Framework for Designing Filters of Spectral Graph Convolutional Neural Networks in the Context of Regularization Theory

Asif Salim; Sumitra S

REVIEW 3 major objections 5 minor 34 references

Framework for Designing Filters of Spectral Graph Convolutional Neural Networks in the Context of Regularization Theory

T0 review · 3 major / 5 minor · reviewed 2026-08-27 · deepseek-v4-flash

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

Pith's one-line read Any monotonically increasing function of Laplacian eigenvalues, inverted, makes a valid low-pass graph filter, and this recipe's diffusion and random-walk filters beat established GCNN filters on citation data.

desk verdict Useful, modest paper on low-pass spectral GCNN filter design; the unifying theorem is stated too broadly and the ChebyNet analysis needs another pass. read the letter →

arxiv 2009.13801 v1 pith:7KK23GCP submitted 2020-09-29 cs.LG stat.ML

Asif Salim , Sumitra S This is my paper · ORCID

classification cs.LGstat.ML

keywords graphconvolutionalnetworksspectralfiltersregularizationtheoryLaplacianlow-passfilteringsemi-supervisednodeclassificationrandomwalkfrequencyresponse

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 proposes a general rule for designing spectral filters in graph convolutional networks: choose any function $r(\lambda)$ that increases monotonically over the Laplacian's eigenvalue range, and define the filter as the inverse of $r$ applied to the normalized graph Laplacian. Because the inverse of an increasing function is decreasing, the resulting filter attenuates high-frequency graph signals and passes low-frequency ones, which the paper argues is the correct regularization behavior for node classification. The framework recovers established filters—GCN, IGCN, GraphHeat, and a diffusion filter—as special cases, and it generates new $p$-step random-walk and cosine filters. On Cora, Citeseer, and Pubmed, the new diffusion and random-walk filters match or beat state-of-the-art spectral GCNN filters. The result matters because it turns filter choice into a checkable condition (monotonicity) plus a closed-form construction.

What carries the argument

The load-bearing object is the regularization function $r(\lambda)$, a real-valued function of Laplacian eigenvalues. The paper demands that $r$ be monotonically increasing on $[0, \lambda_{\max}]$ and expressible in closed form (for example, through a power series), and then sets the graph filter to $F = (r(\tilde L))^{-1}$. In the eigenbasis of the normalized Laplacian $\tilde L = U\Lambda U^T$, this acts on each spectral component with gain $1/r(\lambda_i)$, so monotone increase of $r$ translates directly into monotone decrease of the gain—the defining property of the low-pass filters the framework is designed to produce. The same machinery also carries the regularization interpretation: $\langle f, r(\tilde L) f\rangle$ is the smoothness penalty, and through the support-vector-kernel connection the filter $F$ is a graph kernel whenever the parameterization keeps it positive semidefinite.

What would settle it

Run the same filter-recipe comparison on a graph learning task where connected nodes tend to have opposite labels (a heterophilic setting): if the best achievable linear classifier needs a frequency response that rises at high Laplacian eigenvalues, the claim that every monotone-increasing $r(\lambda)$ yields the valid family of regularized filters is falsified for that setting. A simpler check: compute the optimal linear filter on such a graph and inspect its frequency response; if it is not monotonically decreasing, the framework's restriction is not universal.

Watch

Extended reading notes

Core claim

The central discovery is Theorem 1: if $r: [0,\lambda_{\max}] \to \mathbb{R}$ is monotonically increasing, then $F = (r(\tilde L))^{-1} = \sum_i (1/r(\lambda_i)) u_i u_i^T$ is a valid regularized graph convolution filter, because its frequency response $g(\lambda) = 1/r(\lambda)$ is monotonically decreasing and therefore gives more weight to low-frequency eigenspaces. This gives a design recipe: pick a monotone increasing regularization function with a closed-form power series, and invert it. The paper shows the recipe unifies the frequency responses of GCN (a first-order diffusion approximation), IGCN, GraphHeat, and the regularized Laplacian, and that a $p$-step random-walk filter with frequency response $(a-\lambda)^p$ and $a \ge 2$ improves on GCN by tuning the filter's response location. It also shows that the standard ChebyNet polynomial filter has the opposite regularization behavior—its regularization function decreases with $\lambda$—which the decoupled-filtering experiment confirms by producing accuracy below even a multilayer perceptron. Finally, the new filters, especially diffusion and the random-walk family, achieve the best accuracies among the compared spectral GCNN filters on all three citation datasets.

