Pith. sign in

REVIEW 3 major objections 6 minor 44 references

Virtual Nodes Improve Long-term Traffic Prediction

T0 review · 3 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper shows that adding virtual nodes—extra graph nodes linked to every real node—lets spatio-temporal GNNs aggregate global traffic information in one layer, cutting long-horizon RMSE by ~6.27% with a semi-adaptive adjacency matrix.

desk verdict The paper's headline 6.27% RMSE gain is likely inflated by test-set selection; the idea is sensible but the evidence needs a proper validation split. read the letter →

arxiv 2501.10048 v1 pith:F527E2KA submitted 2025-01-17 cs.LG cs.AI

Xiaoyang Cao , Dingyi Zhuang , Jinhua Zhao , Shenhao Wang This is my paper · ORCID
classification cs.LGcs.AI
keywords trafficpredictionlong-termgraphneuralnetworksvirtualnodesover-squashingsemi-adaptiveadjacencymatrixspatio-temporalforecastingexplainability
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Long-term traffic prediction with spatio-temporal graph neural networks suffers because message passing compresses information from distant nodes, a problem known as over-squashing. The paper proposes adding virtual nodes, which connect to every real node, so that global information can be gathered in a single GNN layer. By combining a geographic distance-based adjacency matrix with learned virtual-node weights, the best configuration (10 virtual nodes) reduces RMSE by about 6.27% and MAPE by about 5.04% at the 75–100 minute horizon relative to the distance-based baseline. The learned virtual-node weights also concentrate on major intersections and high-traffic areas, which gives the model an interpretable spatial focus.

What carries the argument

Virtual nodes are additional graph nodes connected to every real node, so a single message-passing step spreads information across the entire graph, bypassing the multi-hop paths that cause over-squashing. The semi-adaptive adjacency matrix fuses the distance-based adjacency $A_{\text{dist}}$ with learned blocks drawn from an adaptive matrix $A_{\text{adapt}} = \operatorname{ReLU}(E_1 E_2^\top - E_2 E_1^\top)$, where $E_1, E_2$ are learnable node embeddings; weak connections are pruned by a threshold. This construction lets the model integrate global dependencies while preserving geographic locality, and the learned virtual-node weights double as an explainability tool that highlights traffic-critical regions.

What would settle it

Rerun the same experimental setup with a proper three-way train/validation/test split, choose the virtual-node count on the validation set, and evaluate on a test set never seen during model selection; if the 75–100 minute RMSE advantage of Semi-10 V.N. over the distance-based baseline shrinks or vanishes, the central claim is weakened. Alternatively, directly measure pair-wise sensitivity between distant nodes to test whether over-squashing is actually reduced.

Watch

Extended reading notes

Core claim

The central claim is that virtual nodes, when attached to all real nodes of a traffic network, alleviate the over-squashing bottleneck that restricts long-range information flow in ST-GNNs. The paper introduces a semi-adaptive adjacency matrix $A_{\text{semi}} = \begin{pmatrix} A_{\text{dist}} & A_{\text{adapt, real-to-virtual}} \\ A_{\text{adapt, virtual-to-real}} & A_{\text{adapt, virtual nodes}} \end{pmatrix}$ that keeps the fixed distance-based graph while learning virtual-to-real connection weights from node embeddings. Using STGCN as the base model on the San Diego LargeST sub-dataset with 716 sensors, the Semi-10 V.N. configuration reaches an RMSE of 42.32 and a MAPE of 0.1735 in the 75–100 minute average horizon, consistently beating the distance-based, all-ones, and fully adaptive alternatives, and the gap widens as the prediction horizon grows.

Load-bearing premise

The reported improvement assumes the test set used to report errors was not used to choose the number of virtual nodes; the paper does not describe a validation split, so the headline 6.27% gain could be inflated by selection if the same test data guided the choice.

Editorial extensions

If this is right

  • The virtual-node augmentation only changes the adjacency matrix, so existing ST-GNNs that accept an adjacency matrix as input can adopt the method without architectural redesign.
  • The performance gap between semi-adaptive and distance-based models grows with prediction horizon, suggesting that global aggregation becomes more valuable as forecasts extend further into the future.
  • The learned virtual-node weights identify key intersections and high-traffic areas, giving transportation planners a visual diagnostic of which locations most influence long-range forecasts.
  • Fully adaptive configurations underperform the distance-based baseline, implying that geographic structure should be retained rather than discarded when adding learned connectivity.

