Pith. sign in

REVIEW 3 major objections 6 minor 47 references

Dynamic Graph Condensation

T0 review · 3 major / 6 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 This paper introduces DyGC, the first framework for dynamic graph condensation, which distills a time-evolving graph into a compact synthetic graph that preserves spatiotemporal structure; a DGNN trained on 0.5% of the data keeps up to…

desk verdict First dynamic graph condensation method with broad, convincing experiments, but a false theoretical proposition and over-claimed fidelity need fixing. read the letter →

arxiv 2506.13099 v1 pith:BM3AUAW5 submitted 2025-06-16 cs.LG cs.AI

Dong Chen , Shuai Zheng , Yeyu Yan , Muhao Xu , Zhenfeng Zhu , Yao Zhao , Kunlun He This is my paper
classification cs.LGcs.AI
keywords dynamicgraphcondensationspatiotemporaldistributionmatchingspikingneuralnetworksstateevolvingfieldtemporalnodeclassificationmaximummeandiscrepancy
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 record how connections and node attributes change over time, but they are bulky and expensive to train on. This paper introduces DyGC, which compresses an entire sequence of graph snapshots into a tiny synthetic dynamic graph that a dynamic graph neural network (DGNN) can train on instead. The central claim is that the compressed graph can preserve the real graph's spatiotemporal distribution closely enough that a DGNN trained on it keeps most of its accuracy: up to 96.2% of full-graph performance using only 0.5% of the original graph size, with training speedups up to 1,846 times. The guiding insight is that static graph condensation fails on dynamic graphs because matching each snapshot separately ignores temporal dependencies, so DyGC aligns the joint spatiotemporal state rather than per-snapshot marginals. A sympathetic reader would care because this makes large-scale time-evolving graph learning far more data-efficient.

What carries the argument

The three load-bearing pieces are: (1) the spiking structure generator, a leaky integrate-and-fire mechanism (Eqs. 4–5) that turns node features into a temporally evolving binary adjacency matrix, with surrogate-gradient training to handle the non-differentiable firing threshold; (2) the state evolving field (Eq. 6), a recursive operator that mixes spatial message passing $\mathbf{M}^{(t)}$ with temporal smoothing across snapshots, producing a tensor of node states that represents the graph's spatiotemporal dynamics; (3) class-wise maximum mean discrepancy on that field, which implicitly compares all moments of the state distributions and, together with logit alignment, drives the optimization of the synthetic graph. Together they convert the intractable problem of matching two dynamic-graph distributions into matching finite state tensors in a fixed semantic space.

What would settle it

Take the real and condensed graphs for a small dataset and estimate the KL divergence between their full joint distributions over snapshots; if the joint divergence remains large while per-snapshot marginal divergences are near zero, the state evolving field has failed to capture the temporal dependencies that DyGC claims to preserve.

Watch

Extended reading notes

Core claim

DyGC condenses a discrete-time dynamic graph $\mathcal{T}=(\mathbf{A},\mathbf{X},\mathbf{Y})$ into a small synthetic graph $\mathcal{S}$ by jointly optimizing synthetic node features and an evolving binary adjacency structure so that the spatiotemporal distribution of $\mathcal{S}$ matches that of $\mathcal{T}$. Structure is produced by a spiking structure generation module, which treats inter-node affinity like the membrane voltage of a leaky integrate-and-fire neuron: affinity integrates past interaction memory with current feature-based stimulus, and an edge appears when the voltage crosses a threshold, with a soft reset after firing. Distribution matching is done in a 'state evolving field' $\mathbf{H}^{(t,k)}=\alpha\mathbf{H}^{(t-1,k)}+(1-\alpha)\mathbf{M}^{(t)}\mathbf{H}^{(t,k-1)}$, a recursive blend of spatial message passing and temporal smoothing, and the discrepancy between real and synthetic fields is measured class-wise with maximum mean discrepancy, plus a logit alignment term from a pretrained DGNN. On DBLP, Reddit, Arxiv, and Tmall, condensed graphs from DyGC outperform snapshot-wise static condensation baselines across five DGNN architectures, with reported performance fidelity of 86.4%–98.7% and large reductions in storage and training time.