Load-bearing premise

The framework assumes that the ideal graph filter always lets slow, smooth signal variations through and suppresses fast, noisy ones—equivalently, that a monotonically decreasing frequency response is the right inductive bias—and it tests this only on citation networks where neighboring nodes usually share labels.

Editorial extensions

If this is right

If Theorem 1 holds, any monotone increasing $r(\lambda)$ with a convergent series expansion yields a filter with a known regularization meaning; regularized Laplacian, diffusion, $p$-step random walk, and cosine filters all qualify.
The framework identifies GCN and IGCN as approximations of diffusion-type regularization, GraphHeat as a diffusion-plus-identity filter, and the $p$-step random-walk filter with $a \ge 2$ as a tunable, better-regularized replacement for GCN.
The standard ChebyNet polynomial filter violates the monotonicity condition; when applied as a fixed pre-filter followed by a multilayer perceptron, it underperforms all other filters and plain MLP, supporting the framework's diagnosis.
On Cora, Citeseer, and Pubmed, the diffusion filter reaches the highest accuracy on Cora and Pubmed and the 3-step random-walk filter on Citeseer, with all $p$-step random-walk variants outperforming GCN and IGCN.

Reading between the lines

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

The paper only evaluates citation networks, where connected nodes tend to share labels; the monotone-inverse recipe would likely need modification for heterophilic graphs, where the optimal filter is often band-pass or high-pass, and that boundary is a natural test of the framework's scope.
The decoupled experiment suggests a practical diagnostic: pre-filtering features with $F=(r(\tilde L))^{-1}$ isolates whether a GCNN's gains come from the filter's inductive bias or from the learned nonlinearities.
Because the paper connects each monotone $r$ to a graph kernel, the same recipe could generate kernel families for tasks beyond node classification, such as link prediction.
A testable extension is to measure the learned frequency response of a trained network; if it drifts away from monotone decreasing while accuracy rises, that would show the monotonicity condition is sufficient but not necessary in end-to-end learning.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Request a human review

A listed scientist reviews the paper for a fee and the review publishes here regardless of verdict. See the reviewers or get listed.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: this is a genuinely useful but modest paper. The design rule—pick a positive increasing function of the eigenvalue and use its reciprocal as the filter—is simple, and it produces a couple of new filters (p-step random walk with a≥2, cosine) that perform competitively on standard citation benchmarks. The paper is clearly written and the experiments follow the usual protocol with standard deviations. The decoupling experiment in Section 6.4 is a nice touch: it shows that the low-pass property alone helps even without learned filter parameters.

Where the paper gets into trouble is in the strength of its claims. Theorem 1 states that any monotonically increasing function on [0, λmax] is a valid regularization function, but monotonicity alone does not guarantee r(λ) > 0 and finite on the spectrum. Without that, r(˜L) may not be invertible and the kernel interpretation in Appendix A fails. This is not pedantic: Table 2 restricts the random walk function to a≥2, but Section 5.3 derives GCN by setting a=1, p=1, and at a=1 the function has a pole at λ=1 and changes sign on [0,2]. So GCN is not a special case of the theorem as stated. The internal inconsistency should be fixed by either extending the theorem to handle a=1 with care or by weakening the claim.

The ChebyNet analysis in Section 5.4 is also shaky. It assumes the learned polynomial coefficients θ_k are the Taylor coefficients of exp(λ), but nothing in ChebyNet training enforces that. The claim that ChebyNet's regularization is "the exact opposite of the expected behavior" is therefore not established. The empirical decoupling result (ChebyNet performs terribly when fixed) is interesting on its own, but the analytical justification needs another pass.