Reading between the lines

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

  • The mechanism is likely to transfer beyond traffic to other spatio-temporal graph forecasting tasks (e.g., energy demand, crowd flow) where long-range dependencies matter and a meaningful graph distance exists; the paper tests only the San Diego traffic dataset.
  • The optimal virtual-node count of 10 may depend on graph size and density, so a useful extension would relate this hyperparameter to graph diameter or average shortest-path length rather than tuning it per dataset.
  • Because adaptive-only virtual nodes perform worse than the distance baseline, a controlled ablation could isolate whether the gain comes from virtual nodes per se or specifically from the combination of learned and geographic information.
  • The paper does not describe a separate validation split, so an independent evaluation that picks the virtual-node count on a held-out set would clarify whether the reported improvement is robust or partly an artifact of selection.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The manuscript proposes to improve long-term traffic prediction by augmenting an STGCN backbone with virtual nodes. The method constructs a semi-adaptive adjacency matrix that combines a distance-based matrix for real nodes with learned adaptive blocks connecting real and virtual nodes, and the authors evaluate adaptive and semi-adaptive variants with 1, 2, 5, 10, and 20 virtual nodes on the San Diego subset of the LargeST benchmark. The paper reports that the Semi-10 configuration achieves a 6.27% RMSE reduction and a 5.04% MAPE reduction on the 75-100 minute average horizon relative to a distance-based baseline, and it qualitatively claims that virtual nodes mitigate the over-squashing problem and improve explainability via heat-map visualizations.

Significance. If the evaluation were robust, the contribution would be a simple, architecture-agnostic graph augmentation that improves long-horizon traffic forecasting and yields interpretable connection weights. The use of a public benchmark, a fixed STGCN backbone, and an explicit all-ones virtual-node baseline are strengths, and the adaptive-versus-semi-adaptive comparison is a reasonable ablation. However, the central empirical claim currently rests on a single test-set comparison without validation-based model selection, error bars, or significance tests, and the over-squashing claim is not directly measured. The significance of the paper therefore depends on whether the authors can substantiate the reported gain under a proper evaluation protocol.

major comments (3)
  1. [§5.3, Table 2] The reported 6.27% RMSE reduction is selected as the best among the 12 configurations in Table 2, and the text states only that years 2019–2020 are used for training and testing, with no validation split or model-selection procedure described. If the same test set was used to choose the virtual-node count, the headline improvement is the minimum of a family of errors and will be inflated by selection even under a null hypothesis. The peaked pattern in Table 2 (Semi-5 worse than the distance-based baseline, Semi-10 best, Semi-20 worse again) is consistent with test-set selection. Please re-run the comparison with a validation split for choosing the number of virtual nodes, and report the performance of the selected configuration on a held-out test set, or otherwise demonstrate that the choice of Semi-10 was not made on the evaluation set.
  2. [§5.3, Table 2] No error bars, multiple seeds, or significance tests are reported for any configuration. Given that the differences between Semi-10 and the distance-based baseline at Horizon 5 are about 1.25 RMSE (3.5%) and at Horizon 20 about 2.51 RMSE (5.5%), a single-run comparison cannot establish that the improvement is real rather than run-to-run noise. Please report the mean and standard deviation over at least three random seeds, and ideally a paired significance test, for the configurations in Table 2.
  3. [§5.4] The claim that virtual nodes 'improve layer-wise sensitivity to mitigate the over-squashing problem' is not supported by any quantitative sensitivity metric. The sensitivity analysis in Section 5.4 and Figure 6 reports RMSE and MAPE versus the number of virtual nodes, which is a prediction-accuracy analysis, not a measure of over-squashing (e.g., Jacobian sensitivity, information-theoretic bottleneck measures, or receptive-field coverage). Please either add a direct measurement of sensitivity or over-squashing, or revise the claim to a hypothesis rather than a demonstrated result.
minor comments (6)
  1. [§3.1] The phrase 'theover-squashing problem' is missing a space and should read 'the over-squashing problem.'
  2. [§1] The sentence 'The rest of the paper is organized as follow' should be 'as follows.'
  3. [§5.3] The manuscript does not report the hyperparameters needed for reproducibility, including the embedding dimension d, the threshold r, the number of STGCN layers, learning rate, batch size, number of epochs, and the optimization algorithm. Please add a reproducibility section or a detailed hyperparameter table.
  4. [§4.2, Eq. (6)] The adaptive adjacency matrix is computed over the full set of real and virtual nodes, but the text does not explain how the threshold r is chosen or how the sparsity of the resulting matrix changes with the number of virtual nodes nv. Please clarify the threshold selection procedure.
  5. [§5.5] The heat map in Figure 7a is described as showing stronger connections with darker colors, but no color bar is provided, which makes the visual comparisons difficult to interpret.
  6. [§5.2] The 'All-ones' baseline is described by setting A_adapt,real_to_virtual = 1 and A_adapt,virtual_to_real = 1, but the text does not specify the value of A_adapt,virtual_nodes; please clarify whether it is zero, one, or something else.

