Adaptive Spatiotemporal Augmentation for Improving Dynamic Graph Learning

Bingce Wang; Hanlin Xue; Tong Mo; Tuoyu Feng; Weiping Li; Xiaoyang Liu; Xu Chu; Zhijie Tan

REVIEW 3 major objections 5 minor 32 references

Adaptive Spatiotemporal Augmentation for Improving Dynamic Graph Learning

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A wavelet-based activity score identifies noise-prone nodes, and a biased random walk turns that score into a diffusion matrix that improves dynamic graph learning.

desk verdict STAA is a clever wavelet-guided extension of TIARA with consistent empirical wins, but the random walk probabilities as written can go negative and need a clamp before the method is fully coherent. read the letter →

arxiv 2501.10010 v1 pith:XXZTKBFA submitted 2025-01-17 cs.LG cs.AI

Xu Chu , Hanlin Xue , Bingce Wang , Xiaoyang Liu , Weiping Li , Tong Mo , Tuoyu Feng , Zhijie Tan This is my paper · ORCID

classification cs.LGcs.AI

keywords dynamicgraphlearningaugmentationwavelettransformrandomwalkdiffusionneuralnetworksnoiserobustnesstemporallocalitylinkprediction

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

Dynamic graphs contain noisy edges that appear and vanish for reasons unrelated to the underlying process, and most augmentation methods assume recent edges are more important—an assumption that lets such noise dominate learning. This paper proposes STAA, a model-agnostic augmentation that scores every node's spatiotemporal activity from graph wavelet coefficients and their change rate over snapshots, then uses a random walk with node-specific time-travel probabilities to build a diffusion matrix that down-weights edges attached to active nodes. The diffusion matrix replaces the adjacency matrix when training any discrete-time dynamic graph neural network. Experiments on seven datasets report that STAA beats no augmentation, static augmentations, and the temporal-locality baselines TIARA and TGAC on both link prediction and node classification.

What carries the argument

The central object is the SpatioTemporal Activity-Aware Random Walk Diffusion matrix $X_t$, generated by the STAARW random walk. On each snapshot $G_t$, a walker moves to a neighbor with probability $1-\alpha-\beta$, restarts at the seed with probability $\alpha$, or jumps to the same node in the next snapshot with probability $\beta$, where $\beta$ is the node's spatiotemporal activity coefficient. The coefficient combines the normalized high-frequency wavelet coefficient (spatial activity) and the normalized rate of change of the low-frequency coefficient over a window of $W$ snapshots (temporal activity). The stationary distribution of this walk, stacked over all seed nodes, yields the recurrence $X_t = \alpha L_t^{-1} + L_t^{-1} \beta_t X_{t-1}$; the resulting matrix is sparsified by a threshold $\rho$ and then used as the augmented adjacency matrix. The machinery converts a continuous spectral estimate of noise-proneness into concrete edge weights before any GNN forward pass.

What would settle it

Build a dynamic graph with injected noise placed deliberately on nodes with smooth, stable neighborhoods (low high-frequency coefficients) and put stable patterns on irregular nodes; then STAA's diffusion matrix should assign high weight to noise if the activity signal is working as claimed, and link prediction should degrade relative to no augmentation. Alternatively, inspect the learned diffusion matrix on a real dataset and check directly whether noisy edges receive lower weight than stable edges for the same node.

Watch

Extended reading notes

Core claim

The paper's central claim is that noise in dynamic graphs can be identified, without labels, from node-level spectral behavior: nodes whose neighborhoods are structurally irregular (high high-frequency graph wavelet coefficients) and whose low-frequency coefficients change sharply across snapshots are the nodes most likely to carry spurious recent edges. STAA condenses these two signals into a spatiotemporal activity coefficient for each node and time step, and uses it to bias a random walk on the snapshot sequence. The walk either moves to a neighbor, restarts at the seed, or time-travels to the next snapshot with a probability proportional to the node's activity; solving this walk gives a diffusion matrix whose entries re-weight the graph so that active nodes' recent-edge preference is suppressed. This matrix is then used in place of the original adjacency matrix for GNN training. The paper argues this is why STAA outperforms augmentations built purely on temporal decay or spatial denoising.

Load-bearing premise

The load-bearing premise is that high high-frequency wavelet coefficients really do mark nodes whose edges are likely to be noise, and that low-frequency coefficient changes really do mark temporally unstable nodes; if that spectral-to-noise mapping fails on a given graph, STAA will down-weight the wrong edges.

