Pith. sign in

REVIEW 4 major objections 4 minor 59 references

Hyperbolic-PDE GNN: Spectral Graph Neural Networks in the Perspective of A System of Hyperbolic Partial Differential Equations

T0 review · 4 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 The paper claims that reformulating message passing as the wave equation $\partial^2 X/\partial t^2 = a^2 \tilde{L} X$ forces every node feature to live in a space spanned by the graph Laplacian's eigenvectors, and uses that bridge to…

desk verdict A useful experimental wrapper for spectral GNNs, but the central proof has a matrix mismatch in Appendix A.2 that takes down The stated Theorems 3.1 and 3.3. read the letter →

arxiv 2505.23014 v1 pith:WXPIQMJS submitted 2025-05-29 cs.LG

Juwei Yue , Haikuo Li , Jiawei Sheng , Xiaodong Li , Taoyu Su , Tingwen Liu , Li Guo This is my paper
classification cs.LG MSC 68T0705C5035L05
keywords graphneuralnetworkshyperbolicpartialdifferentialequationsmessagepassingspectralLaplacianeigenvectorspolynomialfilterssignalprocessing
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

This paper tries to establish that message passing in graph neural networks can be modelled as a system of hyperbolic partial differential equations, $\partial^2 X/\partial t^2 = a^2 \tilde{L} X$, and that the node features obtained from this system lie in a solution space spanned by eigenvectors of the graph Laplacian. If true, this would mean that GNN representations carry the graph's topological structure by construction, and that the features at any time can be decomposed into a superposition of these eigenvector bases. The paper further argues that replacing the Laplacian by a polynomial of it connects the hyperbolic-PDE paradigm to spectral GNNs, turning the PDE solver into a message-passing enhancement for methods like GCN, GPR-GNN, ChebNet, and BernNet. The experiments report accuracy gains on node classification and large improvements in fitting synthetic spectral filters on images.

What carries the argument

The central object is the graph wave equation $\partial^2 X/\partial t^2 = a^2 \tilde{L} X$ (a hyperbolic PDE), which the paper rewrites as a system of first-order constant-coefficient linear differential equations with state $w = [y; x]$. The block matrix $C$, claimed to be $\mathrm{diag}(I, a^2 \tilde{L})$, carries the derivation: its eigenvalues and eigenvectors would determine the fundamental matrix $\Phi(t)$ of the solution space, and thus would fix the basis in which node features are expressed. The polynomial approximation step then replaces $\tilde{L}$ by $P(L,t) = \sum_k \theta_k p_k(L)$, which is what connects the PDE formalism to spectral GNN filters and yields the leapfrog iteration $X(t_{m+1}) = (2I + \tau^2 P)X(t_m) - X(t_{m-1})$.

What would settle it

Resolve whether $dw/dt = Cw$ with the paper's state vector $w = [y; x]$ and $y = \partial x/\partial t$ reproduces the second-order wave equation: the block-diagonal $C$ gives $y' = y$ and $x' = a^2 \tilde{L} x$, which yields $x'' = x'$ rather than $x'' = a^2 \tilde{L} x$; the consistent coupling matrix $[[0, a^2 \tilde{L}],[I, 0]]$ has eigenvalues $\pm\sqrt{a^2 \tilde{\lambda}}$ instead of $1$ and $a^2 \tilde{\lambda}$. Inspecting whether Appendix A.2 actually uses the coupled matrix or the block-diagonal one settles whether the stated eigen-basis solution space exists.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that the second-order wave equation on a graph can be written as a first-order linear system $dw_l/dt = C w_l$, and that the fundamental matrix of this system is built from products $e^{\lambda' t} u'$ of the eigenvalues and eigenvectors of $C$. Because $C$ is taken to be block-diagonal with entries $I$ and $a^2 \tilde{L}$, the paper concludes that the solution space of the message-passing dynamics is spanned by the eigenvectors of the identity and of the graph Laplacian, so that node features are linear combinations of these topological eigen-bases and messages propagate along eigenvector directions. The paper then approximates the PDE's right-hand side by a polynomial $P(L,t)$, links each polynomial basis to a known spectral GNN, and reports that the hyperbolic reformulation improves both classification accuracy and the fidelity of learned spectral filters.

