

    [Tolstikhinet al., 2021 ] Ilya O. Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, Mario Lucic, and Alexey Dosovitskiy. Mlp-mixer: An all-mlp architecture for vi- sion. InNeurIPS, pages 24261–24272,

    2021

  16. [28]

    Traffic speed predic- tion and congestion source exploration: A deep learning method

    [Wanget al., 2016 ] Jingyuan Wang, Qian Gu, Junjie Wu, Guannan Liu, and Zhang Xiong. Traffic speed predic- tion and congestion source exploration: A deep learning method. In2016 IEEE 16th international conference on data mining (ICDM), pages 499–508. IEEE,

    2016

  17. [29]

    Deep fuzzy cognitive maps for interpretable multivariate time series prediction.IEEE transactions on fuzzy systems,

    [Wanget al., 2020 ] Jingyuan Wang, Zhen Peng, Xiaoda Wang, Chao Li, and Junjie Wu. Deep fuzzy cognitive maps for interpretable multivariate time series prediction.IEEE transactions on fuzzy systems,

    2020

  18. [30]

    Libcity: An open library for traffic prediction

    [Wanget al., 2021 ] Jingyuan Wang, Jiawei Jiang, Wenjun Jiang, Chao Li, and Wayne Xin Zhao. Libcity: An open library for traffic prediction. InSIGSPATIAL/GIS,

    2021

  19. [31]

    Traffic flow prediction based on spa- tiotemporal potential energy fields.IEEE Transactions on Knowledge and Data Engineering, 35(9),

    [Wanget al., 2022 ] Jingyuan Wang, Jiahao Ji, Zhe Jiang, and Leilei Sun. Traffic flow prediction based on spa- tiotemporal potential energy fields.IEEE Transactions on Knowledge and Data Engineering, 35(9),

    2022

  20. [32]

    When: A wavelet-dtw hybrid at- tention network for heterogeneous time series analysis

    [Wanget al., 2023 ] Jingyuan Wang, Chen Yang, Xiaohan Jiang, and Junjie Wu. When: A wavelet-dtw hybrid at- tention network for heterogeneous time series analysis. InProceedings of the 29th ACM SIGKDD conference on knowledge discovery and data mining,

    2023

  21. [33]

    Graph wavenet for deep spatial-temporal graph modeling

    [Wuet al., 2019 ] Zonghan Wu, Shirui Pan, Guodong Long, Jing Jiang, and Chengqi Zhang. Graph wavenet for deep spatial-temporal graph modeling. InIJCAI,

    2019

  22. [34]

    Rpmixer: Shak- ing up time series forecasting with random projections for large spatial-temporal data

    [Yehet al., 2024 ] Chin-Chia Michael Yeh, Yujie Fan, Xin Dai, Uday Singh Saini, Vivian Lai, Prince Osei Aboagye, Junpeng Wang, Huiyuan Chen, Yan Zheng, Zhongfang Zhuang, Liang Wang, and Wei Zhang. Rpmixer: Shak- ing up time series forecasting with random projections for large spatial-temporal data. InKDD. ACM,

    2024

  23. [35]

    Frequency-domain mlps are more effective learners in time series forecasting

    [Yiet al., 2023 ] Kun Yi, Qi Zhang, Wei Fan, Shoujin Wang, Pengyang Wang, Hui He, Ning An, Defu Lian, Long- bing Cao, and Zhendong Niu. Frequency-domain mlps are more effective learners in time series forecasting. In NeurIPS,

    2023

  24. [36]

    L2-GCN: layer-wise and learned efficient training of graph convolutional networks

    [Youet al., 2020 ] Yuning You, Tianlong Chen, Zhangyang Wang, and Yang Shen. L2-GCN: layer-wise and learned efficient training of graph convolutional networks. In CVPR. Computer Vision Foundation / IEEE,

    2020

  25. [37]

    BIGCity: A Universal Spatiotem- poral Model for Unified Trajectory and Traffic State Data Analysis

    [Yuet al., ] Xie Yu, Jingyuan Wang, Yifan Yang, Qian Huang, and Ke Qu. BIGCity: A Universal Spatiotem- poral Model for Unified Trajectory and Traffic State Data Analysis . In41st ICDE. IEEE Computer Society, May. [Yuet al., 2018 ] Bing Yu, Haoteng Yin, and Zhanxing Zhu. Spatio-temporal graph convolutional networks: A deep learning framework for traffic fo...

    2018

  26. [38]

    Joe Qin, and Hongwei Zhao

    [Zhanget al., 2023 ] Zijian Zhang, Ze Huang, Zhiwei Hu, Xiangyu Zhao, Wanyu Wang, Zitao Liu, Junbo Zhang, S. Joe Qin, and Hongwei Zhao. MLPST: MLP is all you need for spatio-temporal prediction. InCIKM, pages 3381–3390. ACM,

    2023

  27. [39]

    Hyperparameter Search.We employ a grid search strat- egy to identify the optimal hyperparameter configurations for our HSTMixer in Table

    A Appendix A.1 Implementation Details Baselines.We compare HSTMixer with the following 19 SOTA methods from two categories:(1)Previous SOTAs contains ASTGCN [Guoet al., 2019 ], DCRNN [Liet al., 2018], STGCN[Yuet al., 2018 ], GWNET[Wuet al., 2019 ], STGODE[Fanget al., 2021 ], DSTAGNN[Lanet al., 2022 ], D2STGNN[Shaoet al., 2022b ], HIEST [Maet al., 2023 ], ...