Circularity Check

1 steps flagged · score 4.0 of 10

Headline 6.27% gain is selected as the best of 12 test-set configurations without a described validation split; the qualitative virtual-node claim retains independent support from the fixed All-ones baseline.

  1. fitted input called prediction [Section 5.3, Table 2 and the paragraph following it]
    "We utilized time series data from the years 2019 to 2020 for training and testing... Semi-10 V .N.consistently provides the lowest RMSE and MAPE... representing an RMSE reduction of approximately 6.27% and a MAPE reduction of approximately 5.04% compared to the distance-based baseline."

    As written, the experimental protocol selects the virtual-node count from 12 configurations using the same 2019-2020 test set on which the 6.27% improvement is then reported; no validation split or model-selection procedure is described. The reported gain is therefore the minimum of a family of test-set errors minus the baseline, a statistic that is biased in favor of virtual nodes even under a null effect. The claim that Semi-10 V.N. achieves this reduction is the result of the test-set argmin, not an independent prediction. The fixed All-ones baseline also improves (45.05 vs 45.15 RMSE long-term), so the qualitative structural claim retains some independent support; only the headline magnitude is forced by selection.

full rationale

The paper contains no mathematical derivation that could be circular; it is an empirical comparison of adjacency-matrix configurations on the LargeST SD dataset. The self-citations (Zhuang et al. 2022, Wu et al. 2021a) appear only as examples of prior ST-GNN applications and are not load-bearing. The over-squashing and layer-wise sensitivity claim is asserted qualitatively rather than quantified, which is an evidentiary gap but not a circular reduction. The only circularity concern is evaluative: Section 5.3 says the 2019-2020 data are used 'for training and testing' and then reports Semi-10 as the best of 12 configurations. Because no validation split or model-selection protocol is described, the reported 6.27% gain appears to be the minimum over configurations evaluated on the same test set used to report the result, making the headline magnitude a selected best-of-family statistic rather than an independent prediction. This inflates the evidence for the headline number, but it does not completely force the qualitative claim, since the fixed, non-adaptive All-ones virtual-node baseline also improves long-term RMSE (45.05 vs 45.15) over the distance-based baseline. With an explicit independent validation split the circularity score would be 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The central claim rests on the adaptive learning of virtual node connections, which introduces several free hyperparameters (nv, d, r) and STGCN training choices. The paper also relies on domain assumptions about the cause of long-term prediction difficulty and the validity of the distance-based graph. Virtual nodes are an invented modeling entity with no external evidential support beyond the paper's own experiments.

