Pith. sign in

REVIEW 3 major objections 4 minor 66 references

Simple yet Effective Graph Distillation via Clustering

T0 review · 3 major / 4 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 ClustGDD claims that graph distillation reduces to K-Means clustering on Laplacian-smoothed embeddings, producing condensed graphs that match or beat state-of-the-art GDD accuracy while running orders of magnitude faster.

desk verdict Solid empirical graph distillation paper undone by a false central theorem; fix the theory or drop it. read the letter →

arxiv 2505.20807 v1 pith:D46RB45B submitted 2025-05-27 cs.LG

Yurui Lai , Taiyan Zhang , Renchi Yang This is my paper
classification cs.LG
keywords graphdatadistillationcondensationneuralnetworksnodeclassificationK-MeansclusteringFréchetInceptionDistanceLaplaciansmoothinghomophily
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

ClustGDD claims that distilling a large graph into a small surrogate for GNN training need not involve expensive gradient, trajectory, or spectrum matching: a plain clustering partition of the original nodes does the job. The paper shows that the Fréchet Inception Distance (FID) between original and condensed node representations—a quality metric borrowed from image generation—is bounded by two quantities that clustering directly controls, namely cluster-size variance and the within-cluster sum of squares (WCSS). That makes K-Means over graph-Laplacian-smoothed, label-supervised embeddings a theoretically grounded condensation algorithm, and a light class-aware refinement step repairs the residual over-smoothing caused by heterophilic edges. If the central claim is right, practitioners can distill graphs such as Cora, arXiv, or Reddit in tens of seconds instead of hours while matching or improving downstream node-classification accuracy.

What carries the argument

The load-bearing object is the Fréchet Inception Distance (FID) between the original and condensed node representations, paired with its two bounding theorems. Theorem 3.1 bounds the mean-shift term $\|\boldsymbol{\mu}_{\mathrm{org}} - \boldsymbol{\mu}_{\mathrm{syn}}\|_2^2$ by $\frac{1}{N^2}\sum_{i=1}^{n}\left(\frac{N}{n} - |C_i|\right)^2$, meaning unbalanced clusters inflate the FID; Theorem 3.2 bounds the covariance term by the within-cluster sum of squares $\frac{1}{N}\sum_{i=1}^n \sum_{v_j \in C_i} \|H_j - H'_i\|_2^2$ plus a constant determined by the original graph. Together the theorems convert graph distillation into a K-Means task on embeddings $H = ZW$, where $Z = \sum_{t=0}^{T}(1-\alpha)\alpha^t \tilde{A}^t X$ is the closed-form graph Laplacian smoothing solution (Lemma 4.1) and $W$ is trained with cross-entropy so that same-class nodes become close and clusters align with classes. The condensation itself is the sketching operation $X' = \tilde{C}^\top Z$, $A' = \tilde{C}^\top \tilde{A} \tilde{C}$, with labels taken as the argmax of cluster-averaged predictions. CAAR (Class-Aware Attribute Refinement) is the residual repair mechanism: it samples class-specific subgraphs whose edges are weighted by predicted-class probability times an estimated effective resistance, builds $K$ condensed views, and learns a small augmentation $\Delta$ that enlarges inter-class attribute distances under the combined loss $\mathcal{L}_{\mathrm{org}} + \gamma \mathcal{L}_{\mathrm{syn}} + \lambda \mathcal{L}_{\mathrm{cst}}$.

What would settle it

Apply ClustGDD to a graph whose homophily has been deliberately broken, for instance by reshuffling node labels so that adjacent nodes rarely share a class, and check whether the K-Means clusters mix ground-truth classes and whether GNNs trained on the condensed graph lose accuracy relative to training on the full graph; as a second check, measure the FID of such a condensed graph and test whether the claimed FID–accuracy correlation from the paper's Fig. 1 still holds.

Watch

Extended reading notes

Core claim

The central claim is that a high-quality condensed graph is the synthetic lift of a balanced, low-WCSS partition of the original graph, rather than the output of matching gradients, training trajectories, or eigenbases. When each synthetic node is the averaged embedding of one cluster of original nodes, the paper proves that the FID between the two graphs is bounded by the cluster-size variance (Theorem 3.1) plus a term controlled by the within-cluster sum of squares (Theorem 3.2), so minimizing WCSS under balanced cluster sizes provably drives condensation quality. ClustGDD realizes this by smoothing attributes with the closed-form graph Laplacian smoothing solution, training a linear classifier so that same-class nodes sit close in embedding space, and running K-Means; the cluster sketching matrix then yields synthetic attributes $X' = \tilde{C}^\top Z$, adjacency $A' = \tilde{C}^\top \tilde{A} \tilde{C}$, and labels. Because this homophily-driven clustering over-smooths the features of adjacent nodes in different classes, a class-aware attribute refinement module (CAAR) learns a small augmentation from class-specific sampled subgraphs, trained with supervised and consistency losses, which restores inter-class attribute distances. The empirical claim is that GNNs trained on the resulting condensed graphs match or beat state-of-the-art distillation baselines on five benchmark datasets, with an average condensation time of 56.8 seconds versus $10^3$–$10^5$ seconds for competitors.

