Pith. sign in

REVIEW 3 major objections 5 minor 40 references

Standard spatial attention yields no consistent accuracy advantage over uniform full-range mixing in traffic forecasting, while costing O(N^2) instead of O(N).

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

2026-08-02 06:29 UTC pith:6P7BKESV

load-bearing objection Clean controlled ablation, but the conclusion overstates: under flow-only inputs, standard attention consistently wins, so the 'no consistent advantage' claim is scoped to the full representation only. the 3 major comments →

arxiv 2607.12462 v2 pith:6P7BKESV submitted 2026-07-14 cs.AI

Do We Really Need Adaptive Global Spatial Attention for Traffic Forecasting?

Qihang Zhang , Siyao Zhang , Letao Kang , Wenzhe Liang , Miao Zhang , Zhao Zhang This is my paper
classification cs.AI
keywords traffic forecastingspatial attentionglobal spatial informationuniform full-range mixingattention decompositionnode-mean broadcastingcontrolled ablationspatio-temporal forecasting
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.

The paper tries to establish that the global spatial information used by traffic forecasting models does not require learned, input-dependent pairwise attention weights. Across six benchmarks, replacing the spatial attention module with a row-uniform full-range averaging operator changes mean absolute error by only -1.58% to +1.26%, with no consistent direction of improvement, while cutting node-scale mixing complexity from O(N^2) to O(N). The authors argue that much of attention's benefit is captured by a shared global background common to all nodes, and that the non-uniform attention residual is dataset-dependent: beneficial on some datasets, harmful on others, and nearly neutral on a third. If right, this means simpler, cheaper global mixers are a strong baseline and a plausible explanation for why complex spatial attention modules often do not outperform simpler representations.

Core claim

The central claim is that adaptive global spatial attention—softmax attention over all node pairs—does not consistently outperform a fixed row-uniform global mixer (each node receives the arithmetic mean of all node representations) for traffic forecasting. In a controlled framework where only the spatial mixing module changes, the learned attention variant yields relative MAE changes between -1.58% and +1.26% versus uniform mixing across four road-sensor and two grid-demand datasets; the sign flips, so there is no consistent advantage. The paper further decomposes spatial attention into a row-uniform global background plus a zero-row-sum non-uniform residual, and shows empirically that the

What carries the argument

The controlled ablation framework isolates the spatial mixing module while fixing input representation, temporal encoder, and output head. The decisive comparison is U-Trans, a row-uniform aggregation matrix with entries A_ij = 1/N implemented as an O(N) node-mean broadcast, against S-Trans, an input-dependent multi-head softmax attention matrix built from queries and keys over all N nodes. The decomposition A_t = U + Δ_t, where Δ_t has zero row sums, splits attention into a shared global background and a non-uniform residual; residual-strength interpolation (RI-Trans) and front global-mean injection then test the residual's predictive role and show that it is dataset-dependent.

Load-bearing premise

The conclusion rests on the assumption that the implemented standard spatial attention is a fair, sufficiently tuned representative of adaptive global attention under the fixed hyperparameter protocol; if that variant is under-tuned or if the fixed temporal encoder and rich input representation already capture global context, the absence of an advantage may not generalize to other attention designs.

What would settle it

Re-running the controlled comparison on a dataset with pronounced, spatially structured local propagation—or a synthetic network with long-range correlated shocks—where a properly tuned standard spatial attention beats uniform full-range mixing by more than 2% relative MAE with non-overlapping confidence intervals across at least three seeds would falsify the claim that the non-uniform residual is not consistently beneficial.

