Pith. sign in

REVIEW 5 major objections 6 minor 52 references

Simplifying Graph Kernels for Efficient

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Replacing GNTK's layer-by-layer recursion with K-step aggregation plus a single kernel update gives near-GNTK accuracy at up to about 50x lower kernel cost.

desk verdict Useful simplification of GNTK with real speedups, but the 'no loss of expressive power' claim is unsupported and contradicted by its own MUTAG result. read the letter →

arxiv 2507.03560 v2 pith:TKVPEC5A submitted 2025-07-04 cs.LG

Lin Wang , Shijie Wang , Sirui Huang , Qing Li This is my paper
classification cs.LG
keywords GaussianProcessKernelGraphNeuralTangentSimpleConvolutionefficiencymessagepassingnodeclassification
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

The paper tries to show that the Graph Neural Tangent Kernel (GNTK), an infinite-width GNN kernel that matches GNN accuracy, can be drastically simplified without losing that accuracy on typical benchmarks. Instead of recursively interleaving message passing with neural-tangent-kernel updates layer by layer, the proposed SGTK performs K rounds of feature aggregation first and then a single kernel update, while the proposed SGNK replaces that update with a closed-form Gaussian-process arcsin kernel. On standard node and graph classification datasets the simplified kernels match or beat GNTK accuracy while cutting kernel-computation time by up to a factor of about 50. The central bet is that in deep GNTK layers the nonlinear transformations contribute little, so only the aggregation pattern matters.

What carries the argument

The central object is the K-step aggregated feature matrix $\hat{X} = \hat{A}^K X$, obtained by applying the normalized adjacency (with self-loops) K times, followed by a single infinite-width kernel update instead of K alternating aggregation/NTK recursions. For SGTK the update is one NTK recursion $\Theta^{(1)}_{ij} = \hat{\Sigma}^{(0)}_{ij} \dot{\Sigma}^{(1)}_{ij} + \hat{\Sigma}^{(1)}_{ij}$ evaluated on the aggregated covariance, and for SGNK the update is the closed-form arcsin kernel $\frac{2}{\pi}\sin^{-1}\left(\frac{2\tilde{x}_i^\top \Sigma_w \tilde{x}_j}{\sqrt{(1+2\tilde{x}_i^\top \Sigma_w \tilde{x}_i)(1+2\tilde{x}_j^\top \Sigma_w \tilde{x}_j)}}\right)$ from the infinite-width GP with error-function activation. This single-update structure is what removes the layer-by-layer kernel iterations and cuts the K-dependence in the cost.

What would settle it

Run SGTK and SGNK against the true K-layer GNTK on a benchmark where removing nonlinearities (SGC versus GCN) causes a clear accuracy drop; if the simplified kernels follow SGC's drop rather than GNTK's accuracy, the claim that aggregation alone preserves GNTK's expressive power is refuted.

Watch

Extended reading notes

Core claim

The paper's central claim is that a K-layer infinite-width GNN's kernel behavior, as captured by the GNTK, can be reproduced by concentrating the K rounds of propagation into one feature aggregation step followed by a single kernel evaluation. Concretely, SGTK computes $\hat{X} = \hat{A}^K X$ and then applies one NTK covariance update, while SGNK replaces that update with the closed-form GP arcsin kernel of Eq. (21). The authors report that on five node-classification and four graph-classification benchmarks these simplified kernels match or exceed GNTK accuracy in most cases, and that SGNK builds the kernel matrix on the Computers dataset in 0.015 s versus GNTK's 0.827 s. They also give a complexity analysis showing the K-dependent NTK term shrinks from GNTK's $O(K(n^2 d + nm d + m^2 d + n k^2))$ to a single occurrence for SGTK and to $O(K n k^2 + nm d)$ for SGNK.

Load-bearing premise

The load-bearing premise is that the nonlinear transformations inside deep GNTK layers are negligible, so K rounds of linear message aggregation followed by a single kernel update can stand in for a K-layer infinite-width GNN; this is adopted from the SGC simplification heuristic rather than derived.