Load-bearing premise

The whole method rests on the assumption that after graph Laplacian smoothing and a label-supervised linear projection, nodes of the same class sit close together in embedding space so that K-Means clusters line up with true classes; when that separation fails—on low-homophily graphs, with noisy labels, or under heterogeneous linking patterns—synthetic nodes will mix classes and the condensed graph will degrade.

Editorial extensions

If this is right

  • GNNs trained on ClustGDD condensed graphs match or exceed state-of-the-art gradient-, trajectory-, eigenbasis-, and kernel-matching distillations in node-classification accuracy on Cora, Citeseer, arXiv, Flickr, and Reddit, sometimes surpassing models trained on the full graph.
  • Condensation cost drops by orders of magnitude: about 56.8 seconds on average across benchmarks, where the compared distillation methods consume $10^3$ to $10^5$ seconds.
  • Condensed graphs transfer across GNN architectures (GCN, SGC, APPNP, ChebyNet, BernNet) with the highest average accuracy and the lowest variance among compared methods, so the distilled data is not tuned to a single model family.
  • FID, which needs no ground-truth labels, tracks downstream node-classification accuracy, providing a label-free way to judge whether a condensed graph is good before training.
  • The method's simplified time complexity $O(ENnd + n^2d + Md + M\log M)$ removes the higher-order terms ($EL$, $L^2$, $N_k N^2$) that make prior distillation methods impractical on large graphs.

Reading between the lines

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

  • Because each condensed node is a supernode over a cluster of real nodes, ClustGDD's synthetic graph carries an interpretable mapping back to the original graph—every synthetic node names the original nodes it absorbed—a property gradient-matched synthetic graphs lack, and one that could serve explanation, unlearning, or debugging of downstream models.
  • The paper's own analysis implies a stress test it does not run: on low-homophily or label-noisy graphs the class-alignment assumption behind K-Means is violated, and although CAAR softens the resulting over-smoothing, the core condensation would be expected to lose accuracy; measuring that drop would delimit the method's applicability.
  • If the FID–accuracy link is as general as the paper suggests, the same clustering pipeline could accept embeddings from any model that maps same-class nodes close together—spectral embeddings, pretrained GNN representations, or feature stores from other modalities—potentially covering heterogeneous and multimodal graphs the authors name as future work.
  • The theory suggests a cheap diagnostic for practice: track the WCSS and cluster-size variance of the K-Means output as a label-free proxy for condensation quality, using them to pick the condensation ratio or to compare candidate embeddings before any downstream training.
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 / 4 minor

Summary. The paper proposes ClustGDD, a graph data distillation method that synthesizes a small condensed graph by (i) computing smoothed node representations via graph Laplacian smoothing plus a supervised linear layer, (ii) clustering those representations with balanced K-Means to create supernodes, and (iii) refining the synthetic attributes with a class-aware augmentation module (CAAR). The authors claim a theoretical connection between low within-cluster sum of squares (WCSS) clustering and low Fréchet Inception Distance (FID) of the condensed graph, and they support the method with an empirical study on five benchmark graphs, comparing against nine baselines, including cross-architecture generalization, ablations, hyper-parameter studies, and timing results. The empirical findings indicate that ClustGDD is substantially faster than most GDD baselines and often achieves accuracy comparable to or better than state-of-the-art methods.

Significance. If the theoretical connection were valid, the paper would make a useful conceptual contribution by replacing expensive gradient/trajectory/spectral matching in GDD with a simple clustering procedure. The empirical study is broad and largely well executed: five datasets, nine baselines, multiple condensation ratios, cross-architecture tests, ablations, and timing comparisons, with code publicly linked. The speed advantage is clearly demonstrated. However, the central theoretical claim is false as stated, and the paper explicitly advertises this claim in the abstract, introduction, and Section 4.1. The empirical results may still stand on their own, but the theoretical grounding that distinguishes the method from a purely heuristic GDD procedure is not established and needs substantial repair or honest re-framing.

