Beyond Attention: Learning Spatio-Temporal Dynamics with Emergent Interpretable Topologies

Sai Vamsi Alisetti; Sanjukta Krishnagopal; Vikas Kalagi

REVIEW 4 major objections 5 minor 47 references

Beyond Attention: Learning Spatio-Temporal Dynamics with Emergent Interpretable Topologies

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A single learnable interaction matrix can replace masked dynamic attention in spatio-temporal forecasting, improving accuracy by at least 21% and 6% on two traffic benchmarks while cutting training time by 60–70%.

desk verdict The forecasting claim is plausibly real, but the interpretability analysis is run on the pre-softmax matrix, so the topological story does not describe the model that actually predicts. read the letter →

arxiv 2506.00770 v1 pith:NDW3BPME submitted 2025-06-01 cs.LG cs.AIcs.SI

Sai Vamsi Alisetti , Vikas Kalagi , Sanjukta Krishnagopal This is my paper

classification cs.LGcs.AIcs.SI

keywords spatio-temporalforecastinggraphattentionnetworkslearnableinteractionmatrixlatentstructurelearninginterpretabilitycommunitydetectionspectralanalysistraffic

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

This paper proposes InterGAT, a graph-attention variant that replaces masked, dynamically computed attention scores with a single learnable symmetric node-to-node interaction matrix shared across time steps and trained end-to-end. Paired with a GRU temporal decoder, the model is claimed to outperform an equivalent GAT-GRU baseline across all forecast horizons on two traffic benchmarks, improving forecasting accuracy by at least 21% on SZ-Taxi and 6% on Los-Loop while reducing training time by 60–70%. The paper further argues that the learned matrix carries interpretable structure: sparse attention patterns, community-aligned weights, and a mix of localized and global spectral modes, so that interpretability is attached to the same object that performs spatial aggregation.

What carries the argument

The load-bearing object is the interaction matrix $I \in \mathbb{R}^{N \times N}$, a fully learnable parameter that starts without structural priors. At each forward pass it is symmetrized as $I \leftarrow \tfrac{1}{2}(I + I^\top)$, stabilized with LayerNorm, and row-normalized with softmax; the resulting nonnegative weights aggregate feature-transformed node embeddings through an ELU activation, replacing the masked attention coefficients of standard GAT. The matrix is shared across time, trained with an $\ell^1$ sparsity penalty alongside the forecasting MSE, and its spatial embeddings feed a GRU-based decoder that produces multi-horizon predictions. The interpretability analyses—eigenvalue spectrum, Dirichlet energy, inverse participation ratio, and community contrast—are all computed on this learned matrix.

What would settle it

Re-run the spectral and community-contrast analyses on the operational weights $\mathrm{softmax}(\mathrm{LayerNorm}(\tfrac{1}{2}(I + I^\top)))$; if the eigenvalue spectrum, Dirichlet energy ordering, and intra-versus-inter community contrast differ materially from the values reported for the raw $I$, the interpretability result does not apply to the predictive model.

Watch

Extended reading notes

Core claim

The central claim is that dynamic masked attention over a fixed adjacency graph is not required for accurate spatio-temporal forecasting. A persistent, fully learnable, symmetric interaction matrix $I \in \mathbb{R}^{N \times N}$, regularized toward sparsity and symmetry, can capture latent spatial dependencies directly from data, and on the SZ-Taxi and Los-Loop datasets it predicts better than the GAT-GRU baseline at all horizons from 15 to 60 minutes. Because the same matrix is reused at every time step, the forward and backward passes involve fewer operations, yielding the reported 60–70% training-time reduction. The paper also contends that the learned matrix is intrinsically interpretable: spectral and community analyses reveal sparse, topology-aware attention patterns, functional clusters that are not just physical neighborhoods, and attention heads that specialize in local versus global structure.

Load-bearing premise

The spectral and community analyses are carried out on the raw symmetrized interaction matrix, which can contain negative values, not on the nonnegative, row-softmaxed matrix that actually computes the spatial aggregation during prediction; if those two objects diverge, the claimed interpretable structure may not belong to the model that makes the forecasts.