I don't think the framework is as novel as the authors suggest—unifying filters as inverses of increasing functions is basically the Smola-Kondor kernel connection, which they cite. But the paper does add two useful filter instances and a clean way to think about low-pass design. The experimental gains are small (1-2 points) and limited to three citation datasets, and no code is provided. Still, the core idea is sound enough to deserve refereeing. I'd send it to review with a request for major revision. The authors should tighten Theorem 1, fix the GCN special case, and either provide evidence for the ChebyNet assumption or drop that analysis.

Referee Report

3 major / 5 minor

Summary. The paper proposes a framework for designing spectral graph convolutional filters: given a monotonically increasing regularization function r(λ) on the spectrum of the normalized graph Laplacian, the filter is defined as F = (r(˜L))^{-1}, so that the frequency response g(λ) = 1/r(λ) is low-pass. Theorem 1 states that any monotonically increasing r on [0, λmax] yields a valid regularized filter. The paper lists four concrete regularization functions (regularized Laplacian, diffusion, p-step random walk, cosine), claims that GCN, IGCN, GraphHeat, and label propagation are special cases, and reports node classification results on Cora, Citeseer, and Pubmed showing that the proposed diffusion and random-walk filters outperform the baselines. A decoupling experiment separates low-pass filtering from network learning to study the effect of the filter alone.

Significance. If the framework were correctly stated, it would offer a clean conceptual bridge between regularization theory and spectral GCNN filter design, and the decoupling experiment in Section 6.4 is a useful diagnostic for isolating filter behavior from learned parameters. The empirical study follows a standard protocol and reports mean accuracies with standard deviations. However, the central claimed contribution is the unification of existing state-of-the-art filters under one theorem, and that claim is currently not established because the theorem omits necessary positivity/finiteness conditions and because two of the claimed instantiations (GCN and IGCN) fall outside the stated hypotheses. The ChebyNet analysis in Section 5.4 also relies on an assumption about learned coefficients that training does not impose. The framework is largely definitional, and its generality is limited to low-pass filters.

major comments (3)

[Section 5.2, Theorem 1 and Eq. (3)] Theorem 1 asserts that a monotonically increasing function on [0, λmax] is a valid regularization function for designing filters via F = (r(˜L))^{-1}. The theorem omits the necessary conditions that r be positive and finite on the spectrum. Without positivity and finiteness, r(˜L) may be singular or indefinite, so F may not exist or may fail to be positive semidefinite, and the RKHS/kernel interpretation in Remark 2 and Appendix A does not apply. Moreover, the reciprocal of an increasing function is not necessarily a well-defined decreasing frequency response on the whole interval if r changes sign. This is not a technicality: Table 2 restricts the p-step random walk family to a ≥ 2, yet Section 5.3 Case 1 derives GCN by setting a = 1, p = 1, where r(λ) = (1 - λ)^{-1} has a pole at λ = 1 and changes sign on [0, 2]. Case 2 derives IGCN with the same problematic choice a = 1. Consequently, GCN and IGCN are not special cases under the stated hypotheses, and the claimed unification in the abstract and Section 5.3 is unsupported.
[Section 5.4, ChebyNet analysis] The analysis of ChebyNet assumes that the learned polynomial coefficients θ_k are the coefficients of the Taylor expansion of exp(λ), writing y = U(Σ θ_k Λ^k) U^T f ≈ exp(˜L) f. ChebyNet training does not impose this constraint; the θ_k are free parameters learned from data, and the polynomial is evaluated on a rescaled Laplacian. The conclusion that ChebyNet's regularization is 'the exact opposite of the expected behavior' is therefore a statement about a hypothetical filter, not about the trained ChebyNet. The same issue affects Section 6.4 and Table 4, where the row labeled ChebyNet with g(λ) = Σ λ^k is the unlearned polynomial, not the ChebyNet filter used in Table 3. The usefulness of the decoupling experiment is undermined by this mislabeling.
[Section 5.1 and Definition 5.1] The framework assumes from the outset that a 'regularized graph convolution filter' must be a low-pass filter with monotonically decreasing frequency response. This is an additional modeling choice, not a consequence of the regularization functional in Eq. (2). As a result, the proposed framework excludes band-pass and high-pass filters, which are useful in some graph signal processing and learning tasks. The paper only evaluates the low-pass assumption on three citation networks and does not test it on tasks where high-frequency or band-pass information is relevant. The general claims in the abstract and conclusion should be scoped to low-pass regularized filters for semi-supervised node classification.