Load-bearing premise

The condensed graph is optimized to match the real graph inside a hand-built 'state evolving field' that mixes one-step spatial and temporal message passing; if that field discards information a particular dynamic GNN needs, the condensed graph will underperform for that architecture.

Editorial extensions

If this is right

  • A DGNN trained on a DyGC-condensed graph can reach 86.4%–98.7% of its full-graph accuracy on temporal node classification across five architectures, so expensive training can run on drastically smaller data.
  • Snapshot-wise application of static condensation methods to dynamic graphs is provably insufficient: Proposition 3.1 shows that matching per-snapshot marginals only aligns the joint distribution when snapshots are temporally independent, which real dynamic graphs are not.
  • Condensation time is lower than static methods in most settings and grows more slowly with condensation ratio, because DyGC processes the whole snapshot sequence jointly rather than repeating condensation per snapshot.
  • The spiking structure generator yields discrete, unweighted edges whose consecutive-snapshot Jaccard similarity tracks that of real graphs, indicating structural continuity rather than rigidity.

Reading between the lines

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

  • The state evolving field is a linear blend of one-step spatial diffusion and one-step temporal smoothing; architectures that rely on long-range temporal memory (for example recurrent or state-space DGNNs) may carry information the field never represents, which would limit the claimed fidelity beyond the five tested backbones.
  • The discrete-time snapshot setting is a genuine scope restriction: event-based or continuous-time graphs require a different condensation objective, and the paper's own appendix acknowledges that discretization can lose temporal precision.
  • The method is evaluated on temporal node classification; for link prediction or anomaly detection, the information a condensed graph must preserve is different, so the same spatiotemporal matching objective would likely need re-design rather than simple reuse.
  • The spiking structure generator is a means to produce discrete evolving edges, not the only one; the transferable contribution is the joint spatiotemporal distribution-matching objective, which could be paired with other differentiable structure generators.
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 paper introduces DyGC, a method for dynamic graph condensation that synthesizes a small discrete-time dynamic graph from a large one. It uses a spiking-neuron-based structure generator (SSG) and a distribution matching objective based on maximum mean discrepancy (MMD) in a recursively constructed spatiotemporal 'state evolving field' (Eq. 6), plus a logit alignment term from a pretrained DGNN. Experiments on four dynamic graph datasets and five DGNN backbones report that DyGC outperforms per-snapshot static condensation and coreset baselines, retaining up to 96.2% fidelity at 0.5% of the original graph size and achieving large training speedups.

Significance. If the empirical results hold, DyGC is a practically valuable contribution to data-centric dynamic graph learning. The paper provides broad evidence: four datasets of varying scale, five diverse DGNN backbones, ablations of each module, temporal continuity analysis, and explicit storage and training-speedup quantification. The setup is described in sufficient detail in Appendix E, and Algorithm 1 in Appendix D clarifies the pipeline. However, the theoretical motivation contains a false proposition (Prop 3.1), the headline fidelity claim is based on the best metric and ratio (Micro-F1 on Arxiv at the largest ratio) while Macro-F1 fidelity drops to 86.8% in the same cell, and the transfer from the hand-designed linear state field to arbitrary DGNN accuracy is an unproven modeling assumption rather than a derived guarantee.