free parameters (4)
  • Number of virtual nodes (nv) = 10 (best of 1,2,5,10,20)
    Selected by comparing test-set RMSE and MAPE across configurations in Table 2; no separate validation set is described.
  • Embedding dimension d = not stated
    Hyperparameter for node embeddings E1 and E2 in Eq. 6; used in the adaptive adjacency construction, but its value is not reported.
  • Adaptive adjacency threshold r = not stated
    Threshold in Eq. 7 used to sparsify the adaptive adjacency; its value is not reported.
  • STGCN architecture hyperparameters = not stated
    Layer count, channel widths, kernel sizes, optimizer, learning rate, batch size, and training epochs are not given, so the experiment is under-specified.
assumptions (4)
  • domain assumption The over-squashing bottleneck (Alon and Yahav 2020) is the principal cause of degraded long-term traffic prediction in ST-GNNs.
    Invoked in Sections 1, 2.3, and 3.1 to motivate virtual nodes; the paper does not directly measure over-squashing.
  • domain assumption The thresholded Gaussian kernel adjacency from LargeST provides a valid distance-based graph for traffic flow.
    Used as Adist in Eq. 8 and as the baseline in Table 2; inherited from the benchmark.
  • domain assumption The uni-directional adaptive adjacency construction ReLU(E1 E2^T - E2 E1^T) is appropriate for traffic relations.
    Borrowed from Graph WaveNet (Wu et al. 2020), which the paper cites; the benefit for this task is assumed, and the fully adaptive variant performs worse in the experiments.
  • domain assumption Road network distance is a meaningful proxy for spatial dependence in the San Diego network.
    Used to predefine Adist; plausible but not tested against alternative similarity measures.
invented entities (1)
  • Virtual nodes (V_virtual)
    purpose: Extra graph vertices connected to all real nodes to aggregate global information in one layer and reduce over-squashing.
    They are a modeling construct added to the graph; their learned weights are internal to the model and only evidenced by in-model performance and post hoc heat maps, not by independent external measurements.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Virtual Nodes Improve Long-term Traffic Prediction." pith.science (2026). https://pith.science/paper/F527E2KA

@misc{pith2026250110048,
  author       = {Pith},
  title        = {Pith review of: Virtual Nodes Improve Long-term Traffic Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F527E2KA}},
  note         = {Machine review of arXiv:2501.10048}
}
read the original abstract

Effective traffic prediction is a cornerstone of intelligent transportation systems, enabling precise forecasts of traffic flow, speed, and congestion. While traditional spatio-temporal graph neural networks (ST-GNNs) have achieved notable success in short-term traffic forecasting, their performance in long-term predictions remains limited. This challenge arises from over-squashing problem, where bottlenecks and limited receptive fields restrict information flow and hinder the modeling of global dependencies. To address these challenges, this study introduces a novel framework that incorporates virtual nodes, which are additional nodes added to the graph and connected to existing nodes, in order to aggregate information across the entire graph within a single GNN layer. Our proposed model incorporates virtual nodes by constructing a semi-adaptive adjacency matrix. This matrix integrates distance-based and adaptive adjacency matrices, allowing the model to leverage geographical information while also learning task-specific features from data. Experimental results demonstrate that the inclusion of virtual nodes significantly enhances long-term prediction accuracy while also improving layer-wise sensitivity to mitigate the over-squashing problem. Virtual nodes also offer enhanced explainability by focusing on key intersections and high-traffic areas, as shown by the visualization of their adjacency matrix weights on road network heat maps. Our advanced approach enhances the understanding and management of urban traffic systems, making it particularly well-suited for real-world applications.

Figures

Figures reproduced from arXiv: 2501.10048 by the authors.

Figure 1
Figure 1. GNN with Virtual Node Augmentation Han et al. [2021], Kong et al. [2020], Yu et al. [2018], Shao et al. [2022]). This limitation arises because the information propagation speed between nodes in ST-GNNs is relatively slow, making it challenging to accurately predict long-term traffic states. Due to the nature of the GNN message-passing mechanism, each layer aggregates information only from neighboring nodes. As show… view at source ↗
Figure 2
Figure 2. Traffic Prediction Framework Incorporating Virtual Nodes [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Visualization of Sensor Locations in San Diego [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Adjacency Matrix Heat Map Despite the relatively large number of nodes in the SD dataset, its graph structure remains sparse, with a graph density of only 0.03, which inhibits effective information transfer across the graph. This sparsity is also evident from the adjac…
Figure 5
Figure 5. Figure 5: Long-term Performance of Virtual Nodes The "All-ones" configuration shows a marginal improvement in both the average and long-term (75-100 min.) results, suggesting that employing a uniform structure, which includes virtual nodes, can be potentially beneficial for traf…
Figure 6
Figure 6. Figure 6: Sensitivity Analysis of Virtual Nodes Number in Adaptive and Semi-Adaptive Adjacency Matrices [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Visualization of virtual node connection weights. [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 29 canonical work pages

  1. [1]