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

If this is right

  • If uniform full-range mixing is a strong global spatial baseline, future traffic forecasting models can drop dense O(N^2) attention for global interaction without consistent forecast loss.
  • The saved computation can be reallocated to input representation or temporal modeling, which the paper shows matters more than adaptive pairwise weighting.
  • Dataset-dependent residual effects mean that attention is not harmful per se, but its value must be demonstrated per dataset rather than assumed.
  • Because the attention residual operates on node-local fluctuations and cancels the shared global state, adding more global context through attention cannot improve the estimate of the common background.
  • Public-baseline comparisons suggest a simple row-uniform mixer is competitive with graph-based and Transformer-based models, shifting the burden of proof to designs that add complex spatial modules.

Where Pith is reading between the lines

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

  • The result likely extends beyond traffic: any forecasting task with shared network-level trends and many nodes may do as well with mean broadcasting as with learned pairwise attention; this could be tested on energy, crowd, or supply-chain networks.
  • The decomposition suggests a testable hypothesis: if node-local fluctuations are approximately noise with no spatial structure, the attention residual has nothing useful to learn, so uniform mixing should win; if local structure is strong and informative, attention should win. A synthetic experiment with controlled spatial autocorrelation could confirm this mechanism.
  • The inference-only interpolation shows that post-hoc reduction of attention strength degrades performance because the trained model has co-adapted to its attention weights; retrained uniform baselines remain competitive, so co-adaptation deserves attention when interpreting attention-removal studies.
  • The O(N) versus O(N^2) gain is most valuable at very large sensor counts; the paper's largest dataset has 1024 nodes, so scaling to tens of thousands of nodes is the natural stress test for the efficiency claim.

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 paper asks whether global spatial information in traffic forecasting requires adaptive pairwise attention or can be captured by a simple uniform full-range aggregation operator. It builds a controlled framework that keeps input representation, temporal encoder, and output head fixed and varies only the spatial mixing module: No-SAB, I-Trans, UD-Trans, U-Trans, and S-Trans. Across six benchmarks (four PEMS sensor networks and two grid datasets), U-Trans matches or improves on S-Trans in most configurations; the reported relative MAE changes of S-Trans vs U-Trans are -1.58% to +1.26%, and U-Trans has O(N) node-scale spatial-mixing complexity versus O(N^2). The paper also proposes a decomposition of row-stochastic attention into a uniform background plus a zero-row-sum residual, with a hypothesized latent shared state, and uses interpolation/front mean-injection diagnostics to argue the residual has dataset-dependent effects.

Significance. If correct, this is a valuable negative result for traffic-forecasting architecture design: it suggests global context can often be supplied by a simple row-uniform mixer, and it provides a clean experimental template for isolating spatial-mixing choices. The controlled setup, five-seed reporting, released code, spatial-depth and embedding robustness checks, and the exact A = U + Delta identity are strengths. The main caveat is that the headline is currently stated more broadly than the evidence, especially in light of the flow-only ablation.