minor comments (5)

[Section 5.4, IGCN paragraph] The sentence 'the regularization happening in GCN is as desired' should refer to IGCN, since the paragraph discusses the IGCN filter y = θ(I - ˜L)^k f.
[Section 6.4, Observations] The text refers to 'Theory 1'; this should be 'Theorem 1'.
[Table 2 and Figure 1] For the p-step random walk family, the value of p should be stated explicitly in the figure captions and in the table discussion; the captions mention one-step and two-step random walk but the parameter p is not defined in the captions.
[Section 6.1] The description of Chebyshev polynomial approximation uses standard recurrence relations but does not specify how many terms are used for the p-step random walk filters; adding this detail would improve reproducibility.
[Table 3] Several of the reported accuracy differences between the proposed filters and the baselines are within one standard deviation (e.g., Cora diffusion 83.12 ± 0.37 vs. ChebyNet 82.16 ± 0.74); a statistical significance analysis would strengthen the claim of superiority.

Circularity Check

1 steps flagged · score 6.0 of 10

Theorem 1 restates Definition 5.1: the central 'valid regularized filter' claim is definitional, while the experiments are independent.

self definitional [Section 5.2, Definition 5.1 and Theorem 1 (Eq. 3)]
"Deﬁnition 5.1 (Regularized graph convolution ﬁlter). : The graph ﬁlter whose frequency response function gθ(λ) behaves like a low-pass ﬁlter, i.e, gθ(λ) should be a monotonically decreasing function in λ or equivalently the associated regularization function r(λ) = (gθ(λ))^{-1} should be a monotonically increasing function in λ. Theorem 1. A monotonically increasing function in the interval [0,λmax] is a valid regularization function to design regularized graph convolution ﬁlters using (3). Proof. ... The validity of r(λ) is established by the low pass ﬁltering and hence the proof."

Definition 5.1 defines a 'regularized graph convolution filter' as a low-pass filter whose inverse frequency response r(λ)=1/g(λ) is monotonically increasing. Theorem 1 then asserts that any monotonically increasing r is a valid regularization function, and its proof observes that g=1/r is decreasing and hence low-pass. The conclusion is exactly the defining property from Definition 5.1, restated as a theorem; no property beyond the definition is established. The framework's central validity claim is therefore self-definitional by construction.

full rationale

The paper's only genuinely circular step is the framework's central theorem, which restates Definition 5.1: a filter is declared 'regularized' precisely when 1/g(λ) is monotonically increasing, and Theorem 1 'proves' validity of such r by observing that g is decreasing. This is a definition presented as a derived result. The experimental part is not circular: Table 3 is an external benchmark comparison on Cora, Citeseer, and Pubmed, hyperparameters are tuned on validation data, and no fitted accuracy values are fed back into the framework's constants. There is no load-bearing self-citation; the Smola et al. citations are external support for the RKHS/kernel connection, and the empirical claims are self-contained against standard benchmarks. Separately, as a correctness risk rather than circularity, Theorem 1 omits positivity/finiteness conditions needed for invertibility and positive semidefiniteness, and the GCN 'special case' in Section 5.3 sets a=1 while Table 2 restricts a>=2, so that claimed derivation goes beyond the theorem's stated assumptions. The score reflects partial circularity: the central framework theorem reduces to its own definition, while the filter constructions and experiments retain independent content.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central framework rests on the smoothness functional of the graph Laplacian and the regularization operator connection from Smola and Kondor. The main domain assumption is that low-pass filtering is always desirable, and the paper adds a specific ad hoc identification for the ChebyNet analysis. The experimental hyperparameters (s, a, p, K) are free parameters tuned on validation data.