Load-bearing premise

The load-bearing premise is that the wave equation $\partial^2 X/\partial t^2 = a^2 \tilde{L} X$ can be rewritten as $dw/dt = Cw$ with $C = \mathrm{diag}(I, a^2 \tilde{L})$ and $w = [y; x]$, a block-diagonal matrix that does not actually couple position to velocity.

Editorial extensions

If this is right

  • Node features produced by the Hyperbolic-PDE GNN iteration lie in a space spanned by Laplacian eigenvectors at every time step, so messages propagate along graph-structural directions rather than coordinate axes.
  • The paradigm acts as a drop-in enhancement for spectral GNNs: replacing the polynomial $P(L,t)$ with any base model's convolution yields hyperbolized versions (Hyperbolic-GCN, Hyperbolic-GPR, Hyperbolic-Cheb, and others) that are claimed to improve accuracy.
  • The polynomial approximation allows arbitrary continuous filter functions to be realized, so the model's expressive reach includes band-pass, comb, and Runge filters with small approximation error.
  • Because the model constrains the solution space explicitly, the learned graph spectral filters can be visualized and compared directly, as in the paper's filter analysis figures.
  • The time-stepping scheme gives a concrete, trainable architecture whose update rule depends only on the polynomial order, time step size, and terminal time.

Reading between the lines

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

  • If the first-order reduction is corrected to the consistent coupling matrix $[[0, a^2 \tilde{L}],[I, 0]]$, the fundamental solution is built from $e^{\pm\sqrt{a^2 \tilde{\lambda}}\,t}$ modes rather than from the paper's stated eigenvalues $1$ and $a^2 \tilde{\lambda}$; the topological-interpretability theorem would then need to be re-derived on the corrected matrix.
  • The empirical gains on heterophilic graphs suggest that the wave-equation discretization may act as a filter regularizer even independently of the solution-space theorem, so the scheme's stability and expressivity could be analysed separately from the claimed eigen-basis result.
  • A testable extension is to check whether the improvements persist when the wave equation is solved with the correct coupling matrix, which would separate the value of the hyperbolic-PDE discretization from the specific (possibly incorrect) spectral conclusion attached to it.
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

4 major / 4 minor

Summary. The paper proposes a graph neural network architecture, Hyperbolic-PDE GNN, built on modeling message passing as the wave equation ∂²X/∂t² = a² bL X on a graph. The authors claim that the solution space of this system is spanned by eigenvectors of the graph Laplacian, which equips node features with topological interpretability and distinguishes the approach from traditional spectral GNNs. They introduce a polynomial approximation of the Laplacian in the PDE, connect the resulting numerical scheme to spectral GNNs such as GCN, ChebNet, BernNet, and JacobiConv, and report extensive experiments on node classification (ten datasets) and image signal filtering (seven filters). The paper includes code, detailed hyperparameter search protocols, and a set of theoretical theorems (3.1, 3.3) and remarks (3.4) intended to establish the solution-space result.

Significance. The manuscript has clear strengths: the experimental effort is substantial, with multiple datasets, base models, and filters; the hyperparameter search is documented; and the code is released. The filter-fitting experiments in Section 4.2 are well suited to probe spectral expressiveness. However, the paper's central theoretical claim is not established: the first-order reduction in Appendix A.2 is algebraically wrong, the fundamental matrix formulas are internally inconsistent, and the proposed method's solution-space property is shared, by construction, with the polynomial spectral GNN baselines. Because the abstract and introduction frame the interpretability and topological-basis guarantee as the main contribution, these issues are load-bearing. The stress-test concern regarding the coupling matrix C is confirmed by direct inspection of Eqs. (25), (10), and (30).

