Pith. sign in

REVIEW 3 major objections 4 minor 84 references

EMAGN claims that spatial self-attention for traffic forecasting can be made linear—O(N·M) instead of O(N²)—by learning two clustering matrices that group keys and values into M super-clusters, losing only 2.7–3.2% MAE against full-attentio

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 05:45 UTC pith:GA6R6RWH

load-bearing objection The 'learned clustering' attention is Linformer with a new story; the efficiency numbers are plausible, but the central novelty claim collapses. the 3 major comments →

arxiv 2607.13241 v1 pith:GA6R6RWH submitted 2026-07-14 cs.LG cs.AI

EMAGN: Efficient Multi-Attention Graph Network via Learned Clustering for Scalable Traffic Forecasting

Mingxing Xu , Rakesh Chowdary Machineni , Ke Liu , Xi Cheng , Chengqi Lu , Xin Hu , Lyuhao Chen , Xiangyu Li
show 2 more authors
Junwei You Oliver Gao
This is my paper
classification cs.LG cs.AI
keywords traffic forecastingself-attentionlinear complexitygraph neural networkslearned clusteringGaussian filteringspatial-temporal modelingmulti-head attention
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 quadratic spatial self-attention used in traffic forecasting can be replaced by a linear-complexity variant without a meaningful accuracy collapse. To do this it transplants the idea of fast high-dimensional Gaussian filtering into attention: two learned matrices C_k and C_v adaptively group N keys and values into M super-clusters, so each query attends to M summaries instead of all N sensors. On PEMS-BAY and METR-LA the resulting model, EMAGN, stays within 2.7–3.2% MAE of the full-attention GMAN while training 32% faster, inferring 38% faster, and using 58% less GPU memory, and it can run with 16 heads where GMAN runs out of memory. If right, it means attention-based forecasting can scale to city-scale sensor networks on commodity hardware, and that learned clustering is the reason, outperforming generic linear-attention methods like Linformer and Performer in the same backbone.

Core claim

The central claim is that the attention formula X̂ = Softmax(αQ(C_k K)^T) C_v V is a linear-complexity self-attention derived from fast Gaussian filtering, and that the learned matrices C_k, C_v implement adaptive clustering of keys and values into M super-clusters, reducing complexity from O(N²d) to O(NMd). The paper argues this preserves attention's flexibility for dynamic dependency modelling because the grouping is learned end-to-end with the forecasting objective, rather than fixed or random as in Linformer or Performer. Empirically, on METR-LA and PEMS-BAY, EMAGN lands within 2.7–3.2% MAE of full-attention GMAN while cutting training time by 32%, inference by 38%, and memory by 58%, an

What carries the argument

The central object is the linear-complexity attention formula X̂ = Softmax(αQ(C_k K)^T) C_v V, with C_k, C_v ∈ R^{M×N} learned clustering matrices that compress N keys and values into M super-clusters. It is presented as an analog of fast high-dimensional Gaussian filtering, in which nearby positions are grouped into cluster centers and their contributions pre-aggregated via a truncated Taylor expansion. In EMAGN the matrices are trained end-to-end, replacing hand-crafted or random grouping, and that learned grouping is what the paper credits for outperforming Linformer and Performer in the same GMAN backbone.

Load-bearing premise

The load-bearing premise is that the learned matrices C_k and C_v actually implement the clustering from fast Gaussian filtering — that unconstrained linear projections M×N, trained only by the forecasting loss, really group keys and values into meaningful super-clusters, with no clustering constraint, locality prior, or assignment loss; if that tie is absent, the formula reduces to a generic low-rank projection like Linformer and the claimed derivation and the reason M=4 wor

What would settle it

One concrete observation: compare the learned C_k rows against the sensor adjacency graph. If the C_k rows do not concentrate weight on geographically close or strongly connected sensors, and if a randomly initialized C_k with the same number of parameters matches EMAGN's accuracy, then the Gaussian-filtering derivation is not doing the work — the model is just a low-rank attention. Also, a direct numerical check: for a small N, compare Eq. (2)'s Taylor-approximated attention with the true softmax attention for random Q, K, V; if the approximation error does not shrink as M grows toward N, the