Editorial extensions

If this is right

  • Kernel-matrix construction time becomes essentially flat in K, so expanding a node's receptive field no longer inflates the computation that dominated GNTK's cost.
  • On the five node datasets, SGTK and SGNK reach or exceed GNTK accuracy while cutting kernel time by at least 3x and up to about 54x, so the efficiency gain comes without a measured accuracy loss on those benchmarks.
  • SGNK's closed-form arcsin kernel eliminates the $O(n^2 d)$ term from the kernel update, reducing the node-count-dependent complexity from $O(n^2 d + nm d + m^2 d)$ to $O(nm d)$, which is the difference that scales to larger graphs.
  • Both kernels inherit the training-free, non-parametric character of graph kernels, so they remain applicable when GNN training is too expensive or data are scarce.
  • The complexity analysis and the K-sensitivity plots together imply that for these benchmarks the number of GNTK layers contributes cost without contributing accuracy, so the simplification captures the useful part of depth.

Reading between the lines

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

  • The MUTAG graph-classification drop (90.0 for GNTK to 85.7/86.7 for SGTK/SGNK) suggests the simplifying assumption is not uniformly safe; a practical recipe would be to screen with the fast kernel and fall back to GNTK where small accuracy margins matter.
  • If the SGC-style premise holds beyond these benchmarks, the same 'aggregate first, kernel once' recipe should extend to other structural diffusions (personalized PageRank, heat kernels) and other activation GP kernels, yielding a family of cheap infinite-width graph kernels.
  • A direct ablation of GNTK—removing intermediate NTK updates one at a time and measuring accuracy—would quantify each layer's contribution and test the paper's core assumption more directly than comparing downstream accuracies.
  • Because SGNK's kernel is a closed form on aggregated features, it implies that the infinite-width GNN limit is separable: structural propagation and feature similarity are computed in two independent stages, which may simplify theoretical analysis of GNN generalization.
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

5 major / 6 minor

Summary. The paper proposes two simplified graph kernel methods, SGTK and SGNK, that replace the deep layer-wise NTK recursion of GNTK with K steps of linear message aggregation followed by a single kernel update (SGTK) or a direct Gaussian-process/arc-sine kernel evaluation (SGNK). The authors report competitive or superior accuracy on five node-classification and four graph-classification benchmarks, with substantially reduced kernel-matrix computation times (up to about 50x faster on Computers). They also give a complexity analysis and provide code.

Significance. If the central claim were established, the paper would offer a practical speed-accuracy tradeoff for kernel-based graph learning: simplified kernels that are dramatically faster than GNTK while remaining competitive on several benchmarks. The empirical efficiency results are the paper's strongest contribution, and the provision of code and reproducibility materials is a clear plus. However, the paper's headline claim of 'without sacrificing expressive power' is not backed by a derivation, a bound, or a controlled ablation, and is contradicted by the paper's own MUTAG results. The theoretical derivations also contain internal inconsistencies in the kernel formulas and in the complexity analysis. The value of the paper lies mainly in an empirical observation, not in the claimed theoretical simplification.