Editorial extensions

If this is right

Because the interaction matrix is learned once and reused rather than recomputed per step, GAT-based spatio-temporal models can drop per-step attention computation without losing accuracy, which makes the simplification attractive for resource-constrained settings.
The gains at longer horizons suggest that a dense learnable interaction structure captures long-range dependencies that sparse adjacency masking suppresses.
Head specialization between local and global modes means multi-head interaction learning can double as an unsupervised probe for functional or community structure in a spatial system.

Reading between the lines

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

A natural next check for the interpretability claims is to repeat the spectral and community analyses on the nonnegative row-softmaxed weights actually used in the forward pass, since the reported analyses run on the raw symmetrized parameter that can contain negative values.
The reported MAE values are above 1 while the data is described as normalized to [0,1]; clarifying the exact preprocessing would settle which numerical comparisons are directly valid.
A directly testable extension would be to apply InterGAT to synthetic data with planted community structure, such as a stochastic block model, to check whether the learned matrix recovers ground-truth communities with high contrast.
The $N \times N$ interaction matrix may become a bottleneck on large graphs; the paper's own suggestion of low-rank reparameterization is a natural next test of whether the efficiency and interpretability results persist at scale.

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.

Desk Editor's Note

read the letter

Here's the short version. InterGAT is a simple, readable idea: replace per-step GAT attention with a single learnable symmetric interaction matrix, regularized with L1 sparsity, shared across time. On SZ-Taxi and Los-Loop, it beats an equivalent GAT-GRU by 21% and 6% in MAE, and trains 60-70% faster. The architecture is clearly specified, the ablation against fixed adjacency, weighted adjacency, and covariance matrices is a nice touch, and the code is linked. I think the core forecasting result is plausible and could be useful.

The soft spots, in order of seriousness. First, the interpretability analysis is on the wrong matrix. The forward pass symmetrizes I, applies LayerNorm, then row-wise softmax, and uses that row-stochastic matrix for aggregation. But Sections 5.5 and 5.6 analyze the raw learned I, even saying 'I can contain negative values.' The eigenvectors, Dirichlet energy, and community contrast are computed on a parameter, not on the attention weights that actually produce the predictions. That drives a wedge between the model and the 'emergent interpretable topologies' claim. The fix is straightforward: re-run the spectral and community analysis on the post-softmax matrix and see if the structure survives. Until then, the paper's headline contribution is unsupported.

Second, the numbers don't match the stated normalization. The paper says data is normalized to [0,1], but Table 1 reports MAE of 4.6 for GAT-GRU and 2.8 for InterGAT. That is impossible unless the metric is computed on de-normalized values. They need to state the evaluation scale clearly.

Third, the experimental comparison is thin. Only one baseline, GAT-GRU, which they designed. They cite Graph WaveNet and STGSL but don't run them. For a forecasting paper, that limits what we can conclude about the method's actual standing.

Minor: the train/test split is described only as 80/20, not chronological, which matters for time series, and the ablation's 'accuracy' metric is Frobenius-based, which is fine but unconventional.

Who is this for? Someone working on adaptive graph structure learning and lightweight attention alternatives. The forecasting result deserves referee time, but the interpretability claims need a major revision. I'd send it to review, but only after the authors re-tie the analysis to the operational matrix.

Referee Report

4 major / 5 minor

Summary. The paper proposes InterGAT, a spatio-temporal forecasting architecture that replaces masked GAT attention with a fully learnable, symmetric node interaction matrix I, combined with a GRU temporal decoder. The authors report that InterGAT-GRU outperforms a GAT-GRU baseline on SZ-Taxi and Los-Loop traffic datasets across 15- to 60-minute horizons, reduces training time by 60-70%, and exhibits interpretable structure, including sparsity, spectral patterns, and community alignment, as analyzed in Sections 5.2, 5.5, and 5.6.