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

If this is right

  • Traffic forecasting with attention becomes practical on larger networks: complexity scales linearly with sensor count N, with M fixed and small (M=4), so city-scale deployments with thousands of sensors remain feasible on a single GPU.
  • Attention heads cease to be a memory bottleneck: with K=16 heads GMAN exceeds an 11 GB GPU while EMAGN runs at 8,568 MB, so multi-head configurations previously impossible become available.
  • Deeper models are affordable: at every STBlock depth L, EMAGN uses roughly 58% less memory and trains 23–32% faster than GMAN, so deeper spatial-temporal stacks fit in the same hardware budget.
  • Learned clustering, rather than any linear-attention scheme, is the source of the efficiency–accuracy balance: within the same GMAN backbone EMAGN beats Linformer and Performer on both accuracy and efficiency.
  • The gap to full attention narrows as the forecast horizon grows (5.1% to 3.2% MAE on METR-LA), suggesting long-range forecasting benefits most from the clustered global structure.

Where Pith is reading between the lines

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

  • If the low-rank spatial structure claim holds, the same learned-clustering attention could be applied to other graph or point-cloud tasks where nodes have spatial organisation, not just road networks — the mechanism does not depend on traffic-specific embeddings.
  • A testable extension: visualize the learned C_k rows against sensor coordinates to check whether they correspond to geographic corridors; the paper does not report this analysis, and it would directly support or weaken the Gaussian-filtering interpretation.
  • Because temporal attention is still quadratic in sequence length T, the method's linearity applies to the spatial dimension; extending the same learned-clustering idea to long historical sequences is the paper's own stated future work and would be the natural next step.
  • The reported efficiency numbers are measured at batch size 16 and on one GPU; a stress test at larger batch sizes or with longer sequences would show whether the memory savings persist when the temporal attention term becomes non-negligible.

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 / 4 minor

Summary. The paper proposes EMAGN, a traffic-forecasting model whose main claimed contribution is a linear-complexity spatial self-attention mechanism derived from fast high-dimensional Gaussian filtering. The mechanism replaces full N×N attention with a low-rank form in which learnable matrices C_k and C_v project keys and values to M "super-clusters," giving complexity O(N M d). The authors build on GMAN, compare against full attention and against Linformer and Performer in a common backbone, and report accuracy within 2.7–3.2% MAE of GMAN at the 60-minute horizon while reducing training time, inference time, and GPU memory substantially. They also provide scaling experiments varying depth, heads, and network size, and an ablation over M. The empirical resource measurements are plausible and internally consistent, but the central methodological novelty collapses because Eq. (5) is mathematically identical to Linformer's low-rank projection attention, and no clustering constraint or tie to the Gaussian-filtering derivation is supplied.

Significance. If the mechanism were genuinely new and its derivation valid, the paper would provide a useful drop-in linear attention for spatio-temporal graphs, with careful memory and runtime measurements. The authors deserve credit for reporting GPU usage, training/inference times, and scaling plots, and for including a same-backbone comparison with Linformer and Performer. However, the central claim of a new clustering-based linear attention is not supported: the equation defining EMAGN's spatial attention is already known as Linformer. The efficiency results could support an engineering-oriented benchmark, but they do not establish a new attention mechanism. Given that the paper's title, abstract, and contributions rest on this derivation, the core contribution is not sound.