Editorial extensions

If this is right

Dynamic GNNs using STAA's diffusion matrix should become less sensitive to spurious single-snapshot edges, since the random walk shifts weight from an active node's newest edges to its earlier, more stable edges.
The augmentation is plug-and-play: the paper demonstrates it with GCN, GCRN, and EvolveGCN, so any discrete-time dynamic GNN can consume the diffusion matrix without architectural changes.
STAA requires no labels, noise annotations, or generative models; the activity coefficient is computed purely from adjacency matrices and node degrees across snapshots.
If the reported gains generalize, augmentation methods built solely on temporal locality will need a node-activity component rather than a uniform recent-edge emphasis.

Reading between the lines

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

I would expect STAA's advantage to shrink on graphs where node activity is nearly uniform, because then the temporal-locality bias and the activity bias coincide and the extra machinery mainly adds variance.
A direct test of the noise-identification claim would be to compare STAA's diffusion weights against an oracle that knows which edges are injected noise, isolating how much of the gain comes from correct identification versus from generic smoothing.
The same wavelet-based activity signal could be reused outside augmentation, for instance as an edge-weighting scheme for static GNN training or as a node-selection heuristic for active learning, though the paper does not explore these uses.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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

Quick take: the paper's augmentation rule is genuinely new—node-specific activity coefficients computed from graph wavelet coefficients steer a TIARA-style random walk—and the empirical story is consistent across seven datasets and two tasks. But the method as written has an internal inconsistency that must be fixed before the theoretical framing holds up.

The problem is in Eq. (6). The walker moves with probability 1−α−β_{t,u}, and β_{t,u} is defined as δ·σ(τ̂) with δ up to 2 and σ the sigmoid. Nothing clamps β to 1−α, which the text claims as an upper limit. For any node with τ̂ near zero and δ=2, β≈1, so 1−α−β is negative for typical α<1; for positive τ̂ it gets worse. That means the diagonal matrix I−αI−β can have negative entries, L_t may not be a nonsingular M-matrix, and the resulting X_t can contain negative values. Either the implementation clamps β somewhere, in which case the paper needs to say so, or the equations do not describe a valid random walk. This is load-bearing, not cosmetic.

What the paper does well: the spatiotemporal activity coefficient is a reasonable design, the random walk extension is a natural way to make time travel node-dependent, and the experiments are run five times with error bars. STAA beats NONE, TIARA, and TGAC on most metrics, and the code is promised. That is real evidence the idea has value.

The other soft spot is Observation (a): the claim that high-frequency wavelet coefficients indicate noise-prone critical nodes. The paper gives a small illustrative example but no formal argument or direct empirical check on real dynamic graphs. That is a genuine limitation, though for an applied augmentation paper it is secondary if the empirical results hold.

Minor issues: TGAC was reproduced by the authors and DropEdge/GDC results are borrowed from the TIARA paper, so the baseline comparison is not fully apples-to-apples. Hyperparameter tuning is otherwise standard.

Net: the empirical claim is plausible and the idea is worth engaging with, but the paper in its current form has a gap in the math that a referee should insist be fixed. I would send it to peer review, with the request that the authors either clamp β explicitly or restrict δ, and show that X_t entries are non-negative in practice.

Referee Report

3 major / 5 minor

Summary. The paper proposes STAA, a model-agnostic augmentation method for discrete-time dynamic graphs. It computes per-node spatiotemporal activity coefficients using graph wavelet coefficients and their rate of change, then biases a random walk with restart and time travel to produce a diffusion matrix that is used as an augmented adjacency matrix for dynamic GNN training. Experiments on seven datasets compare STAA with static and dynamic augmentation baselines on link prediction and node classification, reporting consistent improvements over no augmentation and over the dynamic methods TIARA and TGAC. The central claim is that suppressing recently preferred edges for active nodes reduces the influence of noise and improves dynamic graph learning.