major comments (3)
  1. [Section 3, Prop. 3.1 and Appendix C] The claimed 'iff' in Proposition 3.1 is false. If the two dynamic graphs have identical joint distributions (e.g., P = Q with perfectly dependent snapshots), then the joint KL divergence and the sum of marginal KL divergences are both zero, so equality holds even though neither sequence is temporally independent. The proof in Appendix C only constructs one configuration where non-independence leads to inequality; it does not establish necessity. The motivational point that marginal-only alignment is insufficient can be preserved, but the formal statement and proof must be corrected, for example by restating the result as an impossibility example rather than an equivalence.
  2. [Section 4.2, Eq. (6) and Eq. (7)] The distribution matching loss compares only the hand-designed linear state evolving field H, defined as a fixed spatiotemporal diffusion, and no argument is given that small MMD in this field transfers to small error for arbitrary DGNNs (e.g., DySAT's self-attention or ROLAND's hierarchical states). The paper's claim that DyGC 'faithfully preserves the inherent spatiotemporal characteristics' is therefore stronger than what is established; generalization across the five backbones is empirical and may depend on the chosen field. In addition, the logit alignment in Section 4.3 uses T-GCN as the only pretrained architecture; the paper should clarify whether the same condensed graph is used for all backbones and discuss the effect of this task-model mismatch.
  3. [Abstract and Table 2] The headline claim 'retains up to 96.2% DGNN performance' is the Micro-F1 fidelity for T-GCN on Arxiv at the largest condensation ratio, while the Macro-F1 fidelity for the same cell is only 86.8%. The 'up to' phrasing is technically correct but obscures that Macro-F1 fidelity is substantially lower. Please report both metrics in the abstract/conclusion or explicitly discuss the fidelity range (86.4%–98.7%) and the Micro/Macro gap, which is relevant for class-imbalanced applications.
minor comments (6)
  1. [Section 2] The notation D_KL(P_{G_{1:T}} \| Q_{\tilde{G}_{1:T}}) in the proposition is not formally defined for distributions over different graph sequences; clarify that Q denotes the distribution of the synthetic dynamic graph.
  2. [Section 4.1, Eq. (4)] The update equation for \hat{U}^{(t)}_{ij} uses U^{(t-1)}_{ij} on the right-hand side but does not explicitly define the integration time step or the reset mechanism; please provide the full discretized LIF update.
  3. [Section 4.2, Eq. (8)] The kernel function \psi over tensors of shape K × T × d may suffer from high-dimensional MMD; the paper should specify how the RBF kernel bandwidth is chosen and whether features are normalized.
  4. [Figure 4] The y-axis label 'Scale = ln(epoch/sec)' is inconsistent with the plotted values (which appear to be speedup factors); replace it with a clearly defined speedup metric.
  5. [Table 1 and 2] The 'Whole' column is a full-graph upper bound, not a baseline method; it should be labeled accordingly to avoid confusion.
  6. [Appendix E.4] The sensitivity analysis covers only the number of spatial propagation steps K; please add sensitivity studies for the temporal coefficient \alpha, the membrane decay \tau_m, the firing threshold U_th, and the logit-alignment weight \gamma.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: reported fidelity is measured held-out DGNN accuracy, not the MMD training objective; minor overlaps (T-GCN teacher/evaluation overlap, self-cited datasets) are not load-bearing.

full rationale

The central claim—that DyGC produces condensed dynamic graphs on which DGNNs can be trained with high fidelity—is evaluated by training DGNNs on the condensed graph and testing on the real graph's held-out test nodes (Tables 1 and 2). The distillation objective in Eq. 7 is an MMD loss over the hand-designed state evolving field of Eq. 6; this is a training loss, not the reported fidelity metric. The reported 'Fidelity' values are ratios of measured test accuracies (condensed vs. whole graph), so they are not fitted quantities and are not equal to the MMD loss by construction. Logit alignment (Section 4.3) uses a pretrained T-GCN as teacher, and T-GCN is also one of the five evaluated backbones in Table 2; this is a mild self-distillation overlap, but the same condensed graph is also evaluated on GCRN, STGCN, DySAT, and ROLAND, architectures not used as the teacher, so the generalization claim has independent empirical content. The self-citation to [20] supplies public datasets only and is not load-bearing. The paper does contain a rigor concern: Proposition 3.1's 'iff' assertion is not fully established by the Appendix C proof, since the counterexample shows only that dependence can break additivity, not general necessity; however, this is a mathematical correctness issue, not a circular derivation. Likewise, the assumption that small MMD in the linear field of Eq. 6 transfers to arbitrary DGNN accuracy is an unproven modeling assumption, but an unproven assumption is not circularity because the downstream accuracy is measured rather than defined by the loss. Overall, no step in the claimed derivation reduces to its own inputs by construction, so circularity is minimal.