    Long-term traffic prediction based on lstm encoder-decoder architecture

    Zhumei Wang, Xing Su, and Zhiming Ding. Long-term traffic prediction based on lstm encoder-decoder architecture. IEEE Transactions on Intelligent Transportation Systems, 22 0 (10): 0 6561--6571, 2020

    work page 2020

  2. [2]

    Traffic flow forecasting for urban work zones

    Yi Hou, Praveen Edara, and Carlos Sun. Traffic flow forecasting for urban work zones. IEEE transactions on intelligent transportation systems, 16 0 (4): 0 1761--1770, 2014

    work page 2014

  3. [3]

    Spatio-temporal graph convolutional networks: a deep learning framework for traffic forecasting

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

    work page 2018

  4. [4]

    On the bottleneck of graph neural networks and its practical implications

    Uri Alon and Eran Yahav. On the bottleneck of graph neural networks and its practical implications. In International Conference on Learning Representations, 2020

    work page 2020

  5. [5]

    Oversquashing in gnns through the lens of information contraction and graph expansion

    Pradeep Kr Banerjee, Kedar Karhadkar, Yu Guang Wang, Uri Alon, and Guido Mont \'u far. Oversquashing in gnns through the lens of information contraction and graph expansion. In 2022 58th Annual Allerton Conference on Communication, Control, and Computing (Allerton), pages 1--8. IEEE, 2022

    work page 2022

  6. [6]

    On over-squashing in message passing neural networks: The impact of width, depth, and topology

    Francesco Di Giovanni, Lorenzo Giusti, Federico Barbero, Giulia Luise, Pietro Lio, and Michael M Bronstein. On over-squashing in message passing neural networks: The impact of width, depth, and topology. In International Conference on Machine Learning, pages 7865--7885. PMLR, 2023

    work page 2023

  7. [7]

    Learning Graph-Level Representation for Drug Discovery

    Junying Li, Deng Cai, and Xiaofei He. Learning graph-level representation for drug discovery. arXiv preprint arXiv:1709.03741, 2017

    work page Pith review arXiv 2017

  8. [8]

    Neural message passing for quantum chemistry

    Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing for quantum chemistry. In International conference on machine learning, pages 1263--1272. PMLR, 2017

    2017

Show all 44 references
  1. [9]

    Graph classification via deep learning with virtual nodes

    Trang Pham, Truyen Tran, Hoa Dam, and Svetha Venkatesh. Graph classification via deep learning with virtual nodes. arXiv preprint arXiv:1708.04357, 2017

    2017 arXiv

  2. [10]

    Boosting graph structure learning with dummy nodes

    Xin Liu, Jiayang Cheng, Yangqiu Song, and Xin Jiang. Boosting graph structure learning with dummy nodes. In International Conference on Machine Learning, pages 13704--13716. PMLR, 2022

    2022

  3. [11]

    The graph neural network model

    Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model. IEEE transactions on neural networks, 20 0 (1): 0 61--80, 2008

    2008

  4. [12]

    Inductive graph neural networks for spatiotemporal kriging

    Yuankai Wu, Dingyi Zhuang, Aurelie Labbe, and Lijun Sun. Inductive graph neural networks for spatiotemporal kriging. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 4478--4485, 2021 a

    2021

  5. [13]

    Uncertainty quantification of sparse travel demand prediction with spatial-temporal graph neural networks

    Dingyi Zhuang, Shenhao Wang, Haris Koutsopoulos, and Jinhua Zhao. Uncertainty quantification of sparse travel demand prediction with spatial-temporal graph neural networks. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 4639--4647, 2022

    2022

  6. [14]

    Adaptive graph convolutional recurrent network for traffic forecasting

    Lei Bai, Lina Yao, Can Li, Xianzhi Wang, and Can Wang. Adaptive graph convolutional recurrent network for traffic forecasting. Advances in neural information processing systems, 33: 0 17804--17815, 2020

    2020

  7. [15]

    Highly accurate protein structure prediction with alphafold

    John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Z \' dek, Anna Potapenko, et al. Highly accurate protein structure prediction with alphafold. nature, 596 0 (7873): 0 583--589, 2021

    2021

  8. [16]

    Graph neural networks for social recommendation

    Wenqi Fan, Yao Ma, Qing Li, Yuan He, Eric Zhao, Jiliang Tang, and Dawei Yin. Graph neural networks for social recommendation. In The world wide web conference, pages 417--426, 2019

    2019

  9. [17]

    Lightgcn: Simplifying and powering graph convolution network for recommendation

    Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. Lightgcn: Simplifying and powering graph convolution network for recommendation. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, pages...