major comments (5)
  1. [Abstract and Sections 4.3-4.4] The central claim that replacing GNTK's recursive layer-wise NTK iterations with K-step linear aggregation followed by a single kernel update is achieved 'without sacrificing expressive power' is unsupported and is contradicted by the paper's own Table 5. On MUTAG, GNTK achieves 90.0%, SGTK 85.7%, and SGNK 86.7%. The only justification offered is the SGC heuristic from [42], which concerns finite-width weight collapsing and does not directly apply to the GNTK covariance recursion. No bound, no ablation against a true K-layer GNTK with the same K, and no analysis of the error introduced by omitting intermediate nonlinear covariance updates is provided. Please either provide such an analysis or substantially weaken the claim to 'competitive accuracy' in the abstract, introduction, and conclusion.
  2. [Section 4.4, Eq. (19) and Eq. (21)] There is an internal inconsistency between the two displayed SGNK formulas. Equation (19) contains the matrix Σ_w = diag(1, ..., 1, σ_b^2) in both the numerator and denominator, while Eq. (21), which is the formula used in Algorithm 2 and in the experiments, drops Σ_w entirely and uses 2 \tilde{X}_1 \tilde{X}_2^T. These are equivalent only if σ_b^2 = 1 and all feature dimensions are normalized accordingly. But Eq. (10) and Eq. (16) introduce a bias hyperparameter β, whose square should appear as the bias variance. As written, β is not present in the SGNK kernel at all, so the method's bias dependence is not consistent with the derivation. Clarify the relationship among β, σ_b^2, and the augmented feature vector, and specify whether β is tuned or fixed.
  3. [Section 4.3, Eq. (12)-(15)] The statement 'Suppose the enhanced node features ... are independently and identically distributed' is unjustified and, for graph-structured data after K-step message passing, generally false. The NTK/GP derivation for infinite-width networks does not require the inputs to be random; inputs are fixed features, while randomness resides in the network weights. The i.i.d. assumption appears to be an artifact of conflating input randomness with weight randomness. Please remove this assumption and state the kernel derivation for deterministic aggregated features, or justify why the i.i.d. input model is valid for the graph setting.
  4. [Section 4.5, Computational Complexity Analysis] The complexity formulas are incorrect as written. In GNTK, the message-passing step on the covariance matrix between two graphs with n and m nodes and average degree k costs O(n k m), not O(n k^2); the total per-layer cost should involve n m k. For SGNK, propagating the node feature matrix through K sparse aggregation steps costs O(K n k d), where d is the feature dimension, not O(K n k^2). The stated totals O(K(nk^2 + n^2 d + m^2 d + nmd)) and O(Knk^2 + nmd) therefore misstate the complexity of the very operations the paper claims to simplify. Please correct these formulas; the empirical timing figures can still support the qualitative efficiency conclusion.
  5. [Section 5.1 and Tables 3-5] Hyperparameter K is tuned separately for each dataset and method, but the selected K values are not reported, and the node-classification kernel results in Table 3 are reported without standard deviations over different splits. This makes it difficult to assess whether the accuracy differences between SGTK/SGNK and GNTK are statistically meaningful and to reproduce the exact results. Please report the chosen K for each method/dataset and provide repeated-split results or confidence intervals for the kernel methods, at least for the main comparisons.
minor comments (6)
  1. [Title] The title 'Simplifying Graph Kernels for Efficient' is grammatically incomplete; consider 'Simplifying Graph Kernels for Efficient Graph Learning' or a similar complete phrase.
  2. [Typesetting] The manuscript still contains ACM template placeholders such as 'Conference acronym ’XX, June 03–05, 2018' and 'Make sure to enter the correct conference title'.
  3. [Section 4.2, Eq. (8)] The displayed GNTK recursion is a simplified schematic; the actual GNTK involves a more involved recursion with separate covariance and NTK updates per layer. Consider either expanding the notation or explicitly referring to [5] for the full recursion.
  4. [Table 5] The heading 'standard derivation' should be 'standard deviation'.
  5. [Section 5.5] The text says 'five graph-level datasets' but Table 1 lists only four graph datasets; this is likely a typo.
  6. [References] References [15] and [16] appear to be the same paper (Kriege and Mutzel, 2012) and should be merged or differentiated.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: SGTK/SGNK apply standard NTK and NNGP formulas to K-step aggregated features; no fitted parameter is renamed as prediction, and the self-citation to SNTK is not load-bearing.

full rationale

The derivation chain is self-contained. SGTK is defined by applying the standard infinite-width two-layer NTK recursion (Eqs. 12-15) to features \hat X = \hat A^K X (Eq. 9); this is the textbook NTK form for a two-layer ReLU network, with graph structure entering only through the \hat A^K preprocessing. SGNK is the closed-form arc-sine / NNGP kernel of Williams (1996) (Eqs. 18-21) evaluated on the same aggregated features. Neither kernel contains parameters fitted to benchmark labels; the only hyperparameters (K, SVM C, KRR \lambda) are tuned and reported, and the accuracies are direct evaluations of fixed kernel matrices, not predictions derived from fitted values. The comparison to GNTK is external: the claimed speedups are measured wall-clock times on the same datasets. The only self-citation (SNTK, ref [35]) is used as a baseline and in related work, not as justification for the kernel formulas or for the simplification step. The load-bearing heuristic that deep nonlinear layers contribute diminishing signal is borrowed from SGC (ref [42]); whether that heuristic is valid is a correctness or evidence question, not a circularity one, because the SGTK/SGNK outputs are not defined in terms of GNTK or of the accuracies they are said to predict. The MUTAG gap (GNTK 90.0 vs SGTK 85.7 and SGNK 86.7) undermines the abstract's 'without sacrificing expressive power' claim but does not make the derivation circular.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central methods are composed of existing building blocks (SGC preprocessing, NTK, NNGP/arc-sine kernel), so no new theoretical entities are invented. The main uncharged axioms are the NTK/NNGP limit, the arc-sine formula, and two ad hoc assumptions: the SGC simplification heuristic and the i.i.d. feature assumption.