major comments (3)
  1. [Section 3.1, Eq. (5)] The proposed mechanism is mathematically identical to Linformer's low-rank projection attention. Linformer uses \hat X = Softmax(Q (E K)^T / sqrt(d)) F V with learned E,F; Eq. (5) has exactly the same form with learned C_k,C_v in place of E,F. The Gaussian-filtering derivation in Eqs. (1)–(2) is never used to constrain, initialize, or regularize C_k and C_v: there is no clustering constraint, assignment loss, locality prior, or link between the Taylor-expansion cluster summaries of Eq. (2) and a left multiplication by an unconstrained matrix. Thus the terms "clustering matrices," "super-clusters," and "traffic-network-aware adaptive clustering" are not justified; M is simply the projection rank. This invalidates the claimed novelty and the interpretation of the M-ablation.
  2. [Section 2.2 and Table 3] The text states that Linformer applies "fixed or random projection strategies" that treat positions uniformly. This is a mischaracterization: standard Linformer learns its projection matrices E and F. Since EMAGN's Eq. (5) is the same form, the 0.10 MAE difference in Table 3 cannot be attributed to learned clustering versus fixed projections. Without controlling for initialization, weight sharing, training details, or multiple seeds, the reported advantage over Linformer is an uncontrolled implementation difference, not evidence for a distinct mechanism.
  3. [Section 4.5 and Table 4] The ablation is presented as the "number of clusters," but because C_k and C_v are unconstrained projection matrices, varying M is a projection-dimension ablation, not a clustering analysis. The claim that "M = 4 matches or exceeds M = 128" is not consistently supported across all metrics (e.g., 15-min RMSE is 5.79 for M=4 vs. 5.77 for M=128). More importantly, no error bars or multiple-seed results are reported anywhere in Tables 2–4, so the small accuracy differences used to support algorithmic conclusions are within unknown noise. This weakens the inference that traffic graphs have inherent low-rank spatial structure discoverable by "learned clustering."
minor comments (4)
  1. [Abstract and Section 4.4] The abstract's "2.7–3.2% MAE" claim should state that this refers to the 60-minute horizon and to relative MAE; Table 2 shows larger relative gaps at 15 and 30 minutes (e.g., 5.1% at 15 min on METR-LA).
  2. [Section 3.1, Eq. (2)] The notation P_r and the "r-th order Taylor terms" are not defined. The derivation would need a self-contained statement of the approximation error and how it depends on the chosen cluster centers before it can be used as an analogy for Eq. (3).
  3. [Section 4.2] The experimental methodology is not fully controlled: most baselines are quoted from their original papers, and only GMAN is re-run with published code. A statement about identical hardware/software conditions for the compared numbers is needed, and ideally code or detailed configuration files should be released.
  4. [Section 4.4 and Table 4] The phrase "M = 4 matches or exceeds M = 128" is too strong given that some metrics favor M=128. Please qualify with metric-specific comparisons.

Circularity Check

2 steps flagged

Central 'learned clustering' attention is algebraically Linformer's projection attention, so the Gaussian-filtering derivation is a renaming rather than a derivation; M=4 claim is fit to the same ablation.

specific steps
  1. renaming known result [Section 3.1, Eq. (3); Section 3.2, Eq. (5); Section 2.2]
    "Following this analogy, we replace the hand-crafted clustering of classical filtering with two learnable clustering matrices Ck, Cv ∈ R^{M×N} ... The resulting linear-complexity self-attention is X_hat = Softmax(αQ(CkK)^T)CvV. (3) ... EMAGN learns separate clustering matrices Ck and Cv end-to-end that can adapt to the spatial structure of the traffic network."

    Because Ck and Cv are unconstrained learned linear maps, Eq. (5) is exactly Linformer attention, Softmax(Q(EK)^T)FV with E=Ck and F=Cv. No clustering loss, locality prior, or assignment constraint connects Eq. (2)'s Taylor cluster summaries to these matrices, so the claimed derivation does not produce anything beyond the known low-rank projection mechanism. The paper's contrast with Linformer rests on the incorrect statement that Linformer uses fixed/random projections; the claimed novelty reduces by construction to an existing method.

  2. fitted input called prediction [Section 4.3 and Section 4.5 / Table 4]
    "EMAGN uses M=4 clusters for METR-LA, determined by the ablation in Section 4.5. ... Ablation studies confirm that as few as 4–8 learned clusters suffice to approximate full pairwise attention, validating the design principle derived from fast Gaussian filtering."

    The value M=4 is chosen from Table 4, and the same Table 4 is then cited as 'confirming' that few clusters suffice. The sufficiency claim is therefore a post-hoc reading of the same ablation used to select the hyperparameter, not an independently predicted or externally validated property. This is mild: hyperparameter selection is transparent, but presenting it as validation of the derivation is circular.

full rationale

EMAGN's efficiency measurements, ablations, and comparisons are self-contained and independently measurable, so the paper is not globally circular. However, the central methodological derivation collapses: the paper claims Eq. (3)/(5) is a new linear-complexity attention derived from fast Gaussian filtering with learned clustering, but the unconstrained matrices Ck and Cv make the formula algebraically identical to Linformer's learned low-rank projection attention. The Gaussian-filtering Taylor expansion (Eq. (2)) is never shown to imply, constrain, or be preserved by arbitrary learned linear projections, so the derivation is an unsupported substitution; calling the projections 'clustering matrices' is a renaming of an existing mechanism. The M=4 claim is additionally fit to the same ablation that selects it, though this is a secondary issue. There are no load-bearing self-citations. Overall score 4: the central novelty reduces to a known construction by renaming, but the empirical efficiency content remains independent.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 1 invented entities