Assumptions & free parameters 6 free parameters · 4 assumptions · 1 invented entities

The method relies on several hand-chosen hyperparameters and two unproven modeling assumptions: the sufficiency of the state evolving field for preserving DGNN performance, and the adequacy of spiking neurons as a graph generator. These are the main elements the reader pays for beyond the empirical recipe.

free parameters (6)
  • alpha (temporal propagation coefficient) = 0.5 (fixed)
    Temporal propagation coefficient in the state evolving field (Eq. 6); set by hand, not tuned, and no sensitivity analysis except K.
  • K (number of spatial propagation steps) = not explicitly stated; experiments test 1 to 5
    Controls the depth of spatial message passing in the state evolving field; chosen to balance performance and stability.
  • gamma (logit alignment weight) = not reported
    Trade-off weight between distribution matching loss and logit alignment loss; not disclosed in Appendix E.
  • beta (surrogate gradient sharpness) = not reported
    Sharpness of the surrogate gradient for the Heaviside step function (Appendix D); critical for training stability.
  • tau_m (membrane decay factor) = learnable
    Membrane decay factor in the LIF spiking structure generation (Eq. 4); learned during condensation.
  • U_th (firing threshold) = learnable
    Firing threshold for edge generation (Eq. 5); learned to accommodate varying graph densities.
assumptions (4)
  • standard math KL divergence chain rule
    Used in the proof of Proposition 3.1 in Appendix C.
  • domain assumption Discrete-time dynamic graph with a fixed node set across snapshots
    Stated in Section 2: 'Without loss of generality, we assume that all snapshots share a consistent and fixed node set.' This limits applicability to continuous-time or node-evolving graphs.
  • domain assumption The hand-crafted state evolving field is a sufficient representation for DGNN behavior
    The central heuristic: matching MMD in this field transfers to downstream accuracy. Unproven in the paper.
  • ad hoc to paper Spiking structure generation can produce realistic evolving topologies
    The LIF-based generator is motivated by biological analogy, but there is no independent evidence that binary spiking outputs capture real structural evolution; ablation shows modest gains.
invented entities (1)
  • State evolving field H
    purpose: A tensor representation of dynamic graph states used for distribution matching
    Defined by Eq. 6 in Section 4.2; no falsifiable prediction outside the paper. It is a mathematical construct rather than a physically motivated entity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dynamic Graph Condensation." pith.science (2026). https://pith.science/paper/BM3AUAW5

@misc{pith2026250613099,
  author       = {Pith},
  title        = {Pith review of: Dynamic Graph Condensation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BM3AUAW5}},
  note         = {Machine review of arXiv:2506.13099}
}
read the original abstract

Recent research on deep graph learning has shifted from static to dynamic graphs, motivated by the evolving behaviors observed in complex real-world systems. However, the temporal extension in dynamic graphs poses significant data efficiency challenges, including increased data volume, high spatiotemporal redundancy, and reliance on costly dynamic graph neural networks (DGNNs). To alleviate the concerns, we pioneer the study of dynamic graph condensation (DGC), which aims to substantially reduce the scale of dynamic graphs for data-efficient DGNN training. Accordingly, we propose DyGC, a novel framework that condenses the real dynamic graph into a compact version while faithfully preserving the inherent spatiotemporal characteristics. Specifically, to endow synthetic graphs with realistic evolving structures, a novel spiking structure generation mechanism is introduced. It draws on the dynamic behavior of spiking neurons to model temporally-aware connectivity in dynamic graphs. Given the tightly coupled spatiotemporal dependencies, DyGC proposes a tailored distribution matching approach that first constructs a semantically rich state evolving field for dynamic graphs, and then performs fine-grained spatiotemporal state alignment to guide the optimization of the condensed graph. Experiments across multiple dynamic graph datasets and representative DGNN architectures demonstrate the effectiveness of DyGC. Notably, our method retains up to 96.2% DGNN performance with only 0.5% of the original graph size, and achieves up to 1846 times training speedup.

Figures

Figures reproduced from arXiv: 2506.13099 by the authors.