Significance. If the algorithm is corrected and validated, the paper makes a useful contribution: it introduces a novel, principled mechanism for incorporating spatiotemporal activity into dynamic graph augmentation, with code and a reproducible experimental protocol (five runs, error bars, standard snapshot partitions). The use of graph wavelet coefficient change rates to detect temporally active nodes is a distinctive idea that goes beyond the temporal-locality assumption of prior methods. The experiments cover relevant baselines and datasets. However, the current formulation contains a load-bearing technical inconsistency in the random-walk transition probabilities, and the core Observation linking wavelet coefficients to noise susceptibility is not empirically established. These issues must be addressed before the claims can be accepted.

major comments (3)

[Section III-B, Eqs. (5)-(6)] The transition probabilities are not well-defined as written. In Eq. (5), β_{t,u}=δ·σ(τ̂_{t,u}) with δ∈(0,2] and σ(·)∈(0,1), so β_{t,u} can exceed 1−α for many nodes. The text states that β has an upper limit of 1−α, but no clamping or reparameterization appears in Eqs. (5)-(7). Consequently, the 'random walk' probability 1−α−β_{t,u} can be negative, and L_t = I−Ã_t^T(I−αI−β_{t,∧}) need not be a nonsingular M-matrix. As a result, X_t can contain negative entries and is not guaranteed to be a valid weighted adjacency matrix. Please either clamp β to [0,1−α], reparameterize β as δ·σ(τ̂) with δ∈(0,1−α), or otherwise ensure 0≤β≤1−α, and then verify that the reported results are reproduced with the actual implementation.
[Section II, Observation (a)] The method's entire activity coefficient rests on the unproven mapping from high-frequency wavelet coefficients to noise susceptibility. The manuscript offers intuition via the small example in Figure 2 and cites [18], but it provides no formal argument or direct empirical measurement that nodes with high high-frequency coefficients are actually more likely to be incident to noisy edges in real dynamic graphs. Because this Observation is the basis for the spatial activity term in Eq. (5), please provide empirical support—for example, inject synthetic noise into snapshots and show that high-frequency coefficients concentrate on corrupted nodes, or compare STAA's activity coefficients with ground-truth noisy-edge labels on a dataset with known noise.
[Section III-B, Eqs. (6)-(7)] Even if β is clamped, the recursive update in Eq. (6) is not a standard stationary-distribution equation because the time-travel term β_{t,∧}x_{t−1,s} couples to the previous snapshot. The paper claims that x_{t,s} is a probability distribution, but with this extra term x_{t,s} is not obviously nonnegative or normalized, and no convergence proof for the recurrence is given. Since X_t is used as an augmented adjacency matrix, entrywise nonnegativity is essential. Please provide a proof or state explicit sufficient conditions (e.g., spectral radius of L_t^{-1}β_{t,∧} < 1, nonnegativity of L_t^{-1}) under which X_t is entrywise nonnegative and bounded.

minor comments (5)

[Section III-B, Eq. (6)] The notation β_{t,∧} is confusing; using diag(β_{t,1},...,β_{t,n}) or B_t would be clearer and more standard.
[Tables I-II] No statistical significance tests are reported; with five runs, pairwise differences between some comparable methods (e.g., STAA vs. TIARA on several cells) may not be significant. Please add a significance test or at least error-bar analysis.
[Section IV] The very large improvements on some datasets (e.g., Table I, WikiElec with EvolveGCN, 59.3→85.1) are striking and would benefit from a brief explanation or an ablation to rule out an artifact of the sparsification threshold or hyperparameter search.
[Section IV-A] DropEdge and GDC results are taken from [11] rather than recomputed in the same environment; while this is acknowledged, a short note on the compatibility of the experimental setups would strengthen reproducibility.
[Section III-A, Eq. (5)] The hyperparameter ranges in the implementation details (δ∈(0,2]) directly conflict with the text's assertion that β has an upper limit of 1−α; this contradiction should be resolved in the text, not only in the equations.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: STAA's diffusion matrix is a transformation of the input graph; the noise-susceptibility assumption is external, and all benchmarks are independent.

full rationale

STAA computes node activity coefficients from graph wavelet coefficients of the input adjacency matrices (Eqs. 2-5) and uses them to bias a random walk (Eq. 6). The resulting diffusion matrix X_t is a deterministic function of A_t and X_{t-1}; it is an augmented input to a downstream GNN, not a prediction of an external target. No fitted parameter is renamed as a prediction; hyperparameters (α, δ, γ, W, ρ) are tuned on validation data and applied to held-out test sets, which is standard model selection, not circularity. The assumption that high-frequency wavelet coefficients correspond to noise-susceptible nodes (Observation (a), Section II) is an unproved premise supported by citation [18]; if wrong it would undermine validity, but it is not derived from the method's output and does not make the derivation circular. The paper contains no load-bearing self-citations: references [11] and [20] are by Lee and Jung, not the present authors, and are used as baseline and inspiration. The skeptic's concern that β_{t,u} can exceed 1−α (Section III-B, Eq. 6) is an internal-consistency/correctness issue, not a circularity, and does not change the circularity score.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The paper introduces the spatiotemporal activity coefficient β_{t,j} as a derived scalar, not a new physical entity. The central claim rests on the domain assumption that wavelet coefficients identify noise-prone nodes, and on the validity of the random walk recursion. Five hyperparameters are tuned on validation data, and the method's success depends on their chosen ranges.