major comments (3)
  1. [Section 3.2, Theorem 3.1 and Appendix E] Theorem 3.1 is false. The proof applies Jensen's inequality to the terms ||Σ_j (1/|C_i| − n/N) H_j||^2, but the coefficients a_ij = 1/|C_i| − n/N are not nonnegative and do not sum to 1, so Jensen does not apply. The proof also silently drops the factor ||H_j||^2 when moving to the claimed bound. A concrete counterexample is N=4, n=2, C1={v1,v2,v3}, C2={v4}, H1=H2=H3=e1, H4=−e1. Then H'_1=e1, H'_2=−e1, μ_org=0.5e1, and μ_syn=0, so the left-hand side equals 0.25 while the right-hand side equals (1/16)[(2−3)^2+(2−1)^2]=0.125. Because Theorem 3.1 is the stated basis for connecting balanced, low-WCSS clustering to the mean term of the FID, and Theorem 3.2 explicitly invokes it, the paper's central theoretical claim that ClustGDD is 'theoretically-grounded' is not established. The authors should either prove a correct bound under explicit assumptions on the embeddings or re-frame the clustering construction as a heuristic and adjust the abstract, introduction, and Section 4.1 accordingly.
  2. [Section 3.2, Theorem 3.2 and Eq. (19)] The proof of Theorem 3.2 relies on the inequality Tr((Σ_orgΣ_syn)^{1/2}) ≥ Tr(Σ_org^{1/2}Σ_syn^{1/2}), but the product Σ_orgΣ_syn is not symmetric in general, so the expression (Σ_orgΣ_syn)^{1/2} in Eq. (1) is not the standard PSD square root used in the FID formula, and the displayed inequality is not a direct consequence of the Araki–Lieb–Thirring inequality as cited. Please state explicitly whether Eq. (1) is intended to use Tr((Σ_org^{1/2}Σ_synΣ_org^{1/2})^{1/2}) and supply a proof or reference for the trace inequality. Without this, the covariance-term bound in Theorem 3.2 is unsupported.
  3. [Section 4.1, Eqs. (10) and (11)] The paper formulates an objective for constructing A' and X' by matching \tilde C^T H to the propagated synthetic node representations, but then, 'as a workaround', directly sets X' = \tilde C^T Z and A' = \tilde C^T \tilde A \tilde C. This does not optimize Eq. (10), and no residual or approximation error is reported. The reader is left with an asserted rather than demonstrated structural property of the constructed condensed graph. This is not fatal for the empirical method, but it should be stated plainly that Eq. (10) is motivational rather than actually optimized, or the construction should be validated numerically.
minor comments (4)
  1. [Appendix E] The headings 'Proof of Lemma 3.1' and 'Proof of Lemma 3.2' should read 'Proof of Theorem 3.1' and 'Proof of Theorem 3.2'; the numbering in the theorem statements is inconsistent with the proof labels.
  2. [Section 3.1, Eq. (1)] The FID formula writes (Σ_orgΣ_syn)^{1/2}, but for non-commuting covariance matrices the standard FID uses (Σ_org^{1/2}Σ_synΣ_org^{1/2})^{1/2}. Please correct the notation and define the square root carefully.
  3. [Section 6.1 and Table 6] Table 6 shows that ClustGDD has per-dataset and per-ratio hyper-parameters (T, alpha, E1, beta, rho, T', E3, gamma, lambda, dropout, H), and Section C.2 describes an extensive search. The authors should explicitly state how these values were selected (for example, using a validation split) so that the reader can judge whether the favorable test accuracy could result from tuning on the test set.
  4. [Section 6.2, Table 3] The text states that ClustGDD 'consistently achieve[s] superior or comparable performance', which is fair, but on several rows the differences against the runner-up are within one standard deviation (e.g., Cora 1.30%, Flickr 0.10%). The prose should avoid claiming a clear win on every dataset and instead emphasize the overall accuracy/efficiency trade-off.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the FID-to-clustering link is a surrogate-bound design choice, not an equivalence; the invalid Theorem 3.1 proof is a correctness risk, not a circularity.