major comments (4)
  1. [Appendix A.2, Eq. (25); Theorem 3.3; Table 1] The reduction of the wave equation (9) to a first-order system is incorrect. With w_l = [y_{:l}; x_{:l}] and y_{:l} = ∂x_{:l}/∂t, the identities in the appendix give dw_l/dt = [dy_{:l}/dt; dx_{:l}/dt] = [a² bL x_{:l}; y_{:l}], so the coupling matrix must be C = [[0, a² bL], [I, 0]]. The paper instead uses C = diag(I, a² bL), which yields the decoupled system [dy/dt; dx/dt] = [y; a² bL x] and is not equivalent to Eq. (9). Consequently, the eigenvalue list in Theorem 3.3 (λ'_1 = ... = λ'_n = 1 and λ'_{n+i} = a² bλ_i) and the basis in Table 1 are incorrect; the correct eigenvalues are ±a√(bλ_i). Since Remark 3.4 and the claimed topological decomposition of node features rely on these theorems, the central interpretability result is not proven as written. A corrected derivation could still yield an expansion of x(t) in Laplacian eigenvectors, but that is a different proof and the present statements need revision.
  2. [Eq. (10) and Eq. (30)] Even taking the paper's block-diagonal C as given, the fundamental matrix is computed incorrectly. For C = diag(I, a² bL), exp(Ct) = diag(e^{tI}, e^{a² bL t}), whereas the paper writes Φ(t) = diag(e^{It}, a² e^{bLt}) in Theorem 3.1 and Eq. (10). The second block should not have a multiplicative factor a² outside the exponential, and the exponent must contain a² bL. The series expansion in Eq. (30) is also internally inconsistent: the second diagonal block of C² is (a² bL)² = a^4 bL², but the appendix writes a² bL². These are not cosmetic typographical issues, because the entries of Φ(t) determine the claimed solution space and the time evolution of the system.
  3. [Section 3.4, Eqs. (15)-(20); Table 1] The claimed contrast between the hyperbolic-PDE paradigm and traditional spectral GNNs is not supported by the paper's own construction. In the experiments, P(L,t) in Eq. (15) is set to the spectral convolution of the base model, so the recurrence (20) applies a polynomial in L to the input at each step, and every polynomial in L shares the eigenvectors of L. The output of the recurrence therefore lies in the Laplacian eigenspace automatically; moreover, since the eigenvectors of the symmetric normalized Laplacian form a basis of R^n, any node feature matrix whatsoever is a linear combination of these eigenvectors. Table 1's dichotomy—unit-vector bases for traditional MP versus Laplacian-eigenvector bases for the proposed method—is therefore misleading, and the baselines used in the experiments (GCN, ChebNet, BernNet, JacobiConv, and others) already produce outputs in the Laplacian eigenspace. The paper needs to specify what the hyperbolic-PDE formulation contributes beyond standard polynomial spectral filtering; the present framing does not establish that contribution.
  4. [Table 5] The claim in Section 4.1 that 'upon enhancing all these methods to the system of hyperbolic PDEs, they all show varying degrees of performance improvement' is contradicted by the paper's own results. Table 5 reports several substantial degradations: Hyperbolic-SGC drops 2.02 points on CiteSeer, Hyperbolic-APPNP drops 4.36 points on Cornell, Hyperbolic-Bern drops 1.83 points on Actor and 2.34 points on Cornell, and Hyperbolic-Jacobi drops 1.70 points on Computers and 2.31 points on Cornell. The narrative and conclusions should be revised to reflect where the enhancement helps and where it hurts.
minor comments (4)
  1. [Section 2.1 and Eq. (9)] The matrix bL in Eq. (9) is never formally defined in the main text; from Appendix A.1 it appears to be the unnormalized Laplacian D - A. Since Theorems 3.1 and 3.3 refer to bL, this definition should be stated in the main text.
  2. [Section 3.3] The statement that replacing bL with the symmetric normalized Laplacian L 'differs only in the nonzero elements' is inaccurate: the unnormalized Laplacian D - A and the normalized Laplacian I - D^{-1/2} A D^{-1/2} have different eigenvalues and different eigenvectors, so the solution-space theorems for bL do not automatically transfer to L.
  3. [Abstract and Table 3] There are grammatical errors in the abstract ('formulates', 'a node features'), a typo in Table 3 ('Datesets'), and a typo in Appendix B ('JaocibConv'). These should be corrected in a revision.
  4. [Section 4.2] The text contains 'average of of squared error' and describes the R² score as 'the coefficient of determination (i.e., R2 score)', but Table 6 reports values labeled as both squared error and R²; the table caption and text should clarify which quantity is reported.

Circularity Check

2 steps flagged · score 8.0 of 10

The solution-space result is built into the chosen first-order matrix C and into the polynomial choice of P, rather than derived from the hyperbolic PDE.