major comments (3)
  1. [Conclusion / Appendix C] Appendix C, Tables 7-12, show that in the flow-only setting S-Trans obtains lower MAE than U-Trans on all six datasets (e.g., PEMS03 16.7475 vs 16.8536; T-DRIVE 23.3364 vs 23.8803). The main text states 'S-Trans exhibits a consistent cross-dataset advantage only under the flow-only setting,' but the Conclusion asserts 'embedding ablations across all six benchmarks show the same cross-dataset pattern under the full and partially ablated configurations considered in this work.' If 'same cross-dataset pattern' means 'no consistent advantage,' the statement is contradicted by the flow-only tables; if it means something weaker, it needs to be restated. This is load-bearing because it delimits the central claim: under a representation the paper itself varies, adaptive attention is indeed consistently beneficial. Please revise the conclusion/abstract to scope the claim to the full (or all non-f
  2. [Table 2 / Experimental Settings] No significance tests are reported, and most mean differences in Table 2 are within one standard deviation (e.g., PEMS04: 18.12 +/- 0.02 vs 18.13 +/- 0.03; PEMS08: 13.33 +/- 0.06 vs 13.37 +/- 0.03). The claim 'no consistent advantage' is a negative/equivalence-type claim; mean directions alone are insufficient. If the five seeds are shared across models, report paired per-seed comparisons (e.g., paired t-test or Wilcoxon, or bootstrap CIs on per-dataset differences); if not, state that and use appropriate tests. Also quantify the flow-only 6/6 win direction (e.g., sign test p = 0.031), which would strengthen the paper's own finding. This is needed before the abstract's categorical statement can be accepted.
  3. [Implementation / External validity] The central comparison fixes hyperparameters to STAEformer's settings and uses a single S-Trans instantiation. This is methodologically clean for internal comparison, but the paper's title and abstract generalize to 'adaptive global spatial attention.' S-Trans is not tuned under the new framework, and no alternative spatial-attention design is tested; the absence of consistent advantage may reflect the chosen training protocol or one specific attention parameterization. Please either restrict the claims to 'the evaluated S-Trans implementation under STAEformer-style hyperparameters' throughout (including the abstract), or add a small sensitivity study (e.g., varying heads, dropout, learning rate, or comparing with a second standard attention variant) to support the broader wording.
minor comments (5)
  1. [Appendix C] The embedding ablation tables report only mean MAE/RMSE/MAPE without standard deviations or seed-level spread; adding them would help judge the reliability of the flow-only result and other comparisons.
  2. [Tables 4-5] The public-baseline comparison mixes numbers taken from published papers with the authors' own runs. Please state which settings are matched (e.g., data splits, evaluation horizons, normalization) and how protocol differences might affect the comparison; if the table is intended as indicative, say so.
  3. [Equation (14) / Section 4] The text says RI-Trans becomes MH-U-Trans when r=0, but earlier it is shown that MH-U-Trans is mathematically equivalent to U-Trans. For clarity, state that r=0 corresponds to the row-uniform U-Trans behavior.
  4. [Main Test / Table 2] The sentence 'Table 2 shows the same cross-dataset finding for RMSE and MAPE' is too strong; for example, PEMS04 RMSE moves slightly in the opposite direction to MAE (30.03 vs 30.01). Clarify that the pattern refers to the absence of a consistent directional advantage rather than identical signs on every metric.
  5. [Figure 3 / Appendix E] The interpolation and front-compensation plots report five-seed means; adding confidence intervals or seed-wise points would make the dataset-dependent pattern easier to assess.

Circularity Check

0 steps flagged

No significant circularity: the attention decomposition is a mathematical identity and the central claim rests on controlled ablations, not on fitted predictions or self-citation.

full rationale

This paper is a controlled empirical ablation study and does not exhibit circular reasoning at the level of its central claim. The attention decomposition A_t = U + Δ_t with Δ_t 1 = 0 (Eq. 18) is a mathematical identity for any row-stochastic attention matrix, not a fitted or predicted quantity; Eq. (20) merely exploits the zero row-sum property. The latent-state decomposition v_{j,t} = g_t + ε_{j,t} (Eq. 16) is explicitly labeled a hypothesis, and the main claim—S-Trans shows no consistent advantage over U-Trans—rests on direct endpoint comparisons (Table 2) and embedding ablations (Appendix C), not on that hypothesis. The RI-Trans interpolation and front-injection diagnostics are independently trained experiments, not constructions equivalent to their inputs. The O(N) versus O(N^2) complexity claim is an implementation property, not a fitted result. The STAEformer hyperparameter protocol is an external, non-self citation and does not smuggle in the conclusion. The paper does cite Shao et al. 2022a/b (overlapping author Zhao Zhang), but only as related-work context; these citations are not load-bearing. The internal tension between Appendix C (flow-only input shows a consistent S-Trans advantage) and the Conclusion's statement that 'embedding ablations across all six benchmarks show the same cross-dataset pattern' is a reporting/consistency concern that affects the strength of the generalization claim, not circularity in the derivation chain. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported, and no ansatz is smuggled in via self-citation.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 1 invented entities