free parameters (5)

s (diffusion/GraphHeat strength) = tuned in [0.5, 1.5] on validation split
Controls the decay rate of the diffusion and heat filters; used in the experimental evaluation of Table 3 and Figure 3.
a (random walk offset) = tuned in [2,24] on validation split
Controls the offset in the p-step random walk filter (aI-L)^p; required to keep the filter positive semidefinite and low-pass.
p (random walk order) = 1, 2, 3
Sets the power of the random walk filter; the paper evaluates p=1, 2, and 3.
K (truncation order for matrix approximations) = tuned from {1,2,3,4} for diffusion, ChebyNet, GraphHeat; {1,2,3} for cosine
Truncation order for Taylor or Chebyshev approximations used in the implementations.
network hyperparameters (number of filters, dropout, L2 regularization) = filter count tuned among {16,32,64,128}; dropout 0.8; L2 factor 5e-4
Standard architecture hyperparameters from Kipf and Welling (2016), tuned on the validation set; they affect all models equally.

assumptions (5)

standard math Normalized graph Laplacian is real symmetric positive semidefinite with eigenvalues in [0,2].
Used to define the spectral filter and the interval in Theorem 1.
domain assumption The smoothness functional f^T L f measures how smoothly a signal varies over the graph, and low-frequency eigenvectors are smoother.
Motivates treating low eigenvalues as signal and high eigenvalues as noise; cited to Zhu and Goldberg (2009).
domain assumption Regularization functionals on graphs can be written as <f, r(L)f> following the continuous-domain analogy from Smola et al.
This is the bridge that lets the paper define filters as r(L)^{-1}; it is taken from Smola and Kondor (2003).
ad hoc to paper A valid filter for graph learning is a low-pass filter, equivalently its frequency response is monotonically decreasing.
Definition 5.1 makes the low-pass assumption the definition of a regularized graph convolution filter, without justifying it for all tasks.
ad hoc to paper ChebyNet learned parameters can be equated with the coefficients of the Taylor expansion of exp(lambda).
Section 5.4 uses this to conclude ChebyNet has the opposite regularization behavior; the paper does not justify this identification.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Framework for Designing Filters of Spectral Graph Convolutional Neural Networks in the Context of Regularization Theory." pith.science (2026). https://pith.science/paper/7KK23GCP

@misc{pith2026200913801,
  author       = {Pith},
  title        = {Pith review of: Framework for Designing Filters of Spectral Graph Convolutional Neural Networks in the Context of Regularization Theory},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7KK23GCP}},
  note         = {Machine review of arXiv:2009.13801}
}

read the original abstract

Graph convolutional neural networks (GCNNs) have been widely used in graph learning. It has been observed that the smoothness functional on graphs can be defined in terms of the graph Laplacian. This fact points out in the direction of using Laplacian in deriving regularization operators on graphs and its consequent use with spectral GCNN filter designs. In this work, we explore the regularization properties of graph Laplacian and proposed a generalized framework for regularized filter designs in spectral GCNNs. We found that the filters used in many state-of-the-art GCNNs can be derived as a special case of the framework we developed. We designed new filters that are associated with well-defined regularization behavior and tested their performance on semi-supervised node classification tasks. Their performance was found to be superior to that of the other state-of-the-art techniques.

Figures

Figures reproduced from arXiv: 2009.13801 by the authors.

**Figure 1.** Regularization function, r(λ). (a) regularized Laplacian (s = {0.5, 1, 1.5, 2}), (b) diffusion function (s = {0.5, 1, 1.5, 2}), (c) one-step random walk (a = {2, 3, 4, 5}), (d) 2-step random walk (a = {2, 3, 4, 5}), (e) inverse cosine function. Remark 1. There exists an inverse relationship between the regularization function and frequency response function. To impose high penalization on higher frequencies, regular… view at source ↗

**Figure 2.** Regularization function, r(λ). (a) ChebyNet, (b) GCN, (c) GraphHeat, (d) IGCN for k = 2 , (e) IGCN for k = 3. All graphs are for (c = {0.2, 0.5, 1.0, 1.5}) GCN: GCN filtering [13] operation can be written as, y = θ(I −L˜)f ≈ exp(−λ), where we assume parameter θ is 1 in the exponential approximation. The regularization function r(λ) = c.(1−λ) −1 ≈ c.exp(λ) where c is a constant determined by the parameter θ. Hence th… view at source ↗