  1. self definitional [Section 3.2 / Appendix A.2, Eq. (25) and Theorem 3.3]
    "Subsequently, let w_l = [y_{:l}, x_{:l}]∈R^{2n}, Equation (21) can be rewritten in the following form: dw_l/dt = Cw_l = [I_n 0; 0 a^2 bL] w_l ... matrix C = [I 0; 0 a^2 bL] has eigenvalues λ'_1=...=λ'_n=1, λ'_{n+1}=a^2 bλ_1,..., and corresponding 2n linearly independent eigenvectors ... Φ(t) = [e^{λ'_1 t}u'_1,...,e^{λ'_{2n}t}u'_{2n}] is a fundamental matrix of solution of Equation (9)."

    With w=[y;x] and y=∂x/∂t, Equation (9) gives dw/dt=[a^2 bL x; y], so the correct block matrix is [[0,a^2 bL],[I,0]]. The block-diagonal C used in Eq. (25) instead solves the decoupled system dy/dt=y, dx/dt=a^2 bL x. The eigenbasis in Theorem 3.3 and Table 1—vectors of I and of bL with eigenvalues 1 and a^2 bλ_i—is therefore read off from the chosen C, not derived from the wave equation. The claimed topological solution space is an artifact of the definition of C.

  2. self definitional [Section 3.4 (Eqs. 15, 20) and Section 4.1, paragraph 'To further validate...']
    "we let the polynomial P(x) in Equation (15) to be the spectral graph convolution of the base model ... Our approach explicitly constrains the solution space of all aforementioned base models, ensuring that they are solely expressed by eigenvectors representing the topological structure and do not introduce additional features."

    Every base model named (SGC, APPNP, GPR-GNN, ChebNet, BernNet, JacobiConv, ChebNetII) has a spectral convolution that is a polynomial in L or I−L, e.g., Z=Σθ_k L^k X. Setting P(L,t) to that polynomial makes recurrence (20), X(t_{m+1})=(2I+τ^2 P(L,t_m))X(t_m)−X(t_{m−1}), a polynomial-in-L iteration whose iterates lie in span{L^k X(0), L^k X(t_1)} and hence in the Laplacian eigenspaces. The advertised eigenbasis constraint is imposed by the polynomial-filter construction itself, independent of the hyperbolic PDE, so the 'prediction' reduces to the input design.

full rationale

The paper's central interpretability claim (Remark 3.4) is that modeling message passing as ∂²X/∂t²=a²bLX puts node features in a solution space spanned by Laplacian eigenvectors. Two load-bearing steps make this true by construction rather than by derivation. First, in Appendix A.2 the paper defines w_l=[y;x] with y=∂x/∂t and writes dw_l/dt=Cw_l with C=diag(I,a²bL); the wave equation actually forces the off-diagonal coupling matrix [[0,a²bL],[I,0]], so the block-diagonal C describes a different decoupled system. Theorem 3.3's fundamental matrix and eigenbasis are exactly the eigenbasis of the chosen C, so the conclusion is pre-loaded into Eq. (25). Second, the practical model in Section 3.4 sets P(L,t) to the base model's spectral convolution, which is a polynomial in the Laplacian for every tested base model. Recurrence (20) then iterates polynomial-in-L maps, so the output lies in the Laplacian eigenspaces by definition of P, with no need for the hyperbolic PDE. These are not merely unproven claims; the paper's own equations exhibit the reduction. The self-citation to Yue et al. (2025) (Graph Wave Networks) appears as a related-work/Euler-method citation and is not load-bearing for the solution-space theorem. No machine-checked or externally falsifiable derivation is offered for the central claim. The mathematical inconsistency in the first-order reduction is also a correctness risk, but the circularity is that the asserted eigenbasis result is equivalent to the chosen C and the chosen polynomial P. Overall score 8: the central claim is forced by definition/construction rather than established by the stated derivation.