The central 'derivation' rests on an analogy, not a mathematical implication: after Eq. (2), the paper replaces actual cluster centers with learned matrices without constraints. The remaining claims are empirical, resting on standard benchmark assumptions, post-hoc hyperparameter selection, and a mischaracterization of the Linformer baseline.

free parameters (3)
  • M (number of super-clusters) = 4 (selected via ablation on METR-LA)
    Controls the low-rank bottleneck. M=4 is chosen by validation performance in §4.5 and then used as default; it is not derived from the Gaussian filtering theory.
  • C_k, C_v clustering matrices = learned end-to-end, no constraints
    The central 'clustering' is implemented as unconstrained M×N linear projections. Their values are fit to the forecasting objective; no clustering normalization, locality prior, or assignment loss enforces the super-cluster interpretation.
  • Projection dimension m=4 for Linformer/Performer baselines = 4
    The comparison in §4.4 sets Linformer and Performer projection dimension equal to M. If that setting is suboptimal for these baselines, the claimed advantage of EMAGN may be partly an artifact of the configuration.
axioms (4)
  • ad hoc to paper Fast Gaussian filtering can be generalized by replacing hand-crafted cluster centers with arbitrary learned linear projections.
    Used to justify Eq. (3) in §3.1. Eq. (2)'s Taylor expansion requires actual cluster centers and locality; C_k and C_v encode neither, so the mathematical link is not established.
  • domain assumption Traffic sensor graphs have inherent low-rank spatial structure, so M=4 clusters suffice.
    Invoked in §4.5 to explain why small M works. The only support is the same ablation used to select M, which is post-hoc rather than independent evidence.
  • domain assumption Linformer and Performer use fixed or random projections that treat all positions uniformly.
    This premise underpins the claimed advantage of learned clustering in §4.4. It is inaccurate for Linformer, whose projection matrices are learned, and it mischaracterizes the baseline.
  • domain assumption The fixed 12-step temporal sequence length makes temporal attention's O(T^2) cost negligible.
    Used in §3.2 to justify linearizing only spatial attention. Valid only under the fixed short-horizon benchmark setting, not for longer temporal sequences.
invented entities (1)
  • Super-key / super-value vectors (super-clusters) no independent evidence
    purpose: Compress N keys and values into M summaries so attention cost becomes O(N M d).
    These are outputs of unconstrained learned linear projections C_k and C_v. No clustering loss, locality constraint, or cluster-center interpretation is imposed, so the 'super-cluster' entity is a relabeling of low-rank projected keys/values rather than an independently evidenced structure.

pith-pipeline@v1.3.0-alltime-deepseek · 12298 in / 13963 out tokens · 141454 ms · 2026-08-02T05:45:57.705880+00:00 · methodology

0 comments
read the original abstract

Traffic forecasting is highly challenging due to complex and nonlinear spatial and temporal dependencies. Self-attention mechanisms have been widely adopted to model dynamic and long-range dependencies, achieving state-of-the-art performance, but suffer from limited scalability due to quadratic computational and memory complexity. To address this, we propose an Efficient Multi-Attention Graph Network (EMAGN) that linearises the spatial attention mechanism itself, inspired by the theory of fast high-dimensional Gaussian filtering. Two learned clustering matrices C_k and C_v adaptively group key and value vectors into M super-clusters, reducing complexity from O(N^2 d) to O(NMd) without sacrificing the flexibility of attention for dynamic dependency modelling. Experimental results on PEMS-BAY and METR-LA show that EMAGN achieves accuracy within 2.7-3.2% MAE of full-attention GMAN while reducing training time by 32%, inference time by 38%, and GPU memory by 58%. Critically, at K=16 attention heads, full-attention GMAN runs out of memory on a standard 11 GB GPU entirely while EMAGN continues to operate, demonstrating a categorical expansion of feasible model configurations. EMAGN also surpasses Linformer and Performer in both accuracy and efficiency within the same backbone, owing to its traffic-network-aware adaptive clustering.