full rationale

ClustGDD's central derivation is not circular. The paper motivates clustering by Theorem 3.2, which bounds the FID covariance term by the within-cluster sum of squares (WCSS), and then Eq. (6) minimizes exactly that WCSS. This is a standard surrogate-optimization design: the bound is not identical to FID, and the paper does not enforce balanced K-Means, so the first FID term is not forced to zero by construction. The downstream claim, GNN accuracy after training on the condensed graph, is tested against independent baselines on external benchmark datasets, giving the empirical contribution independent content. The self-citation [27] supplies an effective-resistance approximation in CAAR; it is a published, externally checkable lemma and the CAAR ablation is an empirical check, so it does not constitute load-bearing self-citation under the rubric. However, there is a non-circular correctness concern that should be flagged: Appendix E labels the proof as 'Proof of Lemma 3.1' for Theorem 3.1 and applies Jensen's Inequality to coefficients (1/|C_i| - n/N) that are not nonnegative convex weights and do not sum to 1, so the advertised upper bound is unsupported; this undermines the theoretical grounding but is a proof flaw, not a circularity.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The method's central claims rest on a set of domain assumptions about graph structure and about the validity of FID as a quality proxy, plus standard algorithmic primitives. These are not fully proven for the general case and are only validated on the five benchmark datasets.

free parameters (1)
  • Per-dataset and per-ratio hyperparameters (T, alpha, E1, beta, rho, T', E3, gamma, lambda, dropout, H) = Reported in Table 6 (e.g., Cora 1.3%: T=5, alpha=0.8, beta=0.01, rho=0.06, gamma=7.0)
    Each of the 15 dataset/ratio configurations uses a different hyperparameter vector, selected by an extensive search described in Section C.2. The reported accuracy depends on these fitted values, so the method's performance is not parameter-free.
assumptions (3)
  • domain assumption Adjacent nodes in the original graph tend to share the same class label (homophily assumption).
    Invoked in Section 4.1 to justify that GLS-smoothed, label-supervised embeddings will group same-class nodes together. The benchmark graphs have homophily ratios 0.66 to 0.81 (Table 1), so the assumption is only partially satisfied.
  • domain assumption FID between node embeddings is a precise quality metric for condensed graphs, correlated with downstream node classification accuracy.
    The paper shows empirical correlation in Fig. 1 but provides no theorem connecting FID bounds to classification accuracy; this is an unproven proxy.
  • domain assumption GNN message passing can be approximated by the closed-form graph Laplacian smoothing solution Z = sum_t (1-alpha) alpha^t A^t X.
    Taken from prior work [7, 37, 62], used to build embeddings without training a full GNN; the approximation quality on the five datasets is not separately quantified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Simple yet Effective Graph Distillation via Clustering." pith.science (2026). https://pith.science/paper/D46RB45B

@misc{pith2026250520807,
  author       = {Pith},
  title        = {Pith review of: Simple yet Effective Graph Distillation via Clustering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D46RB45B}},
  note         = {Machine review of arXiv:2505.20807}
}
read the original abstract

Despite plentiful successes achieved by graph representation learning in various domains, the training of graph neural networks (GNNs) still remains tenaciously challenging due to the tremendous computational overhead needed for sizable graphs in practice. Recently, graph data distillation (GDD), which seeks to distill large graphs into compact and informative ones, has emerged as a promising technique to enable efficient GNN training. However, most existing GDD works rely on heuristics that align model gradients or representation distributions on condensed and original graphs, leading to compromised result quality, expensive training for distilling large graphs, or both. Motivated by this, this paper presents an efficient and effective GDD approach, ClustGDD. Under the hood, ClustGDD resorts to synthesizing the condensed graph and node attributes through fast and theoretically-grounded clustering that minimizes the within-cluster sum of squares and maximizes the homophily on the original graph. The fundamental idea is inspired by our empirical and theoretical findings unveiling the connection between clustering and empirical condensation quality using Fr\'echet Inception Distance, a well-known quality metric for synthetic images. Furthermore, to mitigate the adverse effects caused by the homophily-based clustering, ClustGDD refines the nodal attributes of the condensed graph with a small augmentation learned via class-aware graph sampling and consistency loss. Our extensive experiments exhibit that GNNs trained over condensed graphs output by ClustGDD consistently achieve superior or comparable performance to state-of-the-art GDD methods in terms of node classification on five benchmark datasets, while being orders of magnitude faster.

Figures

Figures reproduced from arXiv: 2505.20807 by the authors.