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

The central claim rests on a mathematically incorrect conversion of the wave equation to first-order form, and on the standard but non-transferable assumption that normalized and unnormalized Laplacians are interchangeable. No new physical entities are introduced; the free parameters are all standard model hyperparameters.

free parameters (3)
  • Time step tau = searched over {0.2, 0.5, 1.0, 2.0, 5.0} (node classification) and {0.5, 0.75, ..., 5} (filtering)
    Controls the discretization of the PDE in Eqs. (17)-(20); it changes the effective polynomial and the stability of the recurrence, and is tuned per dataset via wandb.
  • Terminal time T = uniform in [1,20] (node classification), [1,10] or {0.5, 1, 2} (filtering)
    Determines how many recurrence steps are applied, i.e., the effective filter order; selected per dataset by hyperparameter search.
  • Propagation coefficient a = not identified; absorbed into trainable polynomial coefficients
    Introduced in Eq. (9) as the only PDE parameter, but the final model in Eq. (16) drops it in favor of learned W_k, so it is not independently estimated.
assumptions (4)
  • ad hoc to paper The wave equation on a graph, ∂²X/∂t² = a² bL X, can be rewritten as dw/dt = Cw with C = diag(I, a² bL).
    This is the premise of Theorems 3.1 and 3.3 in Appendix A.2, but it is inconsistent with the definitions w=[y;x] and y=∂x/∂t; the correct C is [[0, a² bL],[I,0]]. The results therefore do not follow from the stated equations.
  • domain assumption The unnormalized Laplacian bL can be replaced by the symmetric normalized Laplacian L because they 'differ only in the nonzero elements'.
    Used in Section 3.3 to switch to the standard spectral GNN setting; the two operators have different spectra and eigenvectors, so the theoretical solution-space claim is not transferable without additional argument.
  • standard math Stone-Weierstrass approximation theorem (cited as Stone, 1937) ensures polynomials can approximate any continuous filter.
    Invoked in Section 3.3 to justify polynomial approximation of the Laplacian function.
  • standard math Any polynomial of the Laplacian shares the Laplacian's eigenvectors.
    Implied by the spectral decomposition L=UΛU^T; this is what makes the 'solution space spanned by eigenvectors' claim true for any polynomial spectral filter, not specific to the PDE model.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hyperbolic-PDE GNN: Spectral Graph Neural Networks in the Perspective of A System of Hyperbolic Partial Differential Equations." pith.science (2026). https://pith.science/paper/WXPIQMJS

@misc{pith2026250523014,
  author       = {Pith},
  title        = {Pith review of: Hyperbolic-PDE GNN: Spectral Graph Neural Networks in the Perspective of A System of Hyperbolic Partial Differential Equations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WXPIQMJS}},
  note         = {Machine review of arXiv:2505.23014}
}
read the original abstract

Graph neural networks (GNNs) leverage message passing mechanisms to learn the topological features of graph data. Traditional GNNs learns node features in a spatial domain unrelated to the topology, which can hardly ensure topological features. In this paper, we formulates message passing as a system of hyperbolic partial differential equations (hyperbolic PDEs), constituting a dynamical system that explicitly maps node representations into a particular solution space. This solution space is spanned by a set of eigenvectors describing the topological structure of graphs. Within this system, for any moment in time, a node features can be decomposed into a superposition of the basis of eigenvectors. This not only enhances the interpretability of message passing but also enables the explicit extraction of fundamental characteristics about the topological structure. Furthermore, by solving this system of hyperbolic partial differential equations, we establish a connection with spectral graph neural networks (spectral GNNs), serving as a message passing enhancement paradigm for spectral GNNs.We further introduce polynomials to approximate arbitrary filter functions. Extensive experiments demonstrate that the paradigm of hyperbolic PDEs not only exhibits strong flexibility but also significantly enhances the performance of various spectral GNNs across diverse graph tasks.

Figures

Figures reproduced from arXiv: 2505.23014 by the authors.

Figure 1
Figure 1. The graph spectral filter learned by the methods on graph datasets. ditionally, our method also demonstrates competitive per￾formance on PubMed and Actor. These results attest to the effectiveness of Hyperbolic-PDE GNN. To further validate the performance improvement of model￾ing message passing as a system of hyperbolic partial differ￾ential equations, we choose several base models and modify them accordingly. Spec… view at source ↗
Figure 2
Figure 2. The image signals processed by the real filter (first row), Hyperbolic-GCN (second row), and GCN (third row). to them. Hyperbolic-GPR also circumvents this problem. 4.2. Signal Filtering on Images In this section, we investigate the fitting capability of polyno￾mial spectral filters under the influence of Hyperbolic-PDE GNN. We conduct experiment on 50 real images with a resolution of 100 × 100 from the Image Proces… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