Figures

Figures reproduced from arXiv: 2607.13241 by Chengqi Lu, Junwei You, Ke Liu, Lyuhao Chen, Mingxing Xu, Oliver Gao, Rakesh Chowdary Machineni, Xiangyu Li, Xi Cheng, Xin Hu.

Figure 1
Figure 1. Figure 1: Illustration of the attention mechanisms. (a) Standard self-attention with [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The architecture of the proposed EMAGN. Left: overall encoder-decoder architecture with stacked STBlocks [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Scalability with number of attention heads [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Scalability with traffic network size N on METR-LA (L = 5, K = 8, M = 4 for EMAGN). widens significantly and follows the predicted pattern: GMAN’s costs grow steeply while EMAGN’s costs follow a near-linear trajectory, directly validating the theoretical O(NM) complexity. At N = 207, EMAGN requires 32% less training time and 58% less memory. For deployment on larger real-world networks, city-scale systems … 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

84 extracted references · 6 linked inside Pith

  1. [1]

    Frontiers in Handwriting Recognition (ICFHR), 2014 14th International Conference on , pages=

    Real-time segmentation of on-line handwritten arabic script , author=. Frontiers in Handwriting Recognition (ICFHR), 2014 14th International Conference on , pages=. 2014 , organization=

    2014

  2. [2]

    Soft Computing and Pattern Recognition (SoCPaR), 2014 6th International Conference of , pages=

    Fast classification of handwritten on-line Arabic characters , author=. Soft Computing and Pattern Recognition (SoCPaR), 2014 6th International Conference of , pages=. 2014 , organization=

    2014

  3. [3]

    arXiv preprint arXiv:1804.09028 , year=

    Estimate and Replace: A Novel Approach to Integrating Deep Neural Networks with Existing Applications , author=. arXiv preprint arXiv:1804.09028 , year=

    Pith/arXiv arXiv

  4. [4]

    Transportmetrica A: Transport Science , volume=

    A review of travel time estimation and forecasting for advanced traveller information systems , author=. Transportmetrica A: Transport Science , volume=. 2015 , publisher=

    2015

  5. [5]

    Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting , address=

    Li, Yaguang and Yu, Rose and Shahabi, Cyrus and Liu, Yan , Booktitle =. Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting , address=

  6. [6]

    A simple and effective method for predicting travel times on freeways , Volume =

    J. A simple and effective method for predicting travel times on freeways , Volume =. doi:10.1109/TITS.2004.833765 , Issn =

    arXiv 2004

  7. [7]

    Travel-time prediction with support vector regression , Volume =

    Wu, Chun-Hsin and Ho, Jan-Ming and Lee, Der-Tsai , Journal = IEEE_J_ITS, Number =. Travel-time prediction with support vector regression , Volume =

  8. [8]

    Long Short-Term Memory , Volume =

    Hochreiter, Sepp and Schmidhuber, J\". Long Short-Term Memory , Volume =. doi:10.1162/neco.1997.9.8.1735 , Journal =

    doi:10.1162/neco.1997.9.8.1735 1997

  9. [9]

    NIPS 2014 Workshop Deep Learn

    Chung, Junyoung and Gulcehre, Caglar and Cho, KyungHyun and Bengio, Yoshua , Title =. NIPS 2014 Workshop Deep Learn. , address=

    2014

  10. [10]

    Mostafa Dehghani and Stephan Gouws and Oriol Vinyals and Jakob Uszkoreit and Lukasz Kaiser , Title =. 7th Int. Conf. Learn. Rep. (ICLR) , address=

  11. [11]

    Convolutional neural networks on graphs with fast localized spectral filtering , address=

    Defferrard, Micha. Convolutional neural networks on graphs with fast localized spectral filtering , address=. Adv. Neural Inf. Process. Syst. 29 , Pages =

  12. [12]

    and Towsley, D

    Atwood, J. and Towsley, D. , Booktitle =. Diffusion-convolutional neural networks , address=

  13. [13]

    Li, Yujia and Tarlow, Daniel and Brockschmidt, Marc and Zemel, Richard , Title =. Int. Conf. Learn. Rep. 2016 , address=

    2016

  14. [14]

    Neural message passing for quantum chemistry , author=. Proc. 34th Int. Conf. Mach. Learn. , pages=