**Figure 3.** Accuracy variation with hyper-parameters. (a) Diffusion, (b) 1-step RW, (c) 2-step RW, (d) [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 34 canonical work pages

[1]

Gradient-based learning applied to document recognition,

Y . Lecun, L. Bottou, Y . Bengio, and P. Haffner, “Gradient-based learning applied to document recognition,”Proceedings of the IEEE, vol. 86, pp. 2278–2324, Nov 1998

work page 1998
[2]

Geometric deep learning on graphs and manifolds using mixture model cnns,

F. Monti, D. Boscaini, J. Masci, E. Rodolà, J. Svoboda, and M. M. Bronstein, “Geometric deep learning on graphs and manifolds using mixture model cnns,” in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 5425–5434, July 2017

work page 2017
[3]

Neural message passing for quantum chemistry,

J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, and G. E. Dahl, “Neural message passing for quantum chemistry,” in Proceedings of the 34th International Conference on Machine Learning-Volume 70, pp. 1263–1272, JMLR. org, 2017

work page 2017
[4]

Convolutional networks on graphs for learning molecular ﬁngerprints,

D. K. Duvenaud, D. Maclaurin, J. Iparraguirre, R. Bombarell, T. Hirzel, A. Aspuru-Guzik, and R. P. Adams, “Convolutional networks on graphs for learning molecular ﬁngerprints,” in Advances in neural information processing systems, pp. 2224–2232, 2015

work page 2015
[5]

Molecular graph convolutions: moving beyond ﬁngerprints,

S. Kearnes, K. McCloskey, M. Berndl, V . Pande, and P. Riley, “Molecular graph convolutions: moving beyond ﬁngerprints,” Journal of computer-aided molecular design , vol. 30, no. 8, pp. 595–608, 2016

work page 2016
[6]

Inductive representation learning on large graphs,

W. Hamilton, Z. Ying, and J. Leskovec, “Inductive representation learning on large graphs,” in Advances in neural information processing systems, pp. 1024–1034, 2017

work page 2017
[7]

Gated Graph Sequence Neural Networks

Y . Li, D. Tarlow, M. Brockschmidt, and R. Zemel, “Gated graph sequence neural networks,” arXiv preprint arXiv:1511.05493, 2015

work page Pith review arXiv 2015
[8]

Graph Attention Networks

P. Veliˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P. Lio, and Y . Bengio, “Graph attention networks,”arXiv preprint arXiv:1710.10903, 2017

work page Pith review arXiv 2017

Show all 34 references

[9]

F. R. Chung, Spectral graph theory. No. 92, American Mathematical Soc., 1997

1997
[10]

The emerging ﬁeld of signal processing on graphs: Extending high-dimensional data analysis to networks and other irregular domains,

D. I. Shuman, S. K. Narang, P. Frossard, A. Ortega, and P. Vandergheynst, “The emerging ﬁeld of signal processing on graphs: Extending high-dimensional data analysis to networks and other irregular domains,”IEEE signal processing magazine, vol. 30, no. 3, pp. 83–98, 2013

2013
[11]

Spectral networks and locally connected networks on graphs,

J. Bruna, W. Zaremba, A. Szlam, and Y . LeCun, “Spectral networks and locally connected networks on graphs,”arXiv preprint arXiv:1312.6203, 2013

2013 arXiv
[12]

Convolutional neural networks on graphs with fast localized spectral ﬁltering,

M. Defferrard, X. Bresson, and P. Vandergheynst, “Convolutional neural networks on graphs with fast localized spectral ﬁltering,” in Advances in neural information processing systems, pp. 3844–3852, 2016

2016
[13]

Semi-supervised classiﬁcation with graph convolutional networks,

T. N. Kipf and M. Welling, “Semi-supervised classiﬁcation with graph convolutional networks,” arXiv preprint arXiv:1609.02907, 2016. 10