Significance. If the claims hold, the paper offers a simple and efficient alternative to dynamic attention in GATs, with the appealing property that the learned interaction matrix can be analyzed with standard network-science tools. The experimental design includes multi-seed runs, an ablation over several interaction matrix formulations, and a public code link, which are strengths. However, two load-bearing issues currently undermine the contribution: the spectral/community analyses are performed on a matrix that may not be the one used for prediction, and the reported MAE values contradict the stated [0,1] data normalization. The interpretability conclusions also overstate the degree to which sparsity is 'emergent' given the explicit L1 penalty. These concerns are addressable, but they require substantive revisions.

major comments (4)

[§5.5-§5.6, Eqs. (2)-(5)] The forward pass in Eqs. (2)-(5) uses the matrix A = softmax(LayerNorm((I+I^T)/2), dim=1), which is nonnegative and row-stochastic, and Z_i is computed from this A. However, the spectral analysis in §5.5 defines the Dirichlet energy through L = D - I and explicitly states 'I can contain negative values,' and the community contrast in §5.6 is computed on 'the learned interaction matrix I.' The manuscript never states whether these analyses use the raw I, the symmetrized LayerNormed matrix, or the post-softmax row-stochastic matrix. If they use the raw matrix, then Figures 3-6 and Table 4 describe a parameter, not the matrix that actually aggregates features in Eq. (5). Row-wise softmax is a nonlinear, row-dependent transformation that can change relative interaction strengths and destroys symmetry, so the interpretability claims require re-running the analyses on the matrix that enters the prediction, or providing a rigorous justification for why the raw parameter is the relevant object.
[§4.1 and Table 1] Section 4.1 states that data is normalized to the range [0,1], but Table 1 reports MAE values as high as 7.35 (e.g., GAT-GRU on Los-Loop 60 min) and InterGAT-GRU MAE of 2.86-4.82 on normalized data. These values are impossible for data in [0,1] unless the predictions are extremely poor, which is inconsistent with the reported Accuracy values around 0.57-0.90. This suggests either the normalization statement is incorrect, the MAE is computed on the original (unnormalized) scale, or there is a unit mismatch. The authors must clarify the exact preprocessing and ensure all reported metrics are computed on the same scale; otherwise the quantitative comparisons cannot be trusted.
[Abstract and §3.2.2] The paper repeatedly claims that I is a 'symmetric node interaction matrix,' and this symmetry is presented as a key design principle. However, Eq. (4) applies a row-wise softmax to the symmetrized and LayerNormed matrix. For a general matrix, row-wise softmax produces a row-stochastic matrix that is not symmetric unless all rows are identical or the input satisfies special conditions. Thus the matrix that actually computes the spatial aggregation in Eq. (5) is not symmetric in general. The claim of symmetry therefore holds only for the pre-softmax parameter, not for the effective attention weights. The authors should either use a symmetric normalization (if symmetry of the effective weights is intended) or explicitly qualify the symmetry claim.
[§5.2] The text states that 'this sparsity emerges naturally from the soft L1 regularization without explicit masking.' Since the L1 penalty is an explicit, purpose-designed sparsity regularizer, the resulting 50-70% near-zero entries are a direct consequence of the loss term rather than an emergent property of the learning dynamics. This framing overstates the finding. The title and abstract use 'emergent interpretable topologies,' so the paper should be careful to distinguish between structure that arises from a hand-chosen penalty and structure that arises from the data. At minimum, the authors should present an ablation without the L1 penalty to quantify how much sparsity is actually learned rather than enforced.

minor comments (5)