free parameters (6)

α (restart probability)
Searched in (0,1) on validation; controls the balance between random walk and restart.
δ (scaling factor for activity coefficient)
Searched in (0,2]; can in principle make β exceed 1−α, which would break the probability interpretation.
γ (gating factor)
Searched in (0,1); balances temporal change rate and spatial high-frequency content.
W (time window size)
Searched in [1,10]; determines how many snapshots are used for the low-frequency coefficient change rate.
ρ (sparsification threshold)
Searched in [0.0001, 0.01]; controls sparsity of the diffusion matrix.
learning rate and dropout ratio
Searched in [0.01, 0.05] and [0, 0.5] respectively; standard training hyperparameters.

assumptions (4)

domain assumption Nodes with high high-frequency wavelet coefficients occupy critical topological positions and are susceptible to noise.
State as Observation (a) in Section II. This assumption is load-bearing because it links wavelet coefficients to noise, motivating the entire activity coefficient.
domain assumption Nodes with frequently changing low-frequency coefficients are more likely to be temporally active.
State as Observation (b) in Section II. It justifies using the low-frequency coefficient change rate as a temporal activity signal.
domain assumption The STAARW recursion in Equation (6) defines a valid probability distribution over nodes at each time step.
The paper assumes the walk probabilities are valid, but does not enforce 1−α−β_{t,u} ≥ 0 or prove convergence of the recursion.
standard math Standard spectral graph theory, including graph Laplacian eigendecomposition and the graph wavelet transform, applies to the snapshots.
Used in Section II to define wavelet coefficients; these are well-established mathematical tools.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Adaptive Spatiotemporal Augmentation for Improving Dynamic Graph Learning." pith.science (2026). https://pith.science/paper/XXZTKBFA

@misc{pith2026250110010,
  author       = {Pith},
  title        = {Pith review of: Adaptive Spatiotemporal Augmentation for Improving Dynamic Graph Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XXZTKBFA}},
  note         = {Machine review of arXiv:2501.10010}
}

read the original abstract

Dynamic graph augmentation is used to improve the performance of dynamic GNNs. Most methods assume temporal locality, meaning that recent edges are more influential than earlier edges. However, for temporal changes in edges caused by random noise, overemphasizing recent edges while neglecting earlier ones may lead to the model capturing noise. To address this issue, we propose STAA (SpatioTemporal Activity-Aware Random Walk Diffusion). STAA identifies nodes likely to have noisy edges in spatiotemporal dimensions. Spatially, it analyzes critical topological positions through graph wavelet coefficients. Temporally, it analyzes edge evolution through graph wavelet coefficient change rates. Then, random walks are used to reduce the weights of noisy edges, deriving a diffusion matrix containing spatiotemporal information as an augmented adjacency matrix for dynamic GNN learning. Experiments on multiple datasets show that STAA outperforms other dynamic graph augmentation methods in node classification and link prediction tasks.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 29 canonical work pages

[18]

Discrete signal processing on graphs: Frequency analysis,

A. Sandryhaila and J. M. F. Moura, “Discrete signal processing on graphs: Frequency analysis,” IEEE Transactions on Signal Processing, vol. 62, pp. 3042–3054, 2013. [Online]. Available: https://api.semanticscholar.org/CorpusID:12110057

work page 2013
[1]

Data augmentation for graph neural networks,

T. Zhao, Y . Liu, L. Neves, O. J. Woodford, M. Jiang, and N. Shah, “Data augmentation for graph neural networks,” in AAAI Conference on Artificial Intelligence , 2020. [Online]. Available: https://api.semanticscholar.org/CorpusID:219635816

work page 2020
[2]