    2020

  10. [18]

    Long short-term memory

    Sepp Hochreiter and J \"u rgen Schmidhuber. Long short-term memory. Neural computation, 9 0 (8): 0 1735--1780, 1997

    1997

  11. [19]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017

    2017

  12. [20]

    Diffusion convolutional recurrent neural network: Data-driven traffic forecasting

    Yaguang Li, Rose Yu, Cyrus Shahabi, and Yan Liu. Diffusion convolutional recurrent neural network: Data-driven traffic forecasting. In International Conference on Learning Representations, 2018

    2018

  13. [21]

    Predicting station-level short-term passenger flow in a citywide metro network using spatiotemporal graph convolutional neural networks

    Yong Han, Shukang Wang, Yibin Ren, Cheng Wang, Peng Gao, and Ge Chen. Predicting station-level short-term passenger flow in a citywide metro network using spatiotemporal graph convolutional neural networks. ISPRS International Journal of Geo-Information, 8 0 (6): 0 243, 2019

    2019

  14. [22]

    Dynamic and multi-faceted spatio-temporal deep learning for traffic speed forecasting

    Liangzhe Han, Bowen Du, Leilei Sun, Yanjie Fu, Yisheng Lv, and Hui Xiong. Dynamic and multi-faceted spatio-temporal deep learning for traffic speed forecasting. In Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, pages 547--555, 2021

    2021

  15. [23]

    Stgat: Spatial-temporal graph attention networks for traffic flow forecasting

    Xiangyuan Kong, Weiwei Xing, Xiang Wei, Peng Bao, Jian Zhang, and Wei Lu. Stgat: Spatial-temporal graph attention networks for traffic flow forecasting. IEEE Access, 8: 0 134363--134372, 2020

    2020

  16. [24]

    Decoupled dynamic spatial-temporal graph neural network for traffic forecasting

    Zezhi Shao, Zhao Zhang, Wei Wei, Fei Wang, Yongjun Xu, Xin Cao, and Christian S Jensen. Decoupled dynamic spatial-temporal graph neural network for traffic forecasting. Proceedings of the VLDB Endowment, 15 0 (11): 0 2733--2746, 2022

    2022

  17. [25]

    Learning phrase representations using rnn encoder--decoder for statistical machine translation

    Kyunghyun Cho, Bart van Merri \"e nboer, C a g lar Gu̇l c ehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder--decoder for statistical machine translation. In Proceedings of the 2014 Conference on Empiric...

    2014

  18. [26]

    Temporal convolutional networks for action segmentation and detection

    Colin Lea, Michael D Flynn, Rene Vidal, Austin Reiter, and Gregory D Hager. Temporal convolutional networks for action segmentation and detection. In proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 156--165, 2017

    2017

  19. [27]

    Deep multi-view spatial-temporal network for taxi demand prediction

    Huaxiu Yao, Fei Wu, Jintao Ke, Xianfeng Tang, Yitian Jia, Siyu Lu, Pinghua Gong, Jieping Ye, and Zhenhui Li. Deep multi-view spatial-temporal network for taxi demand prediction. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018

    2018

  20. [28]

    Revisiting spatial-temporal similarity: A deep learning framework for traffic prediction

    Huaxiu Yao, Xianfeng Tang, Hua Wei, Guanjie Zheng, and Zhenhui Li. Revisiting spatial-temporal similarity: A deep learning framework for traffic prediction. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 5668--5675, 2019

    2019

  21. [29]

    Graph wavenet for deep spatial-temporal graph modeling

    Zonghan Wu, Shirui Pan, Guodong Long, Jing Jiang, and Chengqi Zhang. Graph wavenet for deep spatial-temporal graph modeling. In Proceedings of the 28th International Joint Conference on Artificial Intelligence, pages 1907--1913, 2019

    1907

  22. [30]

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

    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 Transactions on Knowledge Discovery from Data, 17 0 (1): 0 1--21, 2023

    2023

  23. [31]

    Gman: A graph multi-attention network for traffic prediction

    Chuanpan Zheng, Xiaoliang Fan, Cheng Wang, and Jianzhong Qi. Gman: A graph multi-attention network for traffic prediction. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 1234--1241, 2020