free parameters (3)
  • K (message-passing steps) = tuned per dataset over {1,2,3,4,5}
    The number of aggregation steps strongly affects accuracy; the paper tunes K on each dataset and reports the best accuracy (Section 5.1, 5.6), which risks overfitting without a held-out validation split.
  • beta (bias scaling) = not reported; implicitly 1 in Eq (21)/Algorithm 2
    beta appears in the definition of the SGTK covariance (Eq 12) and in the SGNK augmented covariance via Sigma_w, but Eq (21) drops it, so either beta=1 is assumed or the implementation omits the parameter.
  • SVM C / KRR lambda = C in [1e-2,1e4], lambda in [1e-2,1e2]
    Classifier hyperparameters are tuned per dataset; standard practice but part of the reported best accuracy.
assumptions (4)
  • domain assumption NTK and NNGP limits of infinite-width neural networks
    The paper relies on Jacot et al. [11] and Lee et al. [19] for the equivalence between infinite-width networks and deterministic kernels; this is standard theory, not derived in the paper.
  • standard math Closed-form arc-sine kernel for erf activation (Williams 1996)
    Eq (19) uses the analytic expectation E[erf(f_i)erf(f_j)] = (2/pi)arcsin(...); a known identity, treated as a black box.
  • ad hoc to paper The 'deep layers contribute little' heuristic from SGC [42]
    The justification for replacing GNTK's stacked NTK iterations with a single kernel update after K-step aggregation rests on an empirical observation about SGC, not on a bound or proof for kernels (Sections 4.2-4.3).
  • ad hoc to paper i.i.d. enhanced node features
    Section 4.3 states 'Suppose the enhanced node features {x_hat_i} are independently and identically distributed', which is false for features produced by message passing on a graph; this assumption is used to justify the kernel form.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Simplifying Graph Kernels for Efficient." pith.science (2026). https://pith.science/paper/TKVPEC5A

@misc{pith2026250703560,
  author       = {Pith},
  title        = {Pith review of: Simplifying Graph Kernels for Efficient},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TKVPEC5A}},
  note         = {Machine review of arXiv:2507.03560}
}
abstract

While kernel methods and Graph Neural Networks offer complementary strengths, integrating the two has posed challenges in efficiency and scalability. The Graph Neural Tangent Kernel provides a theoretical bridge by interpreting GNNs through the lens of neural tangent kernels. However, its reliance on deep, stacked layers introduces repeated computations that hinder performance. In this work, we introduce a new perspective by designing the simplified graph kernel, which replaces deep layer stacking with a streamlined $K$-step message aggregation process. This formulation avoids iterative layer-wise propagation altogether, leading to a more concise and computationally efficient framework without sacrificing the expressive power needed for graph tasks. Beyond this simplification, we propose another Simplified Graph Kernel, which draws from Gaussian Process theory to model infinite-width GNNs. Rather than simulating network depth, this kernel analytically computes kernel values based on the statistical behavior of nonlinear activations in the infinite limit. This eliminates the need for explicit architecture simulation, further reducing complexity. Our experiments on standard graph and node classification benchmarks show that our methods achieve competitive accuracy while reducing runtime. This makes them practical alternatives for learning on graphs at scale. Full implementation and reproducibility materials are provided at: https://anonymous.4open.science/r/SGNK-1CE4/.

Figures

Figures reproduced from arXiv: 2507.03560 by the authors.