59 extracted references · 39 canonical work pages

  1. [1]

    Analyzing the expressive power of graph neural networks in a spectral perspective

    Balcilar, M., Renton, G., H \' e roux, P., Ga \" u z \` e re, B., Adam, S., and Honeine, P. Analyzing the expressive power of graph neural networks in a spectral perspective. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021 . OpenReview.net, 2021

    work page 2021

  2. [2]

    M., Grattarola, D., Livi, L., and Alippi, C

    Bianchi, F. M., Grattarola, D., Livi, L., and Alippi, C. Graph neural networks with convolutional ARMA filters. IEEE Trans. Pattern Anal. Mach. Intell. , 2021. doi:10.1109/TPAMI.2021.3054830

    arXiv 2021

  3. [3]

    Specformer: Spectral graph neural networks meet transformers

    Bo, D., Shi, C., Wang, L., and Liao, R. Specformer: Spectral graph neural networks meet transformers. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023 . OpenReview.net, 2023

    work page 2023

  4. [4]

    Spectral networks and locally connected networks on graphs

    Bruna, J., Zaremba, W., Szlam, A., and LeCun, Y. Spectral networks and locally connected networks on graphs. In Bengio, Y. and LeCun, Y. (eds.), 2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Conference Track Proceedings , 2014

    work page 2014

  5. [5]

    I., Bronstein, M

    Chamberlain, B., Rowbottom, J., Gorinova, M. I., Bronstein, M. M., Webb, S., and Rossi, E. GRAND: graph neural diffusion. In Meila, M. and Zhang, T. (eds.), Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event , volume 139 of Proceedings of Machine Learning Research, pp.\ 1407--1418. PMLR , 2021

    work page 2021

  6. [6]

    Hyperbolic graph convolutional neural networks

    Chami, I., Ying, Z., R \' e , C., and Leskovec, J. Hyperbolic graph convolutional neural networks. In Wallach, H. M., Larochelle, H., Beygelzimer, A., d'Alch \' e - Buc, F., Fox, E. B., and Garnett, R. (eds.), Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14,...

    work page arXiv 2019

  7. [7]

    Q., Rubanova, Y., Bettencourt, J., and Duvenaud, D

    Chen, T. Q., Rubanova, Y., Bettencourt, J., and Duvenaud, D. Neural ordinary differential equations. In Bengio, S., Wallach, H. M., Larochelle, H., Grauman, K., Cesa - Bianchi, N., and Garnett, R. (eds.), Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018,...

    2018

  8. [8]

    Adaptive universal generalized pagerank graph neural network

    Chien, E., Peng, J., Li, P., and Milenkovic, O. Adaptive universal generalized pagerank graph neural network. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021 . OpenReview.net, 2021

    work page 2021

Show all 59 references
  1. [9]

    Convolutional neural networks on graphs with fast localized spectral filtering

    Defferrard, M., Bresson, X., and Vandergheynst, P. Convolutional neural networks on graphs with fast localized spectral filtering. In Lee, D. D., Sugiyama, M., von Luxburg, U., Guyon, I., and Garnett, R. (eds.), Advances in Neural Information Processing Systems 29: Annual Conf...

    2016

  2. [10]

    PDE-GCN: novel architectures for graph neural networks motivated by partial differential equations

    Eliasof, M., Haber, E., and Treister, E. PDE-GCN: novel architectures for graph neural networks motivated by partial differential equations. In Ranzato, M., Beygelzimer, A., Dauphin, Y. N., Liang, P., and Vaughan, J. W. (eds.), Advances in Neural Information Processing Systems...

    2021

  3. [11]

    and Lenssen, J

    Fey, M. and Lenssen, J. E. Fast graph representation learning with PyTorch Geometric . In ICLR Workshop on Representation Learning on Graphs and Manifolds, 2019

    2019

  4. [12]

    Addressing heterophily in graph anomaly detection: A perspective of graph spectrum

    Gao, Y., Wang, X., He, X., Liu, Z., Feng, H., and Zhang, Y. Addressing heterophily in graph anomaly detection: A perspective of graph spectrum. In Ding, Y., Tang, J., Sequeda, J. F., Aroyo, L., Castillo, C., and Houben, G. (eds.), Proceedings of the ACM Web Conference 2023, WW...