The experimental claim is empirical; the main invented construct is the latent state used in the explanatory model. Hand-chosen thresholds, interpolation coefficients, and per-dataset depth settings are listed as free parameters because the reported results are conditioned on them.

free parameters (5)
  • residual strength r = 0, 0.25, 0.5, 0.75, 1
    Interpolation coefficient in RI-Trans (Eq. 14/21); chosen by hand to diagnose the residual effect, not fit to data.
  • mask threshold for grid metrics = 10
    Grid datasets report masked error with threshold 10; this hand-picked cutoff affects the reported grid results.
  • UD-Trans hop count k = 5
    Five-hop local neighborhood on PEMS datasets; chosen by hand and conditions the topology-based baseline.
  • free-flow speed quantile = 0.95
    Congestion degree uses the 95th percentile of training-set speed (Eq. A1); fixed without sensitivity analysis.
  • spatial layer count Ls = 1 on PEMS03/04/08/NYCTAXI, 3 on PEMS07/T-DRIVE
    Per-dataset depth choice; main conclusions use this setting, with robustness checks in Appendix B.
axioms (5)
  • standard math Any row-stochastic attention matrix decomposes as U + Delta with Delta*1 = 0.
    Used to split S-Trans into a row-uniform background and a non-uniform residual (Eq. 18).
  • standard math Row-wise softmax is invariant to row-wise constant logit shifts.
    Used in the front global-mean injection derivation (Appendix D, Eq. D4).
  • ad hoc to paper Value-projected node representations decompose into a latent shared global state plus node-local fluctuation.
    Central to interpreting the residual; Equation (16) introduces g_t with no independent measurement.
  • domain assumption The six datasets and the fixed STAEformer-derived protocol are representative enough to test adaptive attention fairly.
    Supports the generalization from the controlled ablations to 'no consistent advantage'.
  • domain assumption The fixed temporal encoder and input representation are strong enough that a benefit of adaptive spatial attention would be visible.
    If the representation already contains global context, the test is biased against attention; the paper partially checks this via embedding ablations.
invented entities (1)
  • latent shared global state g_t no independent evidence
    purpose: Explains why the uniform mean is a strong global baseline: the attention residual cancels the common state and acts only on node-specific fluctuations.
    Introduced as a hypothesis (Eq. 16); U-Trans mean is called an empirical proxy, but no direct falsifiable measurement of g_t is provided.

pith-pipeline@v1.3.0-alltime-deepseek · 18321 in / 14692 out tokens · 146939 ms · 2026-08-02T06:29:09.130532+00:00 · methodology

0 comments
read the original abstract

Existing traffic forecasting models commonly focus on extracting spatial dependencies, particularly global spatial information, which characterizes the representations obtained through interactions between each node and all nodes across the traffic network. However, the underlying mechanism by which global information is modeled and extracted remains insufficiently investigated. Whether global information must be extracted by high-degree-of-freedom adaptive attention or can be captured by a simple global aggregation operator remains unclear. For this purpose, we design a controlled ablation framework that replaces only the spatial mixing module to test attention-based global interaction. Across six traffic benchmarks, standard spatial attention yields relative MAE changes of $-1.58\%$ to $+1.26\%$ compared with uniform full-range mixing, and we observe no consistent advantage for standard spatial attention, while uniform full-range mixing reduces node-scale spatial-mixing complexity from $O(N^2)$ to $O(N)$. We further propose a hypothesized model that decomposes spatial attention into a row-uniform global background and a non-uniform residual. The residual shows dataset-dependent effects. Overall, uniform full-range mixing provides a strong global spatial baseline, while the non-uniform attention residual is not consistently beneficial across datasets.

Figures

Figures reproduced from arXiv: 2607.12462 by Letao Kang, Miao Zhang, Qihang Zhang, Siyao Zhang, Wenzhe Liang, Zhao Zhang.