Figure 1
Figure 1. The layer stacking structure. In spite of its demonstrated effectiveness, the GNTK exhibits certain inherent limitations. Specifically, the GNTK employs a layer￾stacking strategy to compute the kernel value, as illustrated in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. This kernel structure includes SGTK and SGNK. The SGTK improves computational efficiency by using continuous [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. node classification time consumption and accuracy [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: The changes of node classification time consuming with respect to [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Graph classification time consuming comparison. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Variations in graph classification time consumption [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: The changes of graph classification accuracy with [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 33 canonical work pages

  1. [42]

    Felix Wu, Amauri Souza, Tianyi Zhang, Christopher Fifty, Tao Yu, and Kilian Weinberger. 2019. Simplifying graph convolutional networks. In International conference on machine learning . PMLR, 6861–6871

    2019

  2. [1]

    Sanjeev Arora, Simon S Du, Wei Hu, Zhiyuan Li, Russ R Salakhutdinov, and Ruosong Wang. 2019. On exact computation with an infinitely wide neural net. Advances in neural information processing systems 32 (2019)

    work page 2019

  3. [2]

    Karsten M Borgwardt and Hans-Peter Kriegel. 2005. Shortest-path kernels on graphs. In Fifth IEEE international conference on data mining (ICDM’05) . IEEE, 8–pp

    work page 2005

  4. [3]

    Junhan Chen, Yuan Wang, et al . 2021. Social media use for health purposes: systematic review. Journal of medical Internet research 23, 5 (2021), e17917

    work page 2021

  5. [4]

    Mi Diao, Hui Kong, and Jinhua Zhao. 2021. Impacts of transportation network companies on urban mobility. Nature Sustainability 4, 6 (2021), 494–500

    work page 2021

  6. [5]

    Simon S Du, Kangcheng Hou, Russ R Salakhutdinov, Barnabas Poczos, Ruosong Wang, and Keyulu Xu. 2019. Graph neural tangent kernel: Fusing graph neural networks with graph kernels. Advances in neural information processing systems 32 (2019)

    work page 2019

  7. [6]

    Johannes Gasteiger, Aleksandar Bojchevski, and Stephan Günnemann. 2018. Predict then propagate: Graph neural networks meet personalized pagerank. arXiv preprint arXiv:1810.05997 (2018). Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Trovato et al

    arXiv 2018

  8. [7]

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

    2017

Show all 52 references
  1. [8]

    Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. 2020. Lightgcn: Simplifying and powering graph convolution network for recommendation. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval ...

    2020

  2. [9]

    Jilin Hu, Jianbing Shen, Bin Yang, and Ling Shao. 2020. Infinitely wide graph convolutional networks: semi-supervised learning via Gaussian processes. arXiv preprint arXiv:2002.12168 (2020)

    2020 arXiv

  3. [10]

    Wei Hu, Zhiyuan Li, and Dingli Yu. 2019. Simple and effective regularization methods for training on noisily labeled data with generalization guarantee.arXiv preprint arXiv:1905.11368 (2019)

    2019 arXiv

  4. [11]

    Arthur Jacot, Franck Gabriel, and Clément Hongler. 2018. Neural tangent ker- nel: Convergence and generalization in neural networks. Advances in neural information processing systems 31 (2018)

    2018

  5. [12]

    Shunhua Jiang, Yunze Man, Zhao Song, Zheng Yu, and Danyang Zhuo. 2022. Fast Graph Neural Tangent Kernel via Kronecker Sketching. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 36. 7033–7041

    2022

  6. [13]

    Chuanze Kang, Han Zhang, Zhuo Liu, Shenwei Huang, and Yanbin Yin. 2022. LR-GNN: A graph neural network based on link representation for predicting molecular associations. Briefings in Bioinformatics 23, 1 (2022), bbab513

    2022

  7. [14]

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

    2016 arXiv

  8. [15]

    Nils Kriege and Petra Mutzel. 2012. Subgraph matching kernels for attributed graphs. In Proceedings of the 29th International Coference on International Confer- ence on Machine Learning (ICML’12) . 291–298

    2012

  9. [16]