Figure 1
Figure 1. Comparisons of existing GC vs dynamic graph condensation (DyGC) on dynamic graphs. Unfortunately, although existing GC methods significantly improve data efficiency on static graphs, they show poor adaptability in dy￾namic settings. As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overall pipeline of the proposed Dynamic Graph Condensation (DyGC) framework. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The time complexity comparison of different graph condensation methods. Comparison with Baseline Methods. We evaluate the temporal node classification performance of DyGC in comparison to other graph condensation (GC) meth￾ods, using T-GCN as a unified DGNN architecture. The overall results are summarized in [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Storage reduction and DGNN training speedup (Scale [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 6
Figure 6. Figure 6: Temporal continuity of struc￾tures (Jaccard similarity) between con￾secutive time steps. 5.3 Ablation Study Impact of Key Modules on Overall Performance (Micro-F1 %). As presented in [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Micro-F1 (%) and Macro-F1 (%) results on the Arxiv with various spatial propagation steps [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 34 canonical work pages

  1. [1]

    J. Chen, X. Wang, and X. Xu. Gc-lstm: Graph convolution embedded lstm for dynamic network link prediction.Applied Intelligence, pages 1–16, 2022

    work page 2022

  2. [2]

    T. M. Cover.Elements of information theory. John Wiley & Sons, 1999

    1999

  3. [3]

    Z. Cui, Z. Li, S. Wu, X. Zhang, Q. Liu, L. Wang, and M. Ai. Dygcn: Efficient dynamic graph embedding with graph convolutional network.IEEE Transactions on Neural Networks and Learning Systems, 35(4):4635–4646, 2022

    work page 2022

  4. [4]

    W. Fang, Z. Yu, Y . Chen, T. Masquelier, T. Huang, and Y . Tian. Incorporating learnable membrane time constant to enhance learning of spiking neural networks.arXiv preprint arXiv:2007.05785, 2020

    work page Pith review arXiv 2007

  5. [5]

    X. Gao, G. Ye, T. Chen, W. Zhang, J. Yu, and H. Yin. Rethinking and accelerating graph condensation: A training-free approach with class partition. InProceedings of the ACM on Web Conference 2025, pages 4359–4373, 2025

    work page 2025

  6. [6]

    X. Gao, J. Yu, T. Chen, G. Ye, W. Zhang, and H. Yin. Graph condensation: A survey.IEEE Transactions on Knowledge and Data Engineering, 2025

    work page 2025

  7. [7]

    Gretton, K

    A. Gretton, K. M. Borgwardt, M. J. Rasch, B. Schölkopf, and A. Smola. A kernel two-sample test.The Journal of Machine Learning Research, 13(1):723–773, 2012

    2012

  8. [8]

    Hamilton, Z

    W. Hamilton, Z. Ying, and J. Leskovec. Inductive representation learning on large graphs. Advances in neural information processing systems, 30, 2017

    2017

Show all 47 references
  1. [9]

    Hashemi, S

    M. Hashemi, S. Gong, J. Ni, W. Fan, B. A. Prakash, and W. Jin. A comprehensive survey on graph reduction: Sparsification, coarsening, and condensation. InProceedings of the Thirty- Third International Joint Conference on Artificial Intelligence, IJCAI 2024, Jeju, South Korea, ...

    2024

  2. [10]

    Hochreiter and J

    S. Hochreiter and J. Schmidhuber. Long short-term memory.Neural computation, 9(8):1735– 1780, 1997

    1997

  3. [11]

    C. Huan, Y . Liu, H. Zhang, H. Liu, S. Chen, S. L. Song, and Y . Wu. Tegraph+: Scalable temporal graph processing enabling flexible edge modifications.IEEE Transactions on Parallel and Distributed Systems, 35(8):1469–1487, 2024

    2024

  4. [12]

    Hunsberger and C

    E. Hunsberger and C. Eliasmith. Spiking deep networks with LIF neurons.CoRR, abs/1510.08829, 2015

    2015 arXiv

  5. [13]

    E. M. Izhikevich. Which model to use for cortical spiking neurons?IEEE transactions on neural networks, 15(5):1063–1070, 2004