    2019

  28. [2009]

    Arik, and Tomas Pfister

    [Chenet al., 2023 ] Si-An Chen, Chun-Liang Li, Nate Yo- der, Sercan ¨O. Arik, and Tomas Pfister. Tsmixer: An all-mlp architecture for time series forecasting.CoRR, abs/2303.06053,

    Pith/arXiv arXiv 2023

  29. [2014]

    Attention based spatial- temporal graph convolutional networks for traffic flow forecasting

    [Guoet al., 2019 ] Shengnan Guo, Youfang Lin, Ning Feng, Chao Song, and Huaiyu Wan. Attention based spatial- temporal graph convolutional networks for traffic flow forecasting. InAAAI, pages 922–929. AAAI Press,

    2019

  30. [2015]

    DSTAGNN: dynamic spatial-temporal aware graph neural network for traffic flow forecasting

    [Lanet al., 2022 ] Shiyong Lan, Yitong Ma, Weikang Huang, Wenwu Wang, Hongyu Yang, and Pyang Li. DSTAGNN: dynamic spatial-temporal aware graph neural network for traffic flow forecasting. InICML, volume 162 ofProceed- ings of Machine Learning Research. PMLR,

    2022

  31. [2016]

    [Guoet al., 2014 ] Jianhua Guo, Wei Huang, and Billy M Williams. Adaptive kalman filter approach for stochas- tic short-term traffic flow rate prediction and uncertainty quantification.Transportation Research Part C: Emerging Technologies, 43:50–64,

    2014

  32. [2017]

    Pre-training identification of graph winning tickets in adaptive spatial-temporal graph neural networks

    [Duanet al., 2024 ] Wenying Duan, Tianxiang Fang, Hong Rao, and Xiaoxi He. Pre-training identification of graph winning tickets in adaptive spatial-temporal graph neural networks. In30th KDD,

    2024

  33. [2018]

    Dynamic graph convolutional recurrent network for traffic prediction: Benchmark and solution.ACM Trans

    [Liet al., 2023 ] Fuxian Li, Jie Feng, Huan Yan, Guangyin Jin, Fan Yang, Funing Sun, Depeng Jin, and Yong Li. Dynamic graph convolutional recurrent network for traffic prediction: Benchmark and solution.ACM Trans. Knowl. Discov. Data, 17(1):9:1–9:21,

    2023

  34. [2019]

    Bridging traffic state and trajectory for dynamic road network and trajectory representation learning

    [Hanet al., 2025 ] Chengkai Han, Jingyuan Wang, Yongyao Wang, Xie Yu, Hao Lin, Chao Li, and Junjie Wu. Bridging traffic state and trajectory for dynamic road network and trajectory representation learning. InProceedings of the AAAI Conference on Artificial Intelligence,

    2025

  35. [2020]

    Online-svr for short-term traffic flow prediction under typical and atyp- ical traffic conditions.Expert systems with applications, 36(3):6164–6173,

    [Castro-Netoet al., 2009 ] Manoel Castro-Neto, Young-Seon Jeong, Myong-Kee Jeong, and Lee D Han. Online-svr for short-term traffic flow prediction under typical and atyp- ical traffic conditions.Expert systems with applications, 36(3):6164–6173,

    2009

  36. [2021]

    Efficient large-scale traffic forecasting with trans- formers: A spatial data management perspective.arXiv preprint arXiv:2412.09972,

    [Fanget al., 2024 ] Yuchen Fang, Yuxuan Liang, Bo Hui, Zezhi Shao, Liwei Deng, Xu Liu, Xinke Jiang, and Kai Zheng. Efficient large-scale traffic forecasting with trans- formers: A spatial data management perspective.arXiv preprint arXiv:2412.09972,

    Pith/arXiv arXiv 2024

  37. [2022]

    TES- TAM: A time-enhanced spatio-temporal attention model with mixture of experts

    [Lee and Ko, 2024] Hyunwook Lee and Sungahn Ko. TES- TAM: A time-enhanced spatio-temporal attention model with mixture of experts. InICLR. OpenReview.net,

    2024

  38. [2023]

    Poi-enhancer: An llm-based semantic en- hancement framework for poi representation learning

    [Chenget al., 2025 ] Jiawei Cheng, Jingyuan Wang, Yichuan Zhang, Jiahao Ji, Yuanshao Zhu, Zhibo Zhang, and Xi- angyu Zhao. Poi-enhancer: An llm-based semantic en- hancement framework for poi representation learning. In AAAI,

    2025

  39. [2024]

    Spatial-temporal graph ODE net- works for traffic flow forecasting

    [Fanget al., 2021 ] Zheng Fang, Qingqing Long, Guojie Song, and Kunqing Xie. Spatial-temporal graph ODE net- works for traffic flow forecasting. InKDD. ACM,

    2021

  40. [2025]

    Dauphin, Angela Fan, Michael Auli, and David Grangier

    [Dauphinet al., 2017 ] Yann N. Dauphin, Angela Fan, Michael Auli, and David Grangier. Language modeling with gated convolutional networks. InICML, volume 70 of Proceedings of Machine Learning Research. PMLR,

    2017