Figure 1
Figure 1. Figure 1: Unified forecasting framework. 3.2 Fixed Input Representation All spatial variants use the same input representation, formed by concatenating projected traffic flow, congestion degree, temporal embeddings, and adaptive embeddings. Congestion degree is computed from the training-set free￾flow speed; TOD/DOW settings and projection details are given in Appendix A. The resulting representation is: 𝐙 = Concat(… view at source ↗
Figure 1
Figure 1. Figure 1: Unified forecasting framework. Controlled Spatial Mixing Module The spatial module receives the temporally encoded multi-source node representations and compares five spa￾tial settings: No Spatial Aggregation Block (No-SAB), Identity Transformer (I-Trans), Upstream-Downstream Transformer (UD-Trans), Uniform Transformer (U￾Trans), and Standard Spatial Transformer (S-Trans). Except for No-SAB, all variants s… view at source ↗
Figure 3
Figure 3. Figure 3: provides a trajectory-level comparison between U-Trans and S-Trans on PEMS03 node 0 at 5-min and 60- min horizons, showing similar prediction patterns. Both mixers follow the main drop, recovery, valley-peak pattern, and later decline, with larger deviations at the longer horizon. This example is intended as a qualitative illustra￾tion; the dataset-level comparison is presented in Section 4.4 [PITH_FULL_I… view at source ↗
Figure 2
Figure 2. Figure 2: MAE comparison on six datasets [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: MAE comparison across spatial modules. U-Trans outperforms No-SAB on all six datasets. More importantly, I-Trans and U-Trans share the same spatial￾block scaffold and differ only in node aggregation: I-Trans uses identity aggregation, whereas U-Trans broadcasts the full-range node mean. U-Trans achieves lower MAE than I￾Trans on five of the six datasets, indicating that row-uniform full-range propagation u… view at source ↗
Figure 3
Figure 3. Figure 3: visualizes the two diagnostics under the same MH-U-Trans reference [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Global background and node-local fluctuation [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: visualizes the two diagnostics under the same MH-U-Trans reference [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] 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

40 extracted references · 4 canonical work pages

  1. [1]

    Adaptive Graph Convolutional Recurrent Network for Traffic Forecasting , url =

    Bai, Lei and Yao, Lina and Li, Can and Wang, Xianzhi and Wang, Can , booktitle =. Adaptive Graph Convolutional Recurrent Network for Traffic Forecasting , url =

  2. [2]

    GCNet: Non-Local Networks Meet Squeeze-Excitation Networks and Beyond , year=

    Cao, Yue and Xu, Jiarui and Lin, Stephen and Wei, Fangyun and Hu, Han , booktitle=. GCNet: Non-Local Networks Meet Squeeze-Excitation Networks and Beyond , year=

  3. [3]

    SimpleTM: A Simple Baseline for Multivariate Time Series Forecasting , url =

    Chen, Hui and Luong, Viet and Mukherjee, Lopamudra and Singh, Vikas , booktitle =. SimpleTM: A Simple Baseline for Multivariate Time Series Forecasting , url =

  4. [4]

    Proceedings of the AAAI Conference on Artificial Intelligence , author=

    Designing Specialized Two-Dimensional Graph Spectral Filters for Spatial-Temporal Graph Modeling , volume=. Proceedings of the AAAI Conference on Artificial Intelligence , author=. 2025 , month=. doi:10.1609/aaai.v39i11.33251 , abstractNote=

    doi:10.1609/aaai.v39i11.33251 2025

  5. [5]

    Proceedings of the AAAI Conference on Artificial Intelligence , author=

    Graph Neural Controlled Differential Equations for Traffic Forecasting , volume=. Proceedings of the AAAI Conference on Artificial Intelligence , author=. 2022 , month=. doi:10.1609/aaai.v36i6.20587 , abstractNote=

    doi:10.1609/aaai.v36i6.20587 2022

  6. [6]

    Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence,

    GSTNet: Global Spatial-Temporal Network for Traffic Flow Prediction , author =. Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence,. 2019 , month =. doi:10.24963/ijcai.2019/317 , url =

    work page doi:10.24963/ijcai.2019/317 2019

  7. [7]

    Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.1 , pages =

    Fang, Yuchen and Liang, Yuxuan and Hui, Bo and Shao, Zezhi and Deng, Liwei and Liu, Xu and Jiang, Xinke and Zheng, Kai , title =. Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.1 , pages =. 2025 , isbn =. doi:10.1145/3690624.3709177 , abstract =

    arXiv 2025

  8. [8]

    Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining , pages =

    Fang, Zheng and Long, Qingqing and Song, Guojie and Xie, Kunqing , title =. Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining , pages =. 2021 , isbn =. doi:10.1145/3447548.3467430 , abstract =

    arXiv 2021

  9. [9]

    Proceedings of the AAAI Conference on Artificial Intelligence , author=

    Attention Based Spatial-Temporal Graph Convolutional Networks for Traffic Flow Forecasting , volume=. Proceedings of the AAAI Conference on Artificial Intelligence , author=. 2019 , month=. doi:10.1609/aaai.v33i01.3301922 , abstractNote=

    doi:10.1609/aaai.v33i01.3301922 2019

  10. [10]

    Learning Dynamics and Heterogeneity of Spatial-Temporal Graph Data for Traffic Forecasting , year=

    Guo, Shengnan and Lin, Youfang and Wan, Huaiyu and Li, Xiucheng and Cong, Gao , journal=. Learning Dynamics and Heterogeneity of Spatial-Temporal Graph Data for Traffic Forecasting , year=

  11. [11]

    Jindong Han and Weijia Zhang and Hao Liu and Tao Tao and Naiqiang Tan and Hui Xiong , title =. Proc. 2024 , url =. doi:10.14778/3641204.3641217 , timestamp =

    arXiv 2024

  12. [12]

    Proceedings of the AAAI Conference on Artificial Intelligence , author=

    STD-PLM: Understanding Both Spatial and Temporal Properties of Spatial-Temporal Data with PLM , volume=. Proceedings of the AAAI Conference on Artificial Intelligence , author=. 2025 , month=. doi:10.1609/aaai.v39i11.33286 , abstractNote=

    work page doi:10.1609/aaai.v39i11.33286 2025

  13. [13]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =

    Nam, Hyeon-Woo and Kim, Yu-Ji and Heo, Byeongho and Han, Dongyoon and Oh, Seong Joon and Oh, Tae-Hyun , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =. 2023 , pages =

    2023

  14. [14]

    Proceedings of the AAAI Conference on Artificial Intelligence , author=

    PDFormer: Propagation Delay-Aware Dynamic Long-Range Transformer for Traffic Flow Prediction , volume=. Proceedings of the AAAI Conference on Artificial Intelligence , author=. 2023 , month=. doi:10.1609/aaai.v37i4.25556 , abstractNote=

    doi:10.1609/aaai.v37i4.25556 2023

  15. [15]

    Proceedings of the Third Learning on Graphs Conference , pages =

    Enhancing Topological Dependencies in Spatio-Temporal Graphs With Cycle Message Passing Blocks , author =. Proceedings of the Third Learning on Graphs Conference , pages =. 2025 , editor =

    2025

  16. [16]

    FN et: Mixing Tokens with F ourier Transforms

    Lee-Thorp, James and Ainslie, Joshua and Eckstein, Ilya and Onta \ n \'o n, Santiago. FN et: Mixing Tokens with F ourier Transforms. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022. doi:10.18653/v1/2022.naacl-main.319

    doi:10.18653/v1/2022.naacl-main.319 2022

  17. [17]

    International Conference on Learning Representations , year =

    Li, Yaguang and Yu, Rose and Shahabi, Cyrus and Liu, Yan , title =. International Conference on Learning Representations , year =

  18. [18]