2016 arXiv
[14]

Graph wavelet neural network,

B. Xu, H. Shen, Q. Cao, Y . Qiu, and X. Cheng, “Graph wavelet neural network,”arXiv preprint arXiv:1904.07785, 2019

1904 arXiv
[15]

Graph convolutional networks using heat kernel for semi-supervised learning,

B. Xu, H. Shen, Q. Cao, K. Cen, and X. Cheng, “Graph convolutional networks using heat kernel for semi-supervised learning,” in Proceedings of the 28th International Joint Conference on Artiﬁcial Intelligence, pp. 1928–1934, AAAI Press, 2019

1928
[16]

A regularization framework for learning from graph data,

D. Zhou and B. Schölkopf, “A regularization framework for learning from graph data,” inICML 2004 Workshop on Statistical Relational Learning and Its Connections to Other Fields (SRL 2004), pp. 132–137, 2004

2004
[17]

Regularization and semi-supervised learning on large graphs,

M. Belkin, I. Matveeva, and P. Niyogi, “Regularization and semi-supervised learning on large graphs,” inInternational Conference on Computational Learning Theory, pp. 624–638, Springer, 2004

2004
[18]

Kernels and regularization on graphs,

A. J. Smola and R. Kondor, “Kernels and regularization on graphs,” in Learning theory and kernel machines, pp. 144–158, Springer, 2003

2003
[19]

Label efﬁcient semi-supervised learning via graph ﬁltering,

Q. Li, X.-M. Wu, H. Liu, X. Zhang, and Z. Guan, “Label efﬁcient semi-supervised learning via graph ﬁltering,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 9582–9591, 2019

2019
[20]

Manifold regularization: A geometric framework for learning from labeled and unlabeled examples,

M. Belkin, P. Niyogi, and V . Sindhwani, “Manifold regularization: A geometric framework for learning from labeled and unlabeled examples,”Journal of machine learning research, vol. 7, no. Nov, pp. 2399–2434, 2006

2006
[21]

Semi-supervised learning using gaussian ﬁelds and harmonic functions,

X. Zhu, Z. Ghahramani, and J. D. Lafferty, “Semi-supervised learning using gaussian ﬁelds and harmonic functions,” in Proceedings of the 20th International conference on Machine learning (ICML-03), pp. 912–919, 2003

2003
[22]

Deep learning via semi-supervised embed- ding,

J. Weston, F. Ratle, H. Mobahi, and R. Collobert, “Deep learning via semi-supervised embed- ding,” inNeural networks: Tricks of the trade, pp. 639–655, Springer, 2012

2012
[23]

Simplifying graph convolutional networks,

F. Wu, A. Souza, T. Zhang, C. Fifty, T. Yu, and K. Weinberger, “Simplifying graph convolutional networks,” inProceedings of the 36th International Conference on Machine Learning, pp. 6861– 6871, 2019

2019
[24]

Diffusion improves graph learning,

J. Klicpera, S. Weißenberger, and S. Günnemann, “Diffusion improves graph learning,” in Advances in Neural Information Processing Systems, pp. 13333–13345, 2019

2019
[25]

Stability properties of graph neural networks,

F. Gama, J. Bruna, and A. Ribeiro, “Stability properties of graph neural networks,” arXiv preprint arXiv:1905.04497, 2019

1905 arXiv
[26]

Introduction to semi-supervised learning,

X. Zhu and A. B. Goldberg, “Introduction to semi-supervised learning,”Synthesis lectures on artiﬁcial intelligence and machine learning, vol. 3, no. 1, pp. 1–130, 2009

2009
[27]

The connection between regularization operators and support vector kernels,

A. J. Smola, B. Schölkopf, and K.-R. Müller, “The connection between regularization operators and support vector kernels,”Neural networks, vol. 11, no. 4, pp. 637–649, 1998

1998
[28]

Wavelets on graphs via spectral graph theory,

D. K. Hammond, P. Vandergheynst, and R. Gribonval, “Wavelets on graphs via spectral graph theory,”Applied and Computational Harmonic Analysis, vol. 30, no. 2, pp. 129–150, 2011