Figure 1
Figure 1. Classification accuracy v.s. FID under various condensation ratios. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Example of a balanced clustering of G. When we let 𝑢𝑖 ’s representation be the averaged embeddings of the nodes inside cluster 𝐶𝑖 , i.e., 𝑯 ′ 𝑖 = Í 𝑣𝑗 ∈𝐶𝑖 𝑯 𝑗 |𝐶𝑖 | , our Theorem 3.1 reveals that the first term ∥𝝁 org −𝝁 syn∥ 2 2 in the FID can be bounded by 1 𝑁 2 Í𝑛 𝑖=1  𝑁 𝑛 − |𝐶𝑖 | 2 . Since Í𝑛 𝑖=1 |𝐶𝑖 | = 𝑁, the mean cluster size is thus 𝑁 𝑛 . This upper bound is essentially the variance of the sizes of 𝑛 clust… view at source ↗
Figure 4
Figure 4. Example of the heterophilic over-smoothing issue [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: Hyper-parameter Analysis. which is critical for controlling the degree of information integra￾tion between nodes at different distances in the graph. The model’s performance initially increases with 𝑇 and then decreases on both arXiv and Reddit. This means that aggrega…
Figure 6
Figure 6. Figure 6: Computational time comparison Linux machine equipped with an Intel(R) Xeon(R) Gold 6226 CPU @ 2.70GHz and a 32 GB Nvidia Tesla V100 GPU. We set learning rate of GNNs to 0.01, the weight decay is 5𝑒 − 4. For ease of training, 𝑾 and 𝑾′ are two three-layer linear layers, …
Figure 8
Figure 8. Figure 8: The visualization synthetic graphs E Theoretical Proofs Proof of Lemma 3.1. Using the Cauchy–Schwarz inequality, we can get ∥𝝁 org − 𝝁 syn ∥ 2 2 = [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

66 extracted references · 39 canonical work pages

  1. [1]

    Huzihiro Araki. 1990. On an inequality of Lieb and Thirring. Letters in Mathe- matical Physics 19, 2 (1990), 167–170

    work page 1990

  2. [2]

    Fedor Borisyuk, Shihai He, Yunbo Ouyang, Morteza Ramezani, Peng Du, Xiaochen Hou, Chengming Jiang, Nitin Pasumarthy, Priya Bannur, Birjodh Tiwana, et al

    work page

  3. [3]

    Yingmei Chen, Zhongyu Wei, and Xuanjing Huang. 2018. Incorporating cor- poration relationship via graph convolutional neural networks for stock price prediction. In CIKM. 1655–1658

    work page 2018

  4. [4]

    Michaël Defferrard, Xavier Bresson, and Pierre Vandergheynst. 2016. Convolu- tional neural networks on graphs with fast localized spectral filtering. Advances in neural information processing systems 29 (2016)

    2016

  5. [5]

    Austin Derrow-Pinion, Jennifer She, David Wong, Oliver Lange, Todd Hester, Luis Perez, Marc Nunkesser, Seongjae Lee, Xueying Guo, Brett Wiltshire, et al. 2021. Eta prediction with graph neural networks in google maps. In CIKM. 3767–3776

    work page 2021

  6. [6]

    Mucong Ding, Xiaoyu Liu, Tahseen Rabbani, and Furong Huang. 2022. Faster hyperparameter search on graphs via calibrated dataset condensation. InNeurIPS 2022 Workshop: New Frontiers in Graph Learning

    work page 2022

  7. [7]

    Xiaowen Dong, Dorina Thanou, Pascal Frossard, and Pierre Vandergheynst

    work page

  8. [8]

    DC Dowson and BV666017 Landau. 1982. The Fréchet distance between multivari- ate normal distributions. Journal of multivariate analysis 12, 3 (1982), 450–455

    work page 1982

Show all 66 references
  1. [9]

    Junfeng Fang, Xinglin Li, Yongduo Sui, Yuan Gao, Guibin Zhang, Kun Wang, Xiang Wang, and Xiangnan He. 2024. Exgc: Bridging efficiency and explainability in graph condensation. In Proceedings of the ACM on Web Conference 2024 . 721– 732

    2024

  2. [10]

    Wenzheng Feng, Jie Zhang, Yuxiao Dong, Yu Han, Huanbo Luan, Qian Xu, Qiang Yang, Evgeny Kharlamov, and Jie Tang. 2020. Graph random neural networks for semi-supervised learning on graphs. Advances in neural information processing systems 33 (2020), 22092–22103

    2020

  3. [11]

    Alex Fout, Jonathon Byrd, Basir Shariat, and Asa Ben-Hur. 2017. Protein inter- face prediction using graph convolutional networks. In Proceedings of the 31st International Conference on Neural Information Processing Systems . 6533–6542

    2017

  4. [12]

    Chen Gao, Yu Zheng, Nian Li, Yinfeng Li, Yingrong Qin, Jinghua Piao, Yuhan Quan, Jianxin Chang, Depeng Jin, Xiangnan He, et al. 2023. A survey of graph neural networks for recommender systems: Challenges, methods, and directions. ACM Transactions on Recommender Systems 1, 1 (2...