[Abstract] There is a typo: 'indiate' should be 'indicate'.
[§5.1] The text says 'the largest gains are observed at the 60-minute horizon,' but in Table 1 the absolute MAE reduction for Los-Loop is larger at 15 minutes (7.0257 to 3.7752, a drop of 3.25) than at 60 minutes (7.3485 to 4.8219, a drop of 2.53). The claim should be corrected or clarified (e.g., largest relative improvement on SZ-Taxi).
[Figure 1 caption] The caption says 'Attention coefficients α_ij are computed using a normalized interaction matrix I+I^T,' but the forward pass described in Eqs. (2)-(4) uses LayerNorm and a row-wise softmax, not just I+I^T. The caption should be consistent with the equations.
[§5.4, Eq. (11) and Table 3] The first row of Eq. (11) is labeled I'=0, which is described as 'No bias — standard dense attention.' A zero matrix would produce no interactions at all, so this description is confusing. Clarify what I'=0 means and how it corresponds to the GAT baseline.
[Appendix A.6] The appendix defines five metrics (RMSE, MAE, Accuracy, R², Explained Variance), but the experimental tables only report MAE and Accuracy (Frobenius-norm definition). Either report the other metrics in the main results or remove them from the appendix to avoid promising unreported evaluation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the forecasting claims are evaluated on held-out test data, and the interpretive analyses are descriptive rather than fitted predictions.

full rationale

The paper's central performance claim is supported by held-out test comparisons against an equivalent GAT-GRU baseline, so it is not an artifact of fitting. The interaction matrix is trained with the forecasting loss plus an explicit L1 penalty, and Section 5.2 explicitly credits the sparsity to that regularization; calling it 'emergent' is an overstatement but not circularity. The community-alignment analysis compares the learned matrix against spectral communities of the static adjacency matrix, which is an external reference not used in training. There are no load-bearing self-citations: the reference list contains no prior work by Alisetti, Kalagi, or Krishnagopal. The one substantive concern is that Sections 5.5-5.6 analyze the raw learned matrix, and the paper notes 'I can contain negative values,' while the forward pass in Eq. (5) uses the row-wise softmaxed matrix; this is an object-mismatch for the interpretability claims, not circularity, because neither the spectral quantities nor the community contrast feed back into the training loss or the test-set forecast comparison. No fitted value is renamed as a prediction, and no derivation is equivalent by construction to its input.

Assumptions & free parameters 2 free parameters · 4 assumptions · 1 invented entities

The central free parameter is the interaction matrix I per head (about N^2 per head), learned end-to-end, plus the unreported L1 coefficient lambda_sparse. The paper also assumes the static adjacency matrices yield meaningful communities, that spectral analysis of I (even when the forward pass uses softmax(I)) is valid, and that the 80/20 split does not leak time. The invented entity is the interaction matrix itself, which has no independent external evidence; its interpretability is assessed in-sample.

free parameters (2)

Interaction matrix I (N x N, per attention head) = Learned end-to-end; values not reported
The full pairwise interaction matrix is learned from data; it is the central quantity of the method, replacing dynamic attention scores.
L1 regularization coefficient lambda_sparse = Not reported
Tunable hyperparameter controlling the sparsity penalty; its value is absent from the paper, though the sparsity and interpretability claims depend on it.

assumptions (4)

domain assumption The static adjacency matrices of SZ-Taxi and Los-Loop encode meaningful spatial connectivity; spectral clustering on these matrices yields valid communities.
Used as ground truth for community alignment in Section 5.6; if the static graph is uninformative, the contrast analysis is not meaningful.
domain assumption The learned interaction matrix I can be analyzed as a graph with a Laplacian and eigenvectors, even though it is an attention-like weight matrix and may not be positive semi-definite.
Section 5.5 defines Dirichlet energy and spectral quantities on I; the paper itself flags that the Laplacian is not necessarily PSD, so the standard spectral interpretation is not guaranteed.
domain assumption The 80/20 data split does not leak information from the future.
Section 4.1 does not state whether the split is chronological; a random split of a time series would leak temporal autocorrelation and inflate accuracy.
ad hoc to paper Sparsity can be quantified by the fraction of entries below 1e-4, and the L1 penalty produces the observed 50-70% sparsity.
The threshold 1e-4 and the L1 penalty are choices made by the authors; the "emergent" sparsity is partly induced by this penalty.

invented entities (1)