    2004

  6. [14]

    W. Jin, X. Tang, H. Jiang, Z. Li, D. Zhang, J. Tang, and B. Yin. Condensing graphs via one-step gradient matching. InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 720–730, 2022

    2022

  7. [15]

    W. Jin, L. Zhao, S. Zhang, Y . Liu, J. Tang, and N. Shah. Graph condensation for graph neural networks.arXiv preprint arXiv:2110.07580, 2021

    2021 arXiv

  8. [16]

    S. M. Kazemi, R. Goel, K. Jain, I. Kobyzev, A. Sethi, P. Forsyth, and P. Poupart. Representation learning for dynamic graphs: A survey.Journal of Machine Learning Research, 21(70):1–73, 2020

    2020

  9. [17]

    Y . Kim, Y . Li, H. Park, Y . Venkatesha, A. Hambitzer, and P. Panda. Exploring temporal information dynamics in spiking neural networks. InProceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 8308–8316, 2023

    2023

  10. [18]

    T. N. Kipf and M. Welling. Semi-supervised classification with graph convolutional networks. InInternational Conference on Learning Representations, 2017. 10

    2017

  11. [19]

    J. Li, R. Wu, X. Jin, B. Ma, L. Chen, and Z. Zheng. State space models on temporal graphs: A first-principles study.Advances in Neural Information Processing Systems, 37:127030–127058, 2024

    2024

  12. [20]

    J. Li, Z. Yu, Z. Zhu, L. Chen, Q. Yu, Z. Zheng, S. Tian, R. Wu, and C. Meng. Scaling up dynamic graph representation learning via spiking neural networks. InProceedings of the AAAI conference on artificial intelligence, volume 37, pages 8588–8596, 2023

    2023

  13. [21]

    M. Liu, S. Li, X. Chen, and L. Song. Graph condensation via receptive field distribution matching.CoRR, abs/2206.13697, 2022

    2022 arXiv

  14. [22]

    J. D. Nunes, M. Carvalho, D. Carneiro, and J. S. Cardoso. Spiking neural networks: A survey. IEEE access, 10:60738–60764, 2022

    2022

  15. [23]

    Pareja, G

    A. Pareja, G. Domeniconi, J. Chen, T. Ma, T. Suzumura, H. Kanezashi, T. Kaler, T. B. Schardl, and C. E. Leiserson. Evolvegcn: Evolving graph convolutional networks for dynamic graphs. InProceedings of the AAAI conference on artificial intelligence, pages 5363–5370. AAAI Press, 2020

    2020

  16. [24]

    Rossi, B

    E. Rossi, B. Chamberlain, F. Frasca, D. Eynard, F. Monti, and M. M. Bronstein. Temporal graph networks for deep learning on dynamic graphs.CoRR, abs/2006.10637, 2020

    2006 arXiv

  17. [25]

    Sankar, Y

    A. Sankar, Y . Wu, L. Gou, W. Zhang, and H. Yang. Dysat: Deep neural representation learning on dynamic graphs via self-attention networks. InProceedings of the 13th international conference on web search and data mining, pages 519–527, 2020

    2020

  18. [26]

    Sener and S

    O. Sener and S. Savarese. Active learning for convolutional neural networks: A core-set approach. InInternational Conference on Learning Representations, 2018

    2018

  19. [27]

    Y . Seo, M. Defferrard, P. Vandergheynst, and X. Bresson. Structured sequence modeling with graph convolutional recurrent networks. InNeural information processing: 25th international conference, ICONIP 2018, Siem Reap, Cambodia, December 13-16, 2018, proceedings, part I 25, p...