    Information Fusion , volume =

    Lin, Jiaqi and Ren, Qianqian and Lv, Xingfeng and Xu, Hui and Liu, Yong , title =. Information Fusion , volume =. 2025 , doi =

    2025

  19. [19]

    Proceedings of the 32nd ACM International Conference on Information and Knowledge Management , pages =

    Liu, Hangchen and Dong, Zheng and Jiang, Renhe and Deng, Jiewen and Deng, Jinliang and Chen, Quanjun and Song, Xuan , title =. Proceedings of the 32nd ACM International Conference on Information and Knowledge Management , pages =. 2023 , isbn =. doi:10.1145/3583780.3615160 , abstract =

    arXiv 2023

  20. [20]

    2023 , eprint=

    Do We Really Need Graph Neural Networks for Traffic Forecasting? , author=. 2023 , eprint=

    2023

  21. [21]

    Less but More: Linear Adaptive Graph Learning Empowering Spatiotemporal Forecasting , url =

    Ma, Jiaming and Wang, Binwu and Wang, Guanjun and Yang, Kuo and Zhou, Zhengyang and Wang, Pengkun and Wang, Xu and Wang, Yang , booktitle =. Less but More: Linear Adaptive Graph Learning Empowering Spatiotemporal Forecasting , url =

  22. [22]

    Proceedings of the AAAI Conference on Artificial Intelligence , author=

    FC-GAGA: Fully Connected Gated Graph Architecture for Spatio-Temporal Traffic Forecasting , volume=. Proceedings of the AAAI Conference on Artificial Intelligence , author=. 2021 , month=. doi:10.1609/aaai.v35i10.17114 , abstractNote=

    work page doi:10.1609/aaai.v35i10.17114 2021

  23. [23]

    , title =

    Shao, Zezhi and Zhang, Zhao and Wei, Wei and Wang, Fei and Xu, Yongjun and Cao, Xin and Jensen, Christian S. , title =. Proc. VLDB Endow. , month = jul, pages =. 2022 , issue_date =. doi:10.14778/3551793.3551827 , abstract =

    arXiv 2022

  24. [24]

    Proceedings of the 31st ACM International Conference on Information & Knowledge Management , pages =

    Shao, Zezhi and Zhang, Zhao and Wang, Fei and Wei, Wei and Xu, Yongjun , title =. Proceedings of the 31st ACM International Conference on Information & Knowledge Management , pages =. 2022 , isbn =. doi:10.1145/3511808.3557702 , abstract =

    arXiv 2022

  25. [25]

    MLP-Mixer: An all-MLP Architecture for Vision , url =

    Tolstikhin, Ilya O and Houlsby, Neil and Kolesnikov, Alexander and Beyer, Lucas and Zhai, Xiaohua and Unterthiner, Thomas and Yung, Jessica and Steiner, Andreas and Keysers, Daniel and Uszkoreit, Jakob and Lucic, Mario and Dosovitskiy, Alexey , booktitle =. MLP-Mixer: An all-MLP Architecture for Vision , url =

  26. [26]

    Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence,

    Make Bricks with a Little Straw: Large-Scale Spatio-Temporal Graph Learning with Restricted GPU-Memory Capacity , author =. Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence,. 2024 , month =. doi:10.24963/ijcai.2024/264 , url =

    work page doi:10.24963/ijcai.2024/264 2024

  27. [27]

    Non-local Neural Networks , year=

    Wang, Xiaolong and Girshick, Ross and Gupta, Abhinav and He, Kaiming , booktitle=. Non-local Neural Networks , year=

  28. [28]

    and Poor, H

    Wang, Zepu and Nie, Yuqi and Liu, Yang and Mulvey, John M. and Poor, H. Vincent and Boukerche, Azzedine and Nguyen, Nam H. and Sun, Peng , title =. IEEE Transactions on Intelligent Transportation Systems , volume =. 2026 , doi =

    2026

  29. [29]