    2023

  5. [13]

    R., Soman, J., Regep, C., Liu, G., Hayter, J

    Gaudelet, T., Day, B., Jamasb, A. R., Soman, J., Regep, C., Liu, G., Hayter, J. B. R., Vickers, R., Roberts, C., Tang, J., Roblin, D., Blundell, T. L., Bronstein, M. M., and Taylor - King, J. P. Utilizing graph machine learning within drug discovery and development. Briefings ...

    2021 doi

  6. [14]

    Pyramid graph neural network: A graph sampling and filtering approach for multi-scale disentangled representations

    Geng, H., Chen, C., He, Y., Zeng, G., Han, Z., Chai, H., and Yan, J. Pyramid graph neural network: A graph sampling and filtering approach for multi-scale disentangled representations. In Singh, A. K., Sun, Y., Akoglu, L., Gunopulos, D., Yan, X., Kumar, R., Ozcan, F., and Ye, ...

    2023

  7. [15]

    S., Riley, P

    Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O., and Dahl, G. E. Neural message passing for quantum chemistry. In Precup, D. and Teh, Y. W. (eds.), Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 201...

    2017

  8. [16]

    and Wei, Z

    Guo, Y. and Wei, Z. Graph neural networks with learnable and optimal polynomial bases. In Krause, A., Brunskill, E., Cho, K., Engelhardt, B., Sabato, S., and Scarlett, J. (eds.), International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA , ...

    2023

  9. [17]

    Bernnet: Learning arbitrary graph spectral filters via bernstein approximation

    He, M., Wei, Z., Huang, Z., and Xu, H. Bernnet: Learning arbitrary graph spectral filters via bernstein approximation. In Ranzato, M., Beygelzimer, A., Dauphin, Y. N., Liang, P., and Vaughan, J. W. (eds.), Advances in Neural Information Processing Systems 34: Annual Conference...

    2021

  10. [18]

    Convolutional neural networks on graphs with chebyshev approximation, revisited

    He, M., Wei, Z., and Wen, J. Convolutional neural networks on graphs with chebyshev approximation, revisited. In Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., and Oh, A. (eds.), Advances in Neural Information Processing Systems 35: Annual Conference on Neural In...

    2022

  11. [19]

    Optimizing polynomial graph filters: A novel adaptive krylov subspace approach

    Huang, K., Cao, W., Ta, H., Xiao, X., and Li \` o , P. Optimizing polynomial graph filters: A novel adaptive krylov subspace approach. In Chua, T., Ngo, C., Kumar, R., Lauw, H. W., and Lee, R. K. (eds.), Proceedings of the ACM on Web Conference 2024, WWW 2024, Singapore, May 1...

    2024

  12. [20]

    G., Li, M., and Lio, P

    Huang, K., Wang, Y. G., Li, M., and Lio, P. How universal polynomial bases enhance spectral graph neural networks: Heterophily, over-smoothing, and over-squashing. In Forty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024 . OpenR...

    2024

  13. [21]

    Kingma, D. P. and Ba, J. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings , 2015

    2015

  14. [22]

    Kipf, T. N. and Welling, M. Semi-supervised classification with graph convolutional networks. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings . OpenReview.net, 2017

    2017

  15. [23]

    Predict then propagate: Graph neural networks meet personalized pagerank

    Klicpera, J., Bojchevski, A., and G \" u nnemann, S. Predict then propagate: Graph neural networks meet personalized pagerank. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019 . OpenReview.net, 2019 a

    2019

  16. [24]

    Diffusion improves graph learning

    Klicpera, J., Wei enberger, S., and G \" u nnemann, S. Diffusion improves graph learning. In Wallach, H. M., Larochelle, H., Beygelzimer, A., d'Alch \' e - Buc, F., Fox, E. B., and Garnett, R. (eds.), Advances in Neural Information Processing Systems 32: Annual Conference on N...

    2019

  17. [25]

    Evennet: Ignoring odd-hop neighbors improves robustness of graph neural networks

    Lei, R., Wang, Z., Li, Y., Ding, B., and Wei, Z. Evennet: Ignoring odd-hop neighbors improves robustness of graph neural networks. In Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., and Oh, A. (eds.), Advances in Neural Information Processing Systems 35: Annual Co...