    2020

  24. [32]

    Structure-aware transformer for graph representation learning

    Dexiong Chen, Leslie O’Bray, and Karsten Borgwardt. Structure-aware transformer for graph representation learning. In International Conference on Machine Learning, pages 3469--3489. PMLR, 2022 a

    2022

  25. [33]

    Pure transformers are powerful graph learners

    Jinwoo Kim, Dat Nguyen, Seonwoo Min, Sungjun Cho, Moontae Lee, Honglak Lee, and Seunghoon Hong. Pure transformers are powerful graph learners. Advances in Neural Information Processing Systems, 35: 0 14582--14595, 2022

    2022

  26. [34]

    Long range graph benchmark

    Vijay Prakash Dwivedi, Ladislav Ramp \'a s ek, Michael Galkin, Ali Parviz, Guy Wolf, Anh Tuan Luu, and Dominique Beaini. Long range graph benchmark. Advances in Neural Information Processing Systems, 35: 0 22326--22340, 2022

    2022

  27. [35]

    Representing long-range context for graph neural networks with global attention

    Zhanghao Wu, Paras Jain, Matthew Wright, Azalia Mirhoseini, Joseph E Gonzalez, and Ion Stoica. Representing long-range context for graph neural networks with global attention. Advances in Neural Information Processing Systems, 34: 0 13266--13279, 2021 b

    2021

  28. [36]

    Revisiting virtual nodes in graph neural networks for link prediction, 2022

    EunJeong Hwang, Veronika Thost, Shib Sankar Dasgupta, and Tengfei Ma. Revisiting virtual nodes in graph neural networks for link prediction, 2022. URL https://openreview.net/forum?id=ETiaOyNwJW

    2022

  29. [37]

    Probabilistic graph rewiring via virtual nodes

    Chendi Qian, Andrei Manolache, Christopher Morris, and Mathias Niepert. Probabilistic graph rewiring via virtual nodes. arXiv preprint arXiv:2405.17311, 2024

    2024 arXiv

  30. [38]

    Rewiring with positional encodings for graph neural networks

    Rickard Br \"u el-Gabrielsson, Mikhail Yurochkin, and Justin Solomon. Rewiring with positional encodings for graph neural networks. arXiv preprint arXiv:2201.12674, 2022

    2022 arXiv

  31. [39]

    Probabilistically rewired message-passing neural networks

    Chendi Qian, Andrei Manolache, Kareem Ahmed, Zhe Zeng, Guy Van den Broeck, Mathias Niepert, and Christopher Morris. Probabilistically rewired message-passing neural networks. arXiv preprint arXiv:2310.02156, 2023

    2023 arXiv

  32. [40]

    Timegnn: Temporal dynamic graph learning for time series forecasting

    Nancy Xu, Chrysoula Kosma, and Michalis Vazirgiannis. Timegnn: Temporal dynamic graph learning for time series forecasting. In International Conference on Complex Networks and Their Applications, pages 87--99. Springer, 2023

    2023

  33. [41]

    Balanced graph structure learning for multivariate time series forecasting

    Weijun Chen, Yanze Wang, Chengshuo Du, Zhenglong Jia, Feng Liu, and Ran Chen. Balanced graph structure learning for multivariate time series forecasting. arXiv preprint arXiv:2201.09686, 2022 b

    2022 arXiv

  34. [42]

    Connecting the dots: Multivariate time series forecasting with graph neural networks

    Zonghan Wu, Shirui Pan, Guodong Long, Jing Jiang, Xiaojun Chang, and Chengqi Zhang. Connecting the dots: Multivariate time series forecasting with graph neural networks. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages ...

    2020

  35. [43]

    Largest: A benchmark dataset for large-scale traffic forecasting

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

    2023

  36. [44]

    The emerging field of signal processing on graphs: Extending high-dimensional data analysis to networks and other irregular domains

    David I Shuman, Sunil K Narang, Pascal Frossard, Antonio Ortega, and Pierre Vandergheynst. The emerging field of signal processing on graphs: Extending high-dimensional data analysis to networks and other irregular domains. IEEE signal processing magazine, 30 0 (3): 0 83--98, 2013

    2013

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.