    Proceedings of the 36th International Conference on Machine Learning , pages =

    Simplifying Graph Convolutional Networks , author =. Proceedings of the 36th International Conference on Machine Learning , pages =. 2019 , editor =

    2019

  30. [30]

    Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence,

    Graph WaveNet for Deep Spatial-Temporal Graph Modeling , author =. Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence,. 2019 , month =. doi:10.24963/ijcai.2019/264 , url =

    doi:10.24963/ijcai.2019/264 2019

  31. [31]

    Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , pages =

    Wu, Zonghan and Pan, Shirui and Long, Guodong and Jiang, Jing and Chang, Xiaojun and Zhang, Chengqi , title =. Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , pages =. 2020 , isbn =. doi:10.1145/3394486.3403118 , abstract =

    arXiv 2020

  32. [32]

    2020 , eprint=

    Spatial-Temporal Transformer Networks for Traffic Flow Forecasting , author=. 2020 , eprint=

    2020

  33. [33]

    Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence,

    Spatio-Temporal Graph Convolutional Networks: A Deep Learning Framework for Traffic Forecasting , author =. Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence,. 2018 , month =. doi:10.24963/ijcai.2018/505 , url =

    doi:10.24963/ijcai.2018/505 2018

  34. [34]

    Proceedings of the AAAI Conference on Artificial Intelligence , author=

    Are Transformers Effective for Time Series Forecasting? , volume=. Proceedings of the AAAI Conference on Artificial Intelligence , author=. 2023 , month=. doi:10.1609/aaai.v37i9.26317 , abstractNote=

    doi:10.1609/aaai.v37i9.26317 2023

  35. [35]

    Joe and Zhao, Hongwei , title =

    Zhang, Zijian and Huang, Ze and Hu, Zhiwei and Zhao, Xiangyu and Wang, Wanyu and Liu, Zitao and Zhang, Junbo and Qin, S. Joe and Zhao, Hongwei , title =. Proceedings of the 32nd ACM International Conference on Information and Knowledge Management , pages =. 2023 , isbn =. doi:10.1145/3583780.3614969 , abstract =

    arXiv 2023

  36. [36]

    Proceedings of the AAAI Conference on Artificial Intelligence , author=

    GMAN: A Graph Multi-Attention Network for Traffic Prediction , volume=. Proceedings of the AAAI Conference on Artificial Intelligence , author=. 2020 , month=. doi:10.1609/aaai.v34i01.5477 , abstractNote=

    doi:10.1609/aaai.v34i01.5477 2020

  37. [37]

    Proceedings of the AAAI Conference on Artificial Intelligence , author=

    Spatial-Temporal Synchronous Graph Convolutional Networks: A New Framework for Spatial-Temporal Network Data Forecasting , volume=. Proceedings of the AAAI Conference on Artificial Intelligence , author=. 2020 , month=. doi:10.1609/aaai.v34i01.5438 , abstractNote=

    doi:10.1609/aaai.v34i01.5438 2020

  38. [38]

    Urban Traffic Prediction from Spatio-Temporal Data Using Deep Meta Learning , booktitle =

    Zheyi Pan and Yuxuan Liang and Weifeng Wang and Yong Yu and Yu Zheng and Junbo Zhang , editor =. Urban Traffic Prediction from Spatio-Temporal Data Using Deep Meta Learning , booktitle =. 2019 , url =. doi:10.1145/3292500.3330884 , timestamp =

    arXiv 2019

  39. [39]

    Dynamic Spatial-Temporal Representation Learning for Traffic Flow Prediction , year=

    Liu, Lingbo and Zhen, Jiajie and Li, Guanbin and Zhan, Geng and He, Zhaocheng and Du, Bowen and Lin, Liang , journal=. Dynamic Spatial-Temporal Representation Learning for Traffic Flow Prediction , year=

  40. [40]

    Performance Measurement System (. n.d. , howpublished =