  15. [15]

    2014 , month=oct, publisher=

    Deep architecture for traffic flow prediction: Deep belief networks with multitask learning , author=. 2014 , month=oct, publisher=

    2014

  16. [16]

    Traffic flow prediction with big data: A deep learning approach , author=

  17. [17]

    Deep spatio-temporal residual networks for citywide crowd flows prediction , author=. Proc. 31st AAAI Conf. Artif. Intell. , pages=

  18. [18]

    Convolutional sequence to sequence learning , author=. Proc. 34th Int. Conf. Mach. Learn. , pages=

  19. [19]

    Semi-Supervised Classification with Graph Convolutional Networks , author=. 5th Int. Conf. Learn. Rep. (ICLR) , address=

  20. [20]

    A Field Guide to Dynamical Recurrent Neural Networks , pages=

    Gradient flow in recurrent nets: The difficulty of learning long-term dependencies , author=. A Field Guide to Dynamical Recurrent Neural Networks , pages=. 2001 , publisher=

    2001

  21. [21]

    Attention Based Spatial-Temporal Graph Convolutional Networks for Traffic Flow Forecasting , author=. Proc. 33rd AAAI Conf. Artif. Intell. , pages=

  22. [22]

    , title =

    Krizhevsky, Alex and Sutskever, Ilya and Hinton, Geoffrey E. , title =. Adv. Neural Inf. Process. Syst. 25 , month = dec, year =

  23. [23]

    2015 IEEE Conf

    Fully convolutional networks for semantic segmentation , author=. 2015 IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR) , pages=

    2015

  24. [24]

    Long short-term memory neural network for traffic speed prediction using remote microwave sensor data , author=. Transp. Res. Part C Emerg. Technol. , volume=. 2015 , publisher=

    2015

  25. [25]

    arXiv preprint arXiv:1612.01022 , year=

    Short-term traffic flow forecasting with spatial-temporal correlation in a hybrid deep learning framework , author=. arXiv preprint arXiv:1612.01022 , year=

    Pith/arXiv arXiv

  26. [26]

    Real-time road traffic prediction with spatio-temporal correlations , author=. Transp. Res. Part C Emerg. Technol. , volume=. 2011 , publisher=

    2011

  27. [27]

    Proceedings of the 27th International Joint Conference on Artificial Intelligence , pages=

    Spatio-temporal graph convolutional networks: a deep learning framework for traffic forecasting , author=. Proceedings of the 27th International Joint Conference on Artificial Intelligence , pages=

  28. [28]

    New Bayesian combination method for short-term traffic flow forecasting , author=. Transp. Res. Part C Emerg. Technol. , volume=. 2014 , publisher=

    2014

  29. [29]

    Discovering spatio-temporal causal interactions in traffic data streams , author=. Proc. 17th ACM SIGKDD Int. Conf. Knowl. Disc. Data Min. , pages=

  30. [30]

    Graph attention networks , author=. 6th Int. Conf. Learn. Rep. (ICLR) , address=

  31. [31]

    Effective Deep Memory Networks for Distant Supervised Relation Extraction , author=. Proc. 26th Int. Joint Conf. Artif. Intell. , pages=

  32. [32]

    Dynamic spatial-temporal graph convolutional neural networks for traffic forecasting , author=. Proc. 33rd AAAI Conf. Artif. Intell. , pages=

  33. [33]

    Devlin, Jacob and Chang, Ming-Wei and Lee, Kenton and Toutanova, Kristina , booktitle=

  34. [34]

    Sequence to sequence learning with neural networks , author=. Adv. Neural Inf. Process. Syst. 27 , pages=

  35. [35]

    1997 , publisher=

    Makridakis, Spyros and Hibon, Michele , journal=. 1997 , publisher=

    1997

  36. [36]

    GMAN: A Graph Multi-Attention Network for Traffic Prediction

    Chuanpan Zheng and Xiaoliang Fan and Cheng Wang and Jianzhong Qi. GMAN: A Graph Multi-Attention Network for Traffic Prediction. AAAI. 2020

    2020

  37. [37]

    Proceedings of the 29th ACM international conference on information & knowledge management , pages=

    ST-GRAT: A novel spatio-temporal graph attention networks for accurately forecasting dynamically changing road speed , author=. Proceedings of the 29th ACM international conference on information & knowledge management , pages=