Graph contrastive learning with adaptive augmentation,

Y . Zhu, Y . Xu, F. Yu, Q. Liu, S. Wu, and L. Wang, “Graph contrastive learning with adaptive augmentation,” Proceedings of the Web Conference 2021 , 2020. [Online]. Available: https://api. semanticscholar.org/CorpusID:225094367

work page 2021
[3]

Spectral Feature Augmentation for Graph Contrastive Learning and Beyond

Y . Zhang, H. Zhu, Z. Song, P. Koniusz, and I. King, “Spectral feature augmentation for graph contrastive learning and beyond,” ArXiv, vol. abs/2212.01026, 2022. [Online]. Available: https://api.semanticscholar. org/CorpusID:254221169

work page Pith review arXiv 2022
[4]

Unleashing the power of graph data augmentation on covariate distribution shift,

Y . Sui, X. Wang, J. Wu, A. Zhang, and X. He, “Unleashing the power of graph data augmentation on covariate distribution shift,” in Neural Information Processing Systems , 2022. [Online]. Available: https://api.semanticscholar.org/CorpusID:253384147

work page 2022
[5]

Structured sequence modeling with graph convolutional recurrent networks,

Y . Seo, M. Defferrard, P. Vandergheynst, and X. Bresson, “Structured sequence modeling with graph convolutional recurrent networks,” in International Conference on Neural Information Processing , 2016. [Online]. Available: https://api.semanticscholar.org/CorpusID:2687749

work page 2016
[6]

Evolvegcn: Evolving graph convolutional networks for dynamic graphs,

A. Pareja, G. Domeniconi, J. Chen, T. Ma, T. Suzumura, H. Kanezashi, T. Kaler, and C. E. Leisersen, “Evolvegcn: Evolving graph convolutional networks for dynamic graphs,” ArXiv, vol. abs/1902.10191, 2019. [Online]. Available: https://api.semanticscholar.org/CorpusID:67856459

arXiv 1902
[7]

Dysat: Deep neural representation learning on dynamic graphs via self- attention networks,

A. Sankar, Y . Wu, L. Gou, W. Zhang, and H. Yang, “Dysat: Deep neural representation learning on dynamic graphs via self- attention networks,” Proceedings of the 13th International Conference on Web Search and Data Mining , 2020. [Online]. Available: https://api.semanticscholar.org/CorpusID:210884001

work page 2020

Show all 32 references

[8]

Wingnn: Dynamic graph neural networks with random gradient aggregation window,

Y . Zhu, F. Cong, D. Zhang, W. Gong, Q. Lin, W. Feng, Y . Dong, and J. Tang, “Wingnn: Dynamic graph neural networks with random gradient aggregation window,” Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , 2023. [Online]. Available: https...

2023
[9]

Todynet: temporal dynamic graph neural network for multivariate time series classification,

H. Liu, D. Yang, X. Liu, X. Chen, Z. Liang, H. Wang, Y . Cui, and J. Gu, “Todynet: temporal dynamic graph neural network for multivariate time series classification,” Information Sciences, p. 120914, 2024

2024
[10]

Adaptive data augmentation on temporal graphs,

Y . Wang, Y . Cai, Y . Liang, H. Ding, C. Wang, S. Bhatia, and B. Hooi, “Adaptive data augmentation on temporal graphs,” in Neural Information Processing Systems , 2021. [Online]. Available: https://api.semanticscholar.org/CorpusID:248498293

2021
[11]

Time-aware random walk diffusion to improve dynamic graph learning,

J.-w. Lee and J. Jung, “Time-aware random walk diffusion to improve dynamic graph learning,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 7, 2023, pp. 8473–8481

2023
[12]

Temporal graph representation learning with adaptive augmentation contrastive,

H. Chen, P. Jiao, H. Tang, and H. Wu, “Temporal graph representation learning with adaptive augmentation contrastive,” in ECML/PKDD,
[13]

Latent diffusion- based data augmentation for continuous-time dynamic graph model,

Y . Tian, A. Jiang, Q. Huang, J. Guo, and Y . Qi, “Latent diffusion- based data augmentation for continuous-time dynamic graph model,” in Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , 2024, pp. 2900–2911

2024
[14]

Rdgsl: Dynamic graph representation learning with structure learning,

S. Zhang, Y . Xiong, Y . Zhang, Y . Sun, X. Chen, Y . Jiao, and Y . Zhu, “Rdgsl: Dynamic graph representation learning with structure learning,” Proceedings of the 32nd ACM International Conference on Information and Knowledge Management , 2023. [Online]. Available: https://ap...