Symmetric node interaction matrix I
purpose: Latent pairwise spatial dependency structure that replaces masked attention; the central object for interpretability.
It is a model parameter, not an independently measured quantity. Its claimed community alignment is assessed post hoc on the same data and on a variant of the matrix that may not be the one used in predictions.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Beyond Attention: Learning Spatio-Temporal Dynamics with Emergent Interpretable Topologies." pith.science (2026). https://pith.science/paper/NDW3BPME

@misc{pith2026250600770,
  author       = {Pith},
  title        = {Pith review of: Beyond Attention: Learning Spatio-Temporal Dynamics with Emergent Interpretable Topologies},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NDW3BPME}},
  note         = {Machine review of arXiv:2506.00770}
}

read the original abstract

Spatio-temporal forecasting is critical in applications such as traffic prediction, energy demand modeling, and weather monitoring. While Graph Attention Networks (GATs) are popular for modeling spatial dependencies, they rely on predefined adjacency structures and dynamic attention scores, introducing inductive biases and computational overhead that can obscure interpretability. We propose InterGAT, a simplified alternative to GAT that replaces masked attention with a fully learnable, symmetric node interaction matrix, capturing latent spatial relationships without relying on fixed graph topologies. Our framework, InterGAT-GRU, which incorporates a GRU-based temporal decoder, outperforms the baseline GAT-GRU in forecasting accuracy, achieving at least a 21% improvement on the SZ-Taxi dataset and a 6% improvement on the Los-Loop dataset across all forecasting horizons (15 to 60 minutes). Additionally, we observed reduction in training time by 60-70% compared to GAT-GRU baseline. Crucially, the learned interaction matrix reveals interpretable structure: it recovers sparse, topology-aware attention patterns that align with community structure. Spectral and clustering analyses show that the model captures both localized and global dynamics, offering insights into the functional topology driving predictions. This highlights how structure learning can simultaneously support prediction, computational efficiency, and topological interpretabil-ity in dynamic graph-based domains.

Figures

Figures reproduced from arXiv: 2506.00770 by the authors.

**Figure 2.** Sparsity evolution of I per head, defined as fraction of values < 10−4 . The model promotes selective attention across all heads. Visualizing Interaction matrices: We visualize the interaction matrices produced by each attention head. Each head is min-max normalized (values are scaled to [0, 1] ), and only the top 2% of values are visualized to reveal the most salient attention-driven connections (see [PITH_FULL… view at source ↗

**Figure 4.** Dirichlet Energy of the eigenvectors of the learned [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

Figures from the paper (7 more)

**Figure 3.** Figure 3: Eigenvalue spectrum of the learned interaction [PITH_FULL_IMAGE:figures/full_fig_p007_3.png]

**Figure 5.** Figure 5: Inverse Participation Ratio (IPR) of eigenvectors of [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

**Figure 6.** Figure 6: Top three eigenvectors (156–154) of the final layer [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

**Figure 7.** Figure 7: Frobenius norm of the interaction matrix [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]

**Figure 8.** Figure 8: Intra- and inter-community statistics for each attention head computed using spectral clustering across varying [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]

**Figure 9.** Figure 9: Interaction Matrix Heatmap - Binarized top 2% of nodes [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]

**Figure 10.** Figure 10: Comparison of training and validation loss curves on two datasets using the InterGAT-GRU model. [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 34 canonical work pages

[1]

Lei Bai, Lina Yao, Can Li, Xianzhi Wang, and Can Wang. 2020. Adaptive graph convolutional recurrent network for traffic forecasting.Advances in neural information processing systems33 (2020), 17804–17815

2020
[2]

Danielle S Bassett and Olaf Sporns. 2017. Network neuroscience.Nature neuro- science20, 3 (2017), 353–364. doi:10.1038/nn.4502

doi:10.1038/nn.4502 2017
[3]

Peter Battaglia, Jessica B Hamrick, Victor Bapst, Alvaro Sanchez-Gonzalez, Vini- cius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner, et al. 2018. Relational inductive biases, deep learning, and graph networks.arXiv preprint arXiv:1806.01261(2018)

arXiv 2018
[4]