  38. [38]

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

    Graph WaveNet for Deep Spatial-Temporal Graph Modeling , author=. Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence , pages=. 2019 , organization=

    2019

  39. [39]

    Proceedings of The Web Conference 2020 , pages=

    Traffic Flow Prediction via Spatial Temporal Graph Neural Network , author=. Proceedings of The Web Conference 2020 , pages=

    2020

  40. [40]

    GaAN: Gated Attention Networks for Learning on Large and Spatiotemporal Graphs , booktitle =

    Jiani Zhang and Xingjian Shi and Junyuan Xie and Hao Ma and Irwin King and Dit. GaAN: Gated Attention Networks for Learning on Large and Spatiotemporal Graphs , booktitle =

  41. [41]

    IEEE Transactions on Intelligent Transportation Systems , year=

    T-gcn: A temporal graph convolutional network for traffic prediction , author=. IEEE Transactions on Intelligent Transportation Systems , year=

  42. [42]

    arXiv preprint arXiv:1803.07294 , year=

    Gaan: Gated attention networks for learning on large and spatiotemporal graphs , author=. arXiv preprint arXiv:1803.07294 , year=

    Pith/arXiv arXiv

  43. [43]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Spatial-Temporal Synchronous Graph Convolutional Networks: A New Framework for Spatial-Temporal Network Data Forecasting , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  44. [45]

    Proceedings of the IEEE International Conference on Computer Vision , pages=

    Attention augmented convolutional networks , author=. Proceedings of the IEEE International Conference on Computer Vision , pages=

  45. [46]

    Advances in Neural Information Processing Systems , volume=

    Adaptive graph convolutional recurrent network for traffic forecasting , author=. Advances in Neural Information Processing Systems , volume=

  46. [47]

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

    Connecting the dots: Multivariate time series forecasting with graph neural networks , author=. Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , pages=

  47. [48]

    Jiang, Jiawei and Han, Chengkai and Zhao, Wayne Xin and Wang, Jingyuan , booktitle=

  48. [49]

    Proceedings of the VLDB Endowment , volume=

    Bigst: Linear complexity spatio-temporal graph neural network for traffic forecasting on large-scale road networks , author=. Proceedings of the VLDB Endowment , volume=. 2024 , publisher=

    2024

  49. [50]

    arXiv preprint arXiv:2410.00385 , year=

    Stgformer: Efficient spatiotemporal graph transformer for traffic forecasting , author=. arXiv preprint arXiv:2410.00385 , year=

    Pith/arXiv arXiv

  50. [51]

    arXiv preprint arXiv:2006.04768 , year=

    Linformer: Self-attention with linear complexity , author=. arXiv preprint arXiv:2006.04768 , year=

    Pith/arXiv arXiv 2006

  51. [52]

    2022 , eprint=

    Rethinking Attention with Performers , author=. 2022 , eprint=

    2022

  52. [53]

    Nguyen, Tan and Nguyen, Thanh and Ngo, Dung and Nguyen, Duy and Bui, Hung , booktitle=

  53. [54]

    Xiong, Yunyang and Zeng, Zhanpeng and Chakraborty, Rudrasis and Tan, Mingxing and Fung, Glenn and Li, Yin and Singh, Vikas , booktitle=. Nystr

  54. [55]

    IEEE Transactions on Intelligent Transportation Systems , volume=

    Short-term traffic flow forecasting: An experimental comparison of time-series analysis and supervised learning , author=. IEEE Transactions on Intelligent Transportation Systems , volume=. 2013 , publisher=

    2013

  55. [56]

    Advances in neural information processing systems , volume=

    Fmmformer: Efficient and flexible transformer via decomposed near-field and far-field attention , author=. Advances in neural information processing systems , volume=

  56. [57]

    The fast

    Greengard, Leslie and Strain, John , journal=. The fast

  57. [58]

    and Davis, Larry , booktitle=

    Yang, Changjiang and Duraiswami, Ramani and Gumerov, Nail A. and Davis, Larry , booktitle=. Improved fast

  58. [59]

    Computer Graphics Forum (Eurographics) , volume=

    Fast high-dimensional filtering using the permutohedral lattice , author=. Computer Graphics Forum (Eurographics) , volume=

  59. [60]