    2018

  20. [28]

    Skarding, B

    J. Skarding, B. Gabrys, and K. Musial. Foundations and modeling of dynamic networks using dynamic graph neural networks: A survey.iEEE Access, 9:79143–79168, 2021

    2021

  21. [29]

    Trivedi, M

    R. Trivedi, M. Farajtabar, P. Biswal, and H. Zha. Dyrep: Learning representations over dynamic graphs. InInternational conference on learning representations, 2019

    2019

  22. [30]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017

    2017

  23. [31]

    Veliˇckovi´c, G

    P. Veliˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P. Liò, and Y . Bengio. Graph attention networks. InInternational Conference on Learning Representations, 2018

    2018

  24. [32]

    Wang, J.-Y

    T. Wang, J.-Y . Zhu, A. Torralba, and A. A. Efros. Dataset distillation.arXiv preprint arXiv:1811.10959, 2018

    2018 arXiv

  25. [33]

    M. Welling. Herding dynamical weights to learn. InProceedings of the 26th annual international conference on machine learning, pages 1121–1128, 2009

    2009

  26. [34]

    Z. Wu, S. Pan, F. Chen, G. Long, C. Zhang, and P. S. Yu. A comprehensive survey on graph neural networks.IEEE transactions on neural networks and learning systems, 32(1):4–24, 2020

    2020

  27. [35]

    F. Xia, K. Sun, S. Yu, A. Aziz, L. Wan, S. Pan, and H. Liu. Graph learning: A survey.IEEE Transactions on Artificial Intelligence, 2(2):109–127, 2021

    2021

  28. [36]

    Z. Xiao, Y . Wang, S. Liu, B. Hu, H. Wang, M. Song, and T. Zheng. Disentangled condensation for large-scale graphs. In G. Long, M. Blumestein, Y . Chang, L. Lewin-Eytan, Z. H. Huang, and E. Yom-Tov, editors,Proceedings of the ACM on Web Conference 2025, WWW 2025, Sydney, NSW, ...

    2025

  29. [37]

    Z. Xiao, Y . Wang, S. Liu, H. Wang, M. Song, and T. Zheng. Simple graph condensation. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 53–71. Springer, 2024

    2024

  30. [38]

    D. Xu, C. Ruan, E. Körpeoglu, S. Kumar, and K. Achan. Inductive representation learning on temporal graphs. In8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net, 2020

    2020

  31. [39]

    B. Yang, K. Wang, Q. Sun, C. Ji, X. Fu, H. Tang, Y . You, and J. Li. Does graph distillation see like vision dataset counterpart? In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors,Advances in Neural Information Processing Systems 36: Annual Confer...

    2023

  32. [40]

    L. Yang, C. Chatelain, and S. Adam. Dynamic graph representation learning with neural networks: A survey.IEEE Access, 12:43460–43484, 2024

    2024

  33. [41]

    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, pages 2358–2366, 2022

    2022

  34. [42]

    B. Yu, H. Yin, and Z. Zhu. Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting. In J. Lang, editor,Proceedings of the Twenty-Seventh Inter- national Joint Conference on Artificial Intelligence, IJCAI 2018, July 13-19, 2018, Stockholm...

    2018

  35. [43]

    Zhang, P

    Z. Zhang, P. Cui, and W. Zhu. Deep learning on graphs: A survey.IEEE Transactions on Knowledge and Data Engineering, 34(1):249–270, 2020

    2020

  36. [44]

    L. Zhao, Y . Song, C. Zhang, Y . Liu, P. Wang, T. Lin, M. Deng, and H. Li. T-gcn: A tem- poral graph convolutional network for traffic prediction.IEEE transactions on intelligent transportation systems, 21(9):3848–3858, 2019

    2019

  37. [45]

    Zheng, M

    X. Zheng, M. Zhang, C. Chen, Q. V . H. Nguyen, X. Zhu, and S. Pan. Structure-free graph condensation: From large-scale graphs to condensed graph-free data.Advances in Neural Information Processing Systems, 36:6026–6047, 2023. 12 Appendix This appendix supplements our work,Dyna...

    2023

  38. [46]

    Limitations

    when constructing the state evolution field. As shown in Figure 7, the performance improves noticeably when increasing K from 1 to 3. This is because more propagation steps allow for intro- ducing richer structural context into the state field, thereby achieving a more precise...

  39. [47]

    Guidelines: • The answer NA means that the paper does not involve crowdsourcing nor research with human subjects

    Institutional Review Board (IRB) Approvals or Equivalent for Research with Human Subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals...

Pith tools

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