Shaked Brody and Eran Yahav. 2021. Attentive Neighbors for Graph Neural Networks.International Conference on Machine Learning (ICML)(2021)

work page 2021
[5]

Andrea Cavallo, Mohammad Sabbaqi, and Elvin Isufi. 2024. Spatiotemporal covariance neural networks. InJoint European Conference on Machine Learning and Knowledge Discovery in Databases. Springer, 18–34

work page 2024
[6]

Kyunghyun Cho, Bart van Merriënboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2014. Learning phrase representations using RNN encoder–decoder for statistical machine translation. InProceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP). 1724–1734

work page 2014
[7]

Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. 2014. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555(2014)

arXiv 2014
[8]

Ronald R Coifman and Stéphane Lafon. 2006. Diffusion maps.Applied and computational harmonic analysis21, 1 (2006), 5–30

work page 2006

Show all 47 references

[9]

Vijay Prakash Dwivedi, Renjie Liao, Yoshua Bengio, and Xavier Bresson. 2023. Graph Neural Networks with Adaptive Spectral Filters. InInternational Confer- ence on Learning Representations (ICLR)

2023
[10]

Zheng Fang, Qingqing Long, Guojie Song, and Kunqing Xie. 2021. Spatial- temporal graph ode networks for traffic flow forecasting. InProceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining. 364–373

2021
[11]

Shengnan Guo, Youfang Lin, Ning Feng, Chao Song, and Huaiyu Wan. 2019. Attention based spatial-temporal graph convolutional networks for traffic flow forecasting. InProceedings of the AAAI conference on artificial intelligence, Vol. 33. 922–929

2019
[12]

Hamilton, Rex Ying, and Jure Leskovec

William L. Hamilton, Rex Ying, and Jure Leskovec. 2017. Representation Learning on Graphs: Methods and Applications.IEEE Data Engineering Bulletin40, 3 (2017), 52–74

2017
[13]

Hu, Yelong Shen, Phil Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Lu Wang, and Weizhu Chen

Edward J. Hu, Yelong Shen, Phil Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Lu Wang, and Weizhu Chen. 2021. LoRA: Low-Rank Adaptation of Large Language Models. arXiv preprint arXiv:2106.09685(2021)

2021 arXiv
[14]

Bronstein

Anees Kazi, Luca Cosmo, Seyed-Ahmad Ahmadi, Nassir Navab, and Michael M. Bronstein. 2023. Differentiable Graph Module (DGM) for Graph Convolutional Networks.IEEE Transactions on Pattern Analysis and Machine Intelligence45, 2 (2023), 1606–1617

2023
[15]

Minseo Kim, Seongjun Oh, U Kang, and Jaewoo Kang. 2022. Pure Transformers are Powerful Graph Learners. InNeurIPS

2022
[16]

Thomas Kipf, Ethan Fetaya, Kuan-Chieh Wang, Max Welling, and Richard Zemel
[17]

Thomas N Kipf and Max Welling. 2017. Semi-supervised classification with graph convolutional networks. InInternational Conference on Learning Representations (ICLR)

2017
[18]

Xiangyuan Kong, Weiwei Xing, Xiang Wei, Peng Bao, Jian Zhang, and Wei Lu
[19]

Yaguang Li, Rose Yu, Cyrus Shahabi, and Yan Liu. 2018. Diffusion convolu- tional recurrent neural network: Data-driven traffic forecasting. InInternational Conference on Learning Representations (ICLR)

2018
[20]

Ziyu Li, Shiyu Chen, Wei Zheng, Yuxuan Wang, Yang Liu, and Xiangliang Zhang
[21]

Ziyuan Liu, Guojie Song, Yucheng Cao, Meng Qu, Wenjun Zhao, Xiaowei Hu, and Qiang Yang. 2020. Physical-Information Enhanced Graph Neural Networks for Traffic Forecasting. InProceedings of the 29th ACM International Conference on Information & Knowledge Management (CIKM). 629–638

2020
[22]