    Proceedings of the IEEE International Conference on Computer Vision (ICCV) , pages=

    Conditional random fields as recurrent neural networks , author=. Proceedings of the IEEE International Conference on Computer Vision (ICCV) , pages=

  60. [61]

    Advances in Neural Information Processing Systems 29 , pages=

    Diffusion-convolutional neural networks , author=. Advances in Neural Information Processing Systems 29 , pages=

  61. [62]

    NIPS 2014 Workshop on Deep Learning , address=

    Empirical evaluation of gated recurrent neural networks on sequence modeling , author=. NIPS 2014 Workshop on Deep Learning , address=

    2014

  62. [63]

    Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    BERT: Pre-training of deep bidirectional transformers for language understanding , author=. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    2019

  63. [64]

    Proceedings of the 33rd AAAI Conference on Artificial Intelligence , pages=

    Dynamic spatial-temporal graph convolutional neural networks for traffic forecasting , author=. Proceedings of the 33rd AAAI Conference on Artificial Intelligence , pages=

  64. [65]

    Proceedings of the 34th International Conference on Machine Learning , pages=

    Neural message passing for quantum chemistry , author=. Proceedings of the 34th International Conference on Machine Learning , pages=

  65. [66]

    Proceedings of the 33rd AAAI Conference on Artificial Intelligence , pages=

    Attention based spatial-temporal graph convolutional networks for traffic flow forecasting , author=. Proceedings of the 33rd AAAI Conference on Artificial Intelligence , pages=

  66. [67]

    Neural Computation , volume=

    Long short-term memory , author=. Neural Computation , volume=

  67. [68]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Gman: A graph multi-attention network for traffic prediction , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  68. [69]

    6th International Conference on Learning Representations , address=

    Diffusion convolutional recurrent neural network: Data-driven traffic forecasting , author=. 6th International Conference on Learning Representations , address=

  69. [70]

    Journal of Forecasting , volume=

    ARMA models and the Box-Jenkins methodology , author=. Journal of Forecasting , volume=

  70. [71]

    Advances in Neural Information Processing Systems 27 , pages=

    Sequence to sequence learning with neural networks , author=. Advances in Neural Information Processing Systems 27 , pages=

  71. [72]

    Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining , pages=

    Urban traffic prediction from spatio-temporal data using deep meta learning , author=. Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining , pages=

  72. [73]

    arXiv preprint arXiv:2001.02908 , year=

    Spatial-temporal transformer networks for traffic flow forecasting , author=. arXiv preprint arXiv:2001.02908 , year=

    Pith/arXiv arXiv 2001

  73. [74]

    Proceedings of the Thirty-Fourth Conference on Uncertainty in Artificial Intelligence , pages=

    GAAN: Gated Attention Networks for Learning on Large and Spatiotemporal Graphs , author=. Proceedings of the Thirty-Fourth Conference on Uncertainty in Artificial Intelligence , pages=

  74. [75]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Informer: Beyond efficient transformer for long sequence time-series forecasting , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  75. [76]

    Plos one , volume=

    Spatio-temporal transformer traffic prediction network based on multi-level causal attention , author=. Plos one , volume=. 2025 , publisher=

    2025

  76. [77]

    Proceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining , pages=

    node2vec: Scalable feature learning for networks , author=. Proceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining , pages=

  77. [78]

    2020 IEEE International Conference on Data Mining (ICDM) , pages=

    STGCN: a spatial-temporal aware graph learning method for POI recommendation , author=. 2020 IEEE International Conference on Data Mining (ICDM) , pages=. 2020 , organization=

    2020

  78. [79]

    Proceedings of the 31st ACM international conference on information & knowledge management , pages=

    Spatial-temporal identity: A simple yet effective baseline for multivariate time series forecasting , author=. Proceedings of the 31st ACM international conference on information & knowledge management , pages=

  79. [80]

    IEEE Transactions on Knowledge and Data Engineering , volume=

    Flow prediction in spatio-temporal networks based on multitask deep learning , author=. IEEE Transactions on Knowledge and Data Engineering , volume=. 2019 , publisher=

    2019

  80. [81]

    sensors , volume=

    Learning traffic as images: A deep convolutional neural network for large-scale transportation network speed prediction , author=. sensors , volume=. 2017 , publisher=

    2017

Showing first 80 references.