    Nils Kriege and Petra Mutzel. 2012. Subgraph matching kernels for attributed graphs. arXiv preprint arXiv:1206.6483 (2012)

    2012 arXiv

  10. [17]

    Sanjukta Krishnagopal and Luana Ruiz. 2023. Graph neural tangent kernel: Convergence on large graphs. In International Conference on Machine Learning . PMLR, 17827–17841

    2023

  11. [18]

    Yann LeCun, Léon Bottou, Genevieve B Orr, and Klaus-Robert Müller. 2002. Efficient backprop. In Neural networks: Tricks of the trade . Springer, 9–50

    2002

  12. [19]

    Jaehoon Lee, Yasaman Bahri, Roman Novak, Samuel S Schoenholz, Jeffrey Pen- nington, and Jascha Sohl-Dickstein. 2017. Deep neural networks as gaussian processes. arXiv preprint arXiv:1711.00165 (2017)

    2017 arXiv

  13. [20]

    Jaehoon Lee, Lechao Xiao, Samuel Schoenholz, Yasaman Bahri, Roman Novak, Jascha Sohl-Dickstein, and Jeffrey Pennington. 2019. Wide neural networks of any depth evolve as linear models under gradient descent. Advances in neural information processing systems 32 (2019)

    2019

  14. [21]

    Radford M Neal. 2012. Bayesian learning for neural networks . Vol. 118. Springer Science & Business Media

    2012

  15. [22]

    Timothy Nguyen, Zhourong Chen, and Jaehoon Lee. 2020. Dataset meta-learning from kernel ridge-regression. arXiv preprint arXiv:2011.00050 (2020)

    2020 arXiv

  16. [23]

    Giannis Nikolentzos, Giannis Siglidis, and Michalis Vazirgiannis. 2021. Graph kernels: A survey. Journal of Artificial Intelligence Research 72 (2021), 943–1027

    2021

  17. [24]

    Yu Rong, Yatao Bian, Tingyang Xu, Weiyang Xie, Ying Wei, Wenbing Huang, and Junzhou Huang. 2020. Self-supervised graph transformer on large-scale molecular data. Advances in neural information processing systems 33 (2020), 12559–12571

    2020

  18. [25]

    Michael Schlichtkrull, Thomas N Kipf, Peter Bloem, Rianne Van Den Berg, Ivan Titov, and Max Welling. 2018. Modeling relational data with graph convolu- tional networks. In The semantic web: 15th international conference, ESWC 2018, Heraklion, Crete, Greece, June 3–7, 2018, pro...

    2018

  19. [26]

    Oleksandr Shchur, Maximilian Mumme, Aleksandar Bojchevski, and Stephan Günnemann. 2018. Pitfalls of graph neural network evaluation. arXiv preprint arXiv:1811.05868 (2018)

    2018 arXiv

  20. [27]

    Nino Shervashidze, Pascal Schweitzer, Erik Jan Van Leeuwen, Kurt Mehlhorn, and Karsten M Borgwardt. 2011. Weisfeiler-lehman graph kernels. Journal of Machine Learning Research 12, 9 (2011)

    2011

  21. [28]

    Yehui Tang and Junchi Yan. 2022. GraphQNTK: Quantum Neural Tangent Kernel for Graph Data. Advances in Neural Information Processing Systems 35 (2022), 6104–6118

    2022

  22. [29]

    Shu-Feng Tsao, Helen Chen, Therese Tisseverasinghe, Yang Yang, Lianghua Li, and Zahid A Butt. 2021. What social media told us in the time of COVID-19: a scoping review. The Lancet Digital Health 3, 3 (2021), e175–e194

    2021

  23. [30]

    Petar Veličković. 2023. Everything is connected: Graph neural networks. Current Opinion in Structural Biology 79 (2023), 102538

    2023

  24. [31]

    Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, Yoshua Bengio, et al. 2017. Graph attention networks. stat 1050, 20 (2017), 10–48550

    2017

  25. [32]

    S Vichy N Vishwanathan, Nicol N Schraudolph, Risi Kondor, and Karsten M Borgwardt. 2010. Graph kernels. The Journal of Machine Learning Research 11 (2010), 1201–1242

    2010

  26. [33]