    2023

  5. [13]

    Xinyi Gao, Tong Chen, Yilong Zang, Wentao Zhang, Quoc Viet Hung Nguyen, Kai Zheng, and Hongzhi Yin. 2024. Graph condensation for inductive node repre- sentation learning. In 2024 IEEE 40th International Conference on Data Engineering (ICDE). IEEE, 3056–3069

    2024

  6. [14]

    Xinyi Gao, Tong Chen, Wentao Zhang, Junliang Yu, Guanhua Ye, Quoc Viet Hung Nguyen, and Hongzhi Yin. 2024. Rethinking and Accelerating Graph Con- densation: A Training-Free Approach with Class Partition. arXiv preprint arXiv:2405.13707 (2024)

    2024 arXiv

  7. [15]

    Xinyi Gao, Junliang Yu, Tong Chen, Guanhua Ye, Wentao Zhang, and Hongzhi Yin. 2025. Graph condensation: A survey. IEEE Transactions on Knowledge and Data Engineering (2025)

    2025

  8. [16]

    Johannes Gasteiger, Aleksandar Bojchevski, and Stephan Günnemann. 2018. Predict then Propagate: Graph Neural Networks meet Personalized PageRank. In International Conference on Learning Representations

    2018

  9. [17]

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

    2017

  10. [18]

    Will Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs. Advances in neural information processing systems 30 (2017)

    2017

  11. [19]

    Mohammad Hashemi, Shengbo Gong, Juntong Ni, Wenqi Fan, B Aditya Prakash, and Wei Jin. 2024. A comprehensive survey on graph reduction: Sparsification, coarsening, and condensation. arXiv preprint arXiv:2402.03358 (2024)

    2024 arXiv

  12. [20]

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. 2017. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)

    2017

  13. [21]

    Weihua Hu, Matthias Fey, Marinka Zitnik, Yuxiao Dong, Hongyu Ren, Bowen Liu, Michele Catasta, and Jure Leskovec. 2020. Open graph benchmark: Datasets for machine learning on graphs. Advances in neural information processing systems 33 (2020), 22118–22133

    2020

  14. [22]

    Dejun Jiang, Zhenxing Wu, Chang-Yu Hsieh, Guangyong Chen, Ben Liao, Zhe Wang, Chao Shen, Dongsheng Cao, Jian Wu, and Tingjun Hou. 2021. Could graph neural networks learn better molecular representation for drug discovery? A comparison study of descriptor-based and graph-based ...

    2021

  15. [23]

    Weiwei Jiang and Jiayun Luo. 2022. Graph neural network for traffic forecasting: A survey. ESA 207 (2022), 117921

    2022

  16. [24]

    Wei Jin, Xianfeng Tang, Haoming Jiang, Zheng Li, Danqing Zhang, Jiliang Tang, and Bing Yin. 2022. Condensing graphs via one-step gradient matching. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 720–730

    2022

  17. [25]

    Wei Jin, Lingxiao Zhao, Shichang Zhang, Yozen Liu, Jiliang Tang, and Neil Shah. 2021. Graph condensation for graph neural networks. arXiv preprint arXiv:2110.07580 (2021)

    2021 arXiv

  18. [26]

    Thomas N Kipf and Max Welling. 2016. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907 (2016)

    2016 arXiv

  19. [27]

    Yurui Lai, Xiaoyang Lin, Renchi Yang, and Hongtao Wang. 2024. Efficient topology-aware data augmentation for high-degree graph neural networks. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 1463–1473

    2024

  20. [28]

    Shiye Lei and Dacheng Tao. 2023. A comprehensive survey of dataset distillation. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)

    2023

  21. [29]

    Fan Li, Xiaoyang Wang, Dawei Cheng, Wenjie Zhang, Ying Zhang, and Xuemin Lin. 2024. TCGU: Data-centric Graph Unlearning based on Transferable Conden- sation. arXiv preprint arXiv:2410.06480 (2024)

    2024 arXiv

  22. [30]

    Yongqi Li and Wenjie Li. 2021. Data distillation for text classification. arXiv preprint arXiv:2104.08448 (2021)

    2021 arXiv

  23. [31]

    Mengyang Liu, Shanchuan Li, Xinshi Chen, and Le Song. 2022. Graph conden- sation via receptive field distribution matching. arXiv preprint arXiv:2206.13697 (2022)

    2022 arXiv

  24. [32]

    Yang Liu, Deyu Bo, and Chuan Shi. 2024. Graph Distillation with Eigenbasis Matching. In Forty-first International Conference on Machine Learning

    2024

  25. [33]

    Yilun Liu, Ruihong Qiu, and Zi Huang. 2023. Cat: Balanced continual graph learning with graph condensation. In 2023 IEEE International Conference on Data Mining (ICDM). IEEE, 1157–1162