2011
[29]

Revisiting semi-supervised learning with graph embeddings,

Z. Yang, W. W. Cohen, and R. Salakhutdinov, “Revisiting semi-supervised learning with graph embeddings,” in Proceedings of the 33rd International Conference on International Conference on Machine Learning-Volume 48, pp. 40–48, 2016

2016
[30]

Deepwalk: Online learning of social representations,

B. Perozzi, R. Al-Rfou, and S. Skiena, “Deepwalk: Online learning of social representations,” in Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining, pp. 701–710, 2014

2014
[31]

Link-based classiﬁcation,

Q. Lu and L. Getoor, “Link-based classiﬁcation,” in Proceedings of the 20th International Conference on Machine Learning (ICML-03), pp. 496–503, 2003

2003
[32]

Adam: A method for stochastic optimization,

D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014

2014 arXiv
[33]

Understanding the difﬁculty of training deep feedforward neural networks,

X. Glorot and Y . Bengio, “Understanding the difﬁculty of training deep feedforward neural networks,” inProceedings of the thirteenth international conference on artiﬁcial intelligence and statistics, pp. 249–256, 2010

2010
[34]

Tensorﬂow: Large-scale machine learning on heterogeneous dis- tributed systems,

M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro, G. S. Corrado, A. Davis, J. Dean, M. Devin, et al., “Tensorﬂow: Large-scale machine learning on heterogeneous dis- tributed systems,”arXiv preprint arXiv:1603.04467, 2016. 11 A Regularization in graphs, support vec...

2016 arXiv

Pith tools

Reviewed August 27, 2026 · model on record in the stance chip above.

[1] [1]

Gradient-based learning applied to document recognition,

Y . Lecun, L. Bottou, Y . Bengio, and P. Haffner, “Gradient-based learning applied to document recognition,”Proceedings of the IEEE, vol. 86, pp. 2278–2324, Nov 1998

work page 1998

[2] [2]

Geometric deep learning on graphs and manifolds using mixture model cnns,

F. Monti, D. Boscaini, J. Masci, E. Rodolà, J. Svoboda, and M. M. Bronstein, “Geometric deep learning on graphs and manifolds using mixture model cnns,” in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 5425–5434, July 2017

work page 2017

[3] [3]

Neural message passing for quantum chemistry,

J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, and G. E. Dahl, “Neural message passing for quantum chemistry,” in Proceedings of the 34th International Conference on Machine Learning-Volume 70, pp. 1263–1272, JMLR. org, 2017

work page 2017

[4] [4]

Convolutional networks on graphs for learning molecular ﬁngerprints,

D. K. Duvenaud, D. Maclaurin, J. Iparraguirre, R. Bombarell, T. Hirzel, A. Aspuru-Guzik, and R. P. Adams, “Convolutional networks on graphs for learning molecular ﬁngerprints,” in Advances in neural information processing systems, pp. 2224–2232, 2015

work page 2015

[5] [5]

Molecular graph convolutions: moving beyond ﬁngerprints,

S. Kearnes, K. McCloskey, M. Berndl, V . Pande, and P. Riley, “Molecular graph convolutions: moving beyond ﬁngerprints,” Journal of computer-aided molecular design , vol. 30, no. 8, pp. 595–608, 2016

work page 2016

[6] [6]

Inductive representation learning on large graphs,

W. Hamilton, Z. Ying, and J. Leskovec, “Inductive representation learning on large graphs,” in Advances in neural information processing systems, pp. 1024–1034, 2017

work page 2017

[7] [7]

Gated Graph Sequence Neural Networks

Y . Li, D. Tarlow, M. Brockschmidt, and R. Zemel, “Gated graph sequence neural networks,” arXiv preprint arXiv:1511.05493, 2015

work page Pith review arXiv 2015

[8] [8]

Graph Attention Networks

P. Veliˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P. Lio, and Y . Bengio, “Graph attention networks,”arXiv preprint arXiv:1710.10903, 2017

work page Pith review arXiv 2017