    Vladimir Vovk. 2013. Kernel ridge regression. In Empirical inference: Festschrift in honor of vladimir n. vapnik . Springer, 105–116

    2013

  27. [34]

    Lilapati Waikhom and Ripon Patgiri. 2023. A survey of graph neural networks in various learning paradigms: methods, applications, and challenges. Artificial Intelligence Review 56, 7 (2023), 6295–6364

    2023

  28. [35]

    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

  29. [36]

    Shijie Wang, Wenqi Fan, Xiao-Yong Wei, Xiaowei Mei, Shanru Lin, and Qing Li. 2024. Multi-agent attacks for black-box social recommendations. ACM Transactions on Information Systems 43, 1 (2024), 1–26

    2024

  30. [37]

    Shijie Wang, Jiani Huang, Zhikai Chen, Yu Song, Wenzhuo Tang, Haitao Mao, Wenqi Fan, Hui Liu, Xiaorui Liu, Dawei Yin, et al. 2024. Graph machine learning in the era of large language models (llms).ACM Transactions on Intelligent Systems and Technology (2024)

    2024

  31. [38]

    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

  32. [39]

    Lanning Wei, Huan Zhao, Zhiqiang He, and Quanming Yao. 2023. Neural architec- ture search for GNN-based graph classification. ACM Transactions on Information Systems 42, 1 (2023), 1–29

    2023

  33. [40]

    Christopher Williams. 1996. Computing with infinite networks. Advances in neural information processing systems 9 (1996)

    1996

  34. [41]

    Christopher KI Williams and Carl Edward Rasmussen. 2006. Gaussian processes for machine learning. Vol. 2. MIT press Cambridge, MA

    2006

  35. [43]

    Le Wu, Junwei Li, Peijie Sun, Richang Hong, Yong Ge, and Meng Wang. 2020. Diffnet++: A neural influence and interest diffusion network for social recom- mendation. IEEE Transactions on Knowledge and Data Engineering 34, 10 (2020), 4753–4766

    2020

  36. [44]

    Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and S Yu Philip. 2020. A comprehensive survey on graph neural networks. IEEE transactions on neural networks and learning systems 32, 1 (2020), 4–24

    2020

  37. [45]

    Zhihao Wu, Zhao Zhang, and Jicong Fan. 2024. Graph convolutional kernel machine versus graph convolutional networks. Advances in neural information processing systems 36 (2024)

    2024

  38. [46]

    Feng Xia, Jiaying Liu, Hansong Nie, Yonghao Fu, Liangtian Wan, and Xiangjie Kong. 2019. Random walks: A review of algorithms and applications. IEEE Transactions on Emerging Topics in Computational Intelligence 4, 2 (2019), 95–107

    2019

  39. [47]

    Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. 2018. How powerful are graph neural networks? arXiv preprint arXiv:1810.00826 (2018)

    2018 arXiv

  40. [48]

    Vishwanathan

    Pinar Yanardag and S.V.N. Vishwanathan. 2015. Deep Graph Kernels. In Proceed- ings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (Sydney, NSW, Australia)(KDD ’15). 1365–1374

    2015

  41. [49]

    Zhilin Yang, William Cohen, and Ruslan Salakhudinov. 2016. Revisiting semi- supervised learning with graph embeddings. In International conference on ma- chine learning. PMLR, 40–48

    2016

  42. [50]

    Zhen Zhang, Mianzhi Wang, Yijian Xiang, Yan Huang, and Arye Nehorai. 2018. Retgk: Graph kernels based on return probabilities of random walks. Advances in Neural Information Processing Systems 31 (2018)

    2018

  43. [51]

    Jie Zhou, Ganqu Cui, Shengding Hu, Zhengyan Zhang, Cheng Yang, Zhiyuan Liu, Lifeng Wang, Changcheng Li, and Maosong Sun. 2020. Graph neural networks: A review of methods and applications. AI open 1 (2020), 57–81

    2020

  44. [52]

    Xianchen Zhou and Hongxia Wang. 2023. On the Explainability of Graph Con- volutional Network With GCN Tangent Kernel. Neural Computation 35, 1 (2023), 1–26

    2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.