Andrew Y Ng, Michael I Jordan, and Yair Weiss. 2002. On spectral clustering: Analysis and an algorithm.Advances in neural information processing systems14 (2002)

2002
[23]

Suman Ravuri and et al. 2021. Skilful precipitation nowcasting using deep generative models of radar.Nature597, 7878 (2021), 672–677

2021
[24]

Marta Sales-Pardo, Roger Guimerà, André A Moreira, and Luís A Nunes Amaral
[25]

Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. 2009. The Graph Neural Network Model.IEEE Trans- actions on Neural Networks20, 1 (2009), 61–80. doi:10.1109/TNN.2008.2005605

2009
[26]

Mahya Shahhosseini, Ricardo A Martinez-Feria, Guanqiu Hu, Sotirios V Archon- toulis, and Fernando E Miguez. 2021. Forecasting County-Level Corn Yields Using Deep Temporal Convolutional Neural Networks.Scientific Reports11, 1 (2021), 1–12

2021
[27]

Chao Shang, Jie Chen, and Jinbo Bi. 2021. Discrete graph structure learning for forecasting multiple time series.arXiv preprint arXiv:2101.06861(2021)

2021 arXiv
[28]

Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. 2018. Graph attention networks. InInternational Con- ference on Learning Representations (ICLR)

2018
[29]

Petar Velickovic, William Fedus, William L Hamilton, Pietro Liò, Yoshua Bengio, and R Devon Hjelm. 2020. Deep Graph Infomax. InInternational Conference on Learning Representations (ICLR)

2020
[30]

Hao Wang, Fuyong Sun, Jinxin Si, Qiuzhe Ma, Wenjing Zeng, Xiuhuan Zang, Junxi Cao, Shuaibing Song, and Nan Wang. 2024. Robust Spatio-Temporal Graph Neural Network for Electricity Consumption Forecasting. InProceedings of the 2024 8th International Conference on Big Data and In...

2024
[31]

Liang Wang, Zhiwei Wang, Wenjie Tang, Nikolaus Parulian, Yong Zhang, Fei Shen, Haizhou Liu, Yifan Liu, Hua Xu, and Fei Wang. 2022. DeepCOVIDNet: An interpretable deep learning model for predictive surveillance of COVID-19 using heterogeneous features and their interactions.IEE...

2022
[32]

Jiaheng Wu, Jiale Zhang, Yanfeng Xu, and Chen Wang. 2023. Learning Spatial- Temporal Graph for Forecasting Social Event Participation. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 2107–2117

2023
[33]

Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and Philip S Yu. 2021. A comprehensive survey on graph neural networks.IEEE Transactions on Neural Networks and Learning Systems32, 1 (2021), 4–24

2021
[34]

Zonghan Wu, Shirui Pan, Guodong Long, Jing Jiang, Xiaojun Chang, and Chengqi Zhang. 2019. Graph WaveNet for Deep Spatial-Temporal Graph Modeling. In Proceedings of the 28th International Joint Conference on Artificial Intelligence (IJCAI). 1907–1913

2019
[35]

Peng Xie, Minbo Ma, Tianrui Li, Shenggong Ji, Shengdong Du, Zeng Yu, and Junbo Zhang. 2023. Spatio-temporal dynamic graph relation learning for urban metro flow prediction.IEEE Transactions on Knowledge and Data Engineering35, 10 (2023), 9973–9984

2023
[36]

Mingxing Xu, Wenrui Dai, Chunmiao Liu, Xing Gao, Weiyao Lin, Guo-Jun Qi, and Hongkai Xiong. 2020. Spatial-temporal transformer networks for traffic flow forecasting.arXiv preprint arXiv:2001.02908(2020). 9 Sai Vamsi Alisetti, Vikas Kalagi, and Sanjukta Krishnagopal

2020 arXiv
[37]

Nancy Xu, Chrysoula Kosma, and Michalis Vazirgiannis. 2023. TimeGNN: tempo- ral dynamic graph learning for time series forecasting. InInternational Conference on Complex Networks and Their Applications. Springer, 87–99