2023
[15]

Representation learning for dynamic graphs: A survey,

S. M. Kazemi, R. Goel, K. Jain, I. Kobyzev, A. Sethi, P. Forsyth, P. Poupart, and K. M. Borgwardt, “Representation learning for dynamic graphs: A survey,” J. Mach. Learn. Res. , vol. 21, pp. 70:1–70:73,
[16]

Wavelets on graphs via spectral graph theory,

D. K. Hammond, P. Vandergheynst, and R. Gribonval, “Wavelets on graphs via spectral graph theory,” ArXiv, vol. abs/0912.3848, 2009. [Online]. Available: https://api.semanticscholar.org/CorpusID:5593503

2009 arXiv
[17]

Wavelet-based visual analysis of dynamic networks,

A. D. Col, P. Valdivia, F. Petronetto, F. F. Dias, C. T. Silva, and L. G. Nonato, “Wavelet-based visual analysis of dynamic networks,” IEEE Transactions on Visualization and Computer Graphics , vol. 24, pp. 2456–2469, 2018. [Online]. Available: https://api.semanticscholar. org...

2018
[19]

Strong localization in personalized pagerank vectors,

H. Nassar, K. Kloster, and D. F. Gleich, “Strong localization in personalized pagerank vectors,” ArXiv, vol. abs/1509.00016, 2015. [Online]. Available: https://api.semanticscholar.org/CorpusID:8028648

2015 arXiv
[20]

Time-aware random walk diffusion to improve dynamic graph learning,

J. whi Lee and J. Jung, “Time-aware random walk diffusion to improve dynamic graph learning,” ArXiv, vol. abs/2211.01214, 2022. [Online]. Available: https://api.semanticscholar.org/CorpusID:253254924

2022 arXiv
[21]

Edge weight prediction in weighted signed networks,

S. Kumar, F. Spezzano, V . S. Subrahmanian, and C. Faloutsos, “Edge weight prediction in weighted signed networks,” 2016 IEEE 16th International Conference on Data Mining (ICDM) , pp. 221–230, 2016. [Online]. Available: https://api.semanticscholar.org/CorpusID:14025076

2016
[22]

Predicting positive and negative links in online social networks,

J. Leskovec, D. P. Huttenlocher, and J. M. Kleinberg, “Predicting positive and negative links in online social networks,” ArXiv, vol. abs/1003.2429, 2010. [Online]. Available: https://api.semanticscholar. org/CorpusID:7119014

2010 arXiv
[23]

Community interaction and conflict on the web,

S. Kumar, W. L. Hamilton, J. Leskovec, and D. Jurafsky, “Community interaction and conflict on the web,” Proceedings of the 2018 World Wide Web Conference , 2018. [Online]. Available: https: //api.semanticscholar.org/CorpusID:3854959

2018
[24]

Spatio-temporal attentive rnn for node classification in temporal attributed graphs,

D. Xu, W. Cheng, D. Luo, X. Liu, and X. Zhang, “Spatio-temporal attentive rnn for node classification in temporal attributed graphs,” in International Joint Conference on Artificial Intelligence, 2019. [Online]. Available: https://api.semanticscholar.org/CorpusID:199466175

2019
[25]

Semi-supervised classification with graph convolutional networks,

T. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” ArXiv, vol. abs/1609.02907, 2016. [Online]. Available: https://api.semanticscholar.org/CorpusID:3144218

2016 arXiv
[26]

Roland: graph learning framework for dynamic graphs,

J. You, T. Du, and J. Leskovec, “Roland: graph learning framework for dynamic graphs,” in Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining , 2022, pp. 2358–2366

2022
[27]

High-order topology- enhanced graph convolutional networks for dynamic graphs,

J. Zhu, B. Li, Z. Zhang, L. Zhao, and H. Li, “High-order topology- enhanced graph convolutional networks for dynamic graphs,” Symmetry, vol. 14, no. 10, p. 2218, 2022

2022
[28]

Dergcn: Dynamic-evolving graph convolutional networks for human trajectory prediction,

J. Mi, X. Zhang, H. Zeng, and L. Wang, “Dergcn: Dynamic-evolving graph convolutional networks for human trajectory prediction,” Neuro- computing, vol. 569, p. 127117, 2024

2024
[29]

Pytorch geometric temporal: Spatiotemporal signal processing with neural machine learning models,

B. Rozemberczki, P. Scherer, Y . He, G. Panagopoulos, M. S. Astefanoaei, O. Kiss, F. B ´eres, N. Collignon, and R. Sarkar, “Pytorch geometric temporal: Spatiotemporal signal processing with neural machine learning models,” Proceedings of the 30th ACM International Conference o...