    2022

  18. [26]

    G\( ^ 2 \)cn: Graph gaussian convolution networks with concentrated graph filters

    Li, M., Guo, X., Wang, Y., Wang, Y., and Lin, Z. G\( ^ 2 \)cn: Graph gaussian convolution networks with concentrated graph filters. In Chaudhuri, K., Jegelka, S., Song, L., Szepesv \' a ri, C., Niu, G., and Sabato, S. (eds.), International Conference on Machine Learning, ICML ...

    2022

  19. [27]

    A generalized neural diffusion framework on graphs

    Li, Y., Wang, X., Liu, H., and Shi, C. A generalized neural diffusion framework on graphs. In Wooldridge, M. J., Dy, J. G., and Natarajan, S. (eds.), Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI 2024, Thirty-Sixth Conference on Innovative Applications of Arti...

    2024 doi

  20. [28]

    B., Azizzadenesheli, K., Liu, B., Stuart, A

    Li, Z., Kovachki, N. B., Azizzadenesheli, K., Liu, B., Stuart, A. M., Bhattacharya, K., and Anandkumar, A. Multipole graph neural operator for parametric partial differential equations. In Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., and Lin, H. (eds.), Advances in Ne...

    2020

  21. [29]

    Revgnn: Negative sampling enhanced contrastive graph learning for academic reviewer recommendation

    Liao, W., Zhu, Y., Li, Y., Zhang, Q., Ou, Z., and Li, X. Revgnn: Negative sampling enhanced contrastive graph learning for academic reviewer recommendation. ACM Trans. Inf. Syst. , 43 0 (1): 0 1:1--1:26, 2025. doi:10.1145/3679200. URL https://doi.org/10.1145/3679200

    2025 doi

  22. [30]

    A graph-enhanced click model for web search

    Lin, J., Liu, W., Dai, X., Zhang, W., Li, S., Tang, R., He, X., Hao, J., and Yu, Y. A graph-enhanced click model for web search. In Diaz, F., Shah, C., Suel, T., Castells, P., Jones, R., and Sakai, T. (eds.), SIGIR '21: The 44th International ACM SIGIR Conference on Research a...

    2021

  23. [31]

    Hyperbolic graph neural networks

    Liu, Q., Nickel, M., and Kiela, D. Hyperbolic graph neural networks. In Wallach, H. M., Larochelle, H., Beygelzimer, A., d'Alch \' e - Buc, F., Fox, E. B., and Garnett, R. (eds.), Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Pro...

    2019

  24. [32]

    Z., Xiong, H., and Akoglu, L

    Ma, X., Wu, J., Xue, S., Yang, J., Zhou, C., Sheng, Q. Z., Xiong, H., and Akoglu, L. A comprehensive survey on graph anomaly detection with deep learning. IEEE Trans. Knowl. Data Eng. , 35 0 (12): 0 12012--12038, 2023. doi:10.1109/TKDE.2021.3118815. URL https://doi.org/10.1109...

    2023

  25. [33]

    Nguyen, T., Honda, H., Sano, T., Nguyen, V., Nakamura, S., and Nguyen, T. M. From coupled oscillators to graph neural networks: Reducing over-smoothing via a kuramoto model-based approach. In Dasgupta, S., Mandt, S., and Li, Y. (eds.), International Conference on Artificial In...

    2024

  26. [34]

    The pagerank citation ranking: Bringing order to the web

    Page, L., Brin, S., Motwani, R., and Winograd, T. The pagerank citation ranking: Bringing order to the web. Technical report, Stanford infolab, 1999

    1999

  27. [35]

    C., Lei, Y., and Yang, B

    Pei, H., Wei, B., Chang, K. C., Lei, Y., and Yang, B. Geom-gcn: Geometric graph convolutional networks. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020 . OpenReview.net, 2020

    2020

  28. [36]

    Platonov, O., Kuznedelev, D., Diskin, M., Babenko, A., and Prokhorenkova, L. A critical look at the evaluation of gnns under heterophily: Are we really making progress? In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 20...

    2023

  29. [37]

    Graph neural ordinary differential equations

    Poli, M., Massaroli, S., Park, J., Yamashita, A., Asama, H., and Park, J. Graph neural ordinary differential equations. CoRR, abs/1911.07532, 2019

    1911 arXiv

  30. [38]