2023
[38]

Zhitao Ying, Dylan Bourgeois, Jiaxuan You, Marinka Zitnik, and Jure Leskovec
[39]

Hanzhi Yu, Jian Wang, Xuesong Li, Yaguang Liu, and Cyrus Shahabi. 2023. DGFormer: Dynamic Graph Structure Learning for Spatio-Temporal Forecasting. InInternational Conference on Learning Representations (ICLR)

2023
[40]

Jiawei Zhang, Yuxuan Zheng, Lingbo Qi, Puyang Li, and Philip S Yu. 2021. Spatio- temporal graph structure learning for traffic forecasting. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 35. 4189–4197

2021
[41]

Jianzhong Zhang, Yuxuan Zheng, Lin Qi, Yu Zheng, Xuan Yi, and Tianrui Li. 2020. Spatio-Temporal Graph Attention Networks: A Deep Learning Framework for Traffic Forecasting. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 34. 1119–1126

2020
[42]

Ling Zhao, Yujiao Song, Chao Zhang, Yu Liu, Pu Wang, Tao Lin, Min Deng, and Haifeng Li. 2019. T-GCN: A temporal graph convolutional network for traffic prediction.IEEE transactions on intelligent transportation systems21, 9 (2019), 3848–3858. A Appendix A.1 Spectral Clustering...

2019
[2007]

doi:10.1073/ pnas.0703740104

Extracting the hierarchical organization of complex systems.Proceedings of the National Academy of Sciences104, 39 (2007), 15224–15229. doi:10.1073/ pnas.0703740104

2007
[2018]

InInternational confer- ence on machine learning

Neural relational inference for interacting systems. InInternational confer- ence on machine learning. Pmlr, 2688–2697
[2019]

Gnnexplainer: Generating explanations for graph neural networks.Ad- vances in neural information processing systems32 (2019)

2019
[2020]

doi:10.1109/ACCESS.2020

STGAT: Spatial-Temporal Graph Attention Networks for Traffic Flow Forecasting.IEEE Access8 (2020), 134363–134372. doi:10.1109/ACCESS.2020. 3011186

2020 doi
[2023]

InProceedings of the AAAI Conference on Artificial Intelligence, Vol

UrbanFlow: Spatiotemporal Graph Neural Networks for Traffic Forecasting in Smart Cities. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 37. 9266–9274

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Lei Bai, Lina Yao, Can Li, Xianzhi Wang, and Can Wang. 2020. Adaptive graph convolutional recurrent network for traffic forecasting.Advances in neural information processing systems33 (2020), 17804–17815

2020

[2] [2]

Danielle S Bassett and Olaf Sporns. 2017. Network neuroscience.Nature neuro- science20, 3 (2017), 353–364. doi:10.1038/nn.4502

doi:10.1038/nn.4502 2017

[3] [3]

Peter Battaglia, Jessica B Hamrick, Victor Bapst, Alvaro Sanchez-Gonzalez, Vini- cius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner, et al. 2018. Relational inductive biases, deep learning, and graph networks.arXiv preprint arXiv:1806.01261(2018)

arXiv 2018

[4] [4]

Shaked Brody and Eran Yahav. 2021. Attentive Neighbors for Graph Neural Networks.International Conference on Machine Learning (ICML)(2021)

work page 2021

[5] [5]

Andrea Cavallo, Mohammad Sabbaqi, and Elvin Isufi. 2024. Spatiotemporal covariance neural networks. InJoint European Conference on Machine Learning and Knowledge Discovery in Databases. Springer, 18–34

work page 2024

[6] [6]

Kyunghyun Cho, Bart van Merriënboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2014. Learning phrase representations using RNN encoder–decoder for statistical machine translation. InProceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP). 1724–1734

work page 2014

[7] [7]

Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. 2014. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555(2014)

arXiv 2014

[8] [8]

Ronald R Coifman and Stéphane Lafon. 2006. Diffusion maps.Applied and computational harmonic analysis21, 1 (2006), 5–30

work page 2006