2021
[30]

Deep graph library: A graph-centric, highly-performant package for graph neural networks,

M. Wang, D. Zheng, Z. Ye, Q. Gan, M. Li, X. Song, J. Zhou, C. Ma, L. Yu, Y . Gai, T. Xiao, T. He, G. Karypis, J. Li, and Z. Zhang, “Deep graph library: A graph-centric, highly-performant package for graph neural networks,” arXiv: Learning , 2019. [Online]. Available: https://a...

2019
[2019]

Available: https://api.semanticscholar.org/CorpusID: 216608194

[Online]. Available: https://api.semanticscholar.org/CorpusID: 216608194
[2023]

Available: https://api.semanticscholar.org/CorpusID: 262088334

[Online]. Available: https://api.semanticscholar.org/CorpusID: 262088334

Pith tools

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

[1] [18]

Discrete signal processing on graphs: Frequency analysis,

A. Sandryhaila and J. M. F. Moura, “Discrete signal processing on graphs: Frequency analysis,” IEEE Transactions on Signal Processing, vol. 62, pp. 3042–3054, 2013. [Online]. Available: https://api.semanticscholar.org/CorpusID:12110057

work page 2013

[2] [1]

Data augmentation for graph neural networks,

T. Zhao, Y . Liu, L. Neves, O. J. Woodford, M. Jiang, and N. Shah, “Data augmentation for graph neural networks,” in AAAI Conference on Artificial Intelligence , 2020. [Online]. Available: https://api.semanticscholar.org/CorpusID:219635816

work page 2020

[3] [2]

Graph contrastive learning with adaptive augmentation,

Y . Zhu, Y . Xu, F. Yu, Q. Liu, S. Wu, and L. Wang, “Graph contrastive learning with adaptive augmentation,” Proceedings of the Web Conference 2021 , 2020. [Online]. Available: https://api. semanticscholar.org/CorpusID:225094367

work page 2021

[4] [3]

Spectral Feature Augmentation for Graph Contrastive Learning and Beyond

Y . Zhang, H. Zhu, Z. Song, P. Koniusz, and I. King, “Spectral feature augmentation for graph contrastive learning and beyond,” ArXiv, vol. abs/2212.01026, 2022. [Online]. Available: https://api.semanticscholar. org/CorpusID:254221169

work page Pith review arXiv 2022

[5] [4]

Unleashing the power of graph data augmentation on covariate distribution shift,

Y . Sui, X. Wang, J. Wu, A. Zhang, and X. He, “Unleashing the power of graph data augmentation on covariate distribution shift,” in Neural Information Processing Systems , 2022. [Online]. Available: https://api.semanticscholar.org/CorpusID:253384147

work page 2022

[6] [5]

Structured sequence modeling with graph convolutional recurrent networks,

Y . Seo, M. Defferrard, P. Vandergheynst, and X. Bresson, “Structured sequence modeling with graph convolutional recurrent networks,” in International Conference on Neural Information Processing , 2016. [Online]. Available: https://api.semanticscholar.org/CorpusID:2687749

work page 2016

[7] [6]

Evolvegcn: Evolving graph convolutional networks for dynamic graphs,

A. Pareja, G. Domeniconi, J. Chen, T. Ma, T. Suzumura, H. Kanezashi, T. Kaler, and C. E. Leisersen, “Evolvegcn: Evolving graph convolutional networks for dynamic graphs,” ArXiv, vol. abs/1902.10191, 2019. [Online]. Available: https://api.semanticscholar.org/CorpusID:67856459

arXiv 1902

[8] [7]

Dysat: Deep neural representation learning on dynamic graphs via self- attention networks,

A. Sankar, Y . Wu, L. Gou, W. Zhang, and H. Yang, “Dysat: Deep neural representation learning on dynamic graphs via self- attention networks,” Proceedings of the 13th International Conference on Web Search and Data Mining , 2020. [Online]. Available: https://api.semanticscholar.org/CorpusID:210884001

work page 2020