    2023

  26. [34]

    Zhanyu Liu, Chaolv Zeng, and Guanjie Zheng. 2024. Graph data condensation via self-expressive graph structure reconstruction. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 1992–2002

    2024

  27. [35]

    Stuart Lloyd. 1982. Least squares quantization in PCM. IEEE transactions on information theory 28, 2 (1982), 129–137

    1982

  28. [36]

    László Lovász. 1993. Random walks on graphs. Combinatorics, Paul erdos is eighty 2, 1-46 (1993), 4

    1993

  29. [37]

    Yao Ma, Xiaorui Liu, Tong Zhao, Yozen Liu, Jiliang Tang, and Neil Shah. 2021. A unified view on graph neural networks as graph signal denoising. InProceedings of the 30th ACM International Conference on Information & Knowledge Management . 1202–1211

    2021

  30. [38]

    Ilija Radosavovic, Piotr Dollár, Ross Girshick, Georgia Gkioxari, and Kaiming He

  31. [39]

    Noveen Sachdeva and Julian McAuley. 2023. Data distillation: A survey. arXiv preprint arXiv:2301.04272 (2023)

    2023 arXiv

  32. [40]

    Prithviraj Sen, Galileo Namata, Mustafa Bilgic, Lise Getoor, Brian Galligher, and Tina Eliassi-Rad. 2008. Collective classification in network data. AI magazine 29, 3 (2008), 93–93

    2008

  33. [41]

    Ozan Sener and Silvio Savarese. 2017. Active learning for convolutional neural networks: A core-set approach. arXiv preprint arXiv:1708.00489 (2017)

    2017 arXiv

  34. [42]

    Jonathan M Stokes, Kevin Yang, Kyle Swanson, Wengong Jin, Andres Cubillos- Ruiz, Nina M Donghia, Craig R MacNair, Shawn French, Lindsey A Carfrae, Zohar Bloom-Ackermann, et al. 2020. A deep learning approach to antibiotic discovery. Cell 180, 4 (2020), 688–702

    2020

  35. [43]

    Yefan Tao, Luyang Kong, Andrey Kan, and Laurent Callot. 2024. Textual Dataset Distillation via Language Model Embedding. In Findings of the Association for Computational Linguistics: EMNLP 2024 . 12557–12569

    2024

  36. [44]

    Lin Wang, Wenqi Fan, Jiatong Li, Yao Ma, and Qing Li. 2024. Fast graph conden- sation with structure-based neural tangent kernel. In Proceedings of the ACM on Web Conference 2024. 4439–4448

    2024

  37. [45]

    Lin Wang and Qing Li. 2025. Efficient Graph Condensation via Gaussian Process. arXiv preprint arXiv:2501.02565 (2025)

    2025

  38. [46]

    Yuyang Wang, Jianren Wang, Zhonglin Cao, and Amir Barati Farimani. 2022. Molecular contrastive learning of representations via graph neural networks. Nature Machine Intelligence 4, 3 (2022), 279–287

    2022

  39. [47]

    Max Welling. 2009. Herding dynamical weights to learn. In Proceedings of the 26th annual international conference on machine learning . 1121–1128

    2009

  40. [48]

    Lirong Wu, Haitao Lin, Zhangyang Gao, Guojiang Zhao, and Stan Z. Li. 2024. A Teacher-Free Graph Knowledge Distillation Framework With Dual Self- Distillation. IEEE Transactions on Knowledge and Data Engineering 36, 9 (2024), 4375–4385. https://doi.org/10.1109/TKDE.2024.3374773

    2024

  41. [49]

    Shiwen Wu, Fei Sun, Wentao Zhang, Xu Xie, and Bin Cui. 2022. Graph neural networks in recommender systems: a survey. Comput. Surveys 55, 5 (2022), 1–37

    2022

  42. [50]

    Zhenbang Xiao, Yu Wang, Shunyu Liu, Huiqiong Wang, Mingli Song, and Tongya Zheng. 2024. Simple graph condensation. InJoint European Conference on Machine Learning and Knowledge Discovery in Databases . Springer, 53–71

    2024

  43. [51]

    Hongjia Xu, Liangliang Zhang, Yao Ma, Sheng Zhou, Zhuonan Zheng, and Bu Jiajun. 2024. A survey on graph condensation. arXiv preprint arXiv:2402.02000 (2024). 10 Simple yet Effective Graph Distillation via Clustering KDD’25, August 3–7, 2025, Toronto,ON, Canada

    2024 arXiv

  44. [52]

    Keyulu Xu, Chengtao Li, Yonglong Tian, Tomohiro Sonobe, Ken-ichi Kawarabayashi, and Stefanie Jegelka. 2018. Representation learning on graphs with jumping knowledge networks. In International conference on machine learn- ing. PMLR, 5453–5462

    2018

  45. [53]

    Zhe Xu, Yuzhong Chen, Menghai Pan, Huiyuan Chen, Mahashweta Das, Hao Yang, and Hanghang Tong. 2023. Kernel ridge regression-based graph dataset distillation. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 2850–2861

    2023

  46. [54]

    Beining Yang, Kai Wang, Qingyun Sun, Cheng Ji, Xingcheng Fu, Hao Tang, Yang You, and Jianxin Li. 2023. Does graph distillation see like vision dataset counterpart? Advances in Neural Information Processing Systems 36 (2023), 53201– 53226

    2023

  47. [55]

    Zeyuan Yin and Zhiqiang Shen. 2024. Dataset distillation via curriculum data synthesis in large data era. Transactions on Machine Learning Research (2024)

    2024

  48. [56]

    Zeyuan Yin, Eric Xing, and Zhiqiang Shen. 2024. Squeeze, recover and relabel: Dataset condensation at imagenet scale from a new perspective. Advances in Neural Information Processing Systems 36 (2024)

    2024

  49. [57]

    Hanqing Zeng, Hongkuan Zhou, Ajitesh Srivastava, Rajgopal Kannan, and Viktor Prasanna. 2019. Graphsaint: Graph sampling based inductive learning method. arXiv preprint arXiv:1907.04931 (2019)

    2019 arXiv

  50. [58]

    Ge Zhang, Zhao Li, Jiaming Huang, Jia Wu, Chuan Zhou, Jian Yang, and Jianliang Gao. 2022. efraudcom: An e-commerce fraud detection system via competitive graph neural networks. TOIS 40, 3 (2022), 1–29

    2022

  51. [59]

    Xin Zheng, Miao Zhang, Chunyang Chen, Quoc Viet Hung Nguyen, Xingquan Zhu, and Shirui Pan. 2024. Structure-free graph condensation: From large-scale graphs to condensed graph-free data. Advances in Neural Information Processing Systems 36 (2024)

    2024

  52. [60]

    Ziang Zhou, Jieming Shi, Renchi Yang, Yuanhang Zou, and Qing Li. 2023. SlotGAT: slot-based message passing for heterogeneous graphs. InInternational Conference on Machine Learning. PMLR, 42644–42657

    2023

  53. [61]

    Jiong Zhu, Yujun Yan, Lingxiao Zhao, Mark Heimann, Leman Akoglu, and Danai Koutra. 2020. Beyond homophily in graph neural networks: Current limitations and effective designs. Advances in neural information processing systems 33 (2020), 7793–7804

    2020

  54. [62]

    Meiqi Zhu, Xiao Wang, Chuan Shi, Houye Ji, and Peng Cui. 2021. Interpreting and Unifying Graph Neural Networks with An Optimization Framework. Proceedings of the Web Conference 2021 (2021). 11 KDD’25, August 3–7, 2025, Toronto,ON, Canada Yurui Lai, et al. A A Detailed Discussi...

    2021

  55. [66]

    Plugging the above inequality into Eq

    (21) From Lemma E.2, we can further derive that Tr(𝚺syn)≤ 1 𝑛 𝑛∑︁ 𝑖=1 ∥𝑯′ 𝑖− 𝝁org∥2 2≤ 1 𝑛· 𝑁 𝑐min 𝑛∑︁ 𝑖=1 |𝐶𝑖| 𝑁 ∥𝑯′ 𝑖− 𝝁org∥2 2 ≤ 𝑁 𝑛𝑐min · Tr(𝚺org). Plugging the above inequality into Eq. (21) finishes the proof. □ Proof of Lemma E.1. Recall that 𝚺org 𝑎,𝑏 represents the cov...

    2025

  56. [2016]

    IEEE Transactions on Signal Processing 64, 23 (2016), 6160–6173

    Learning Laplacian matrix in smooth graph signal representations. IEEE Transactions on Signal Processing 64, 23 (2016), 6160–6173

    2016

  57. [2018]

    In Proceedings of the IEEE conference on computer vision and pattern recognition

    Data distillation: Towards omni-supervised learning. In Proceedings of the IEEE conference on computer vision and pattern recognition . 4119–4128

  58. [2024]

    In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining

    Lignn: Graph neural networks at linkedin. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 4793–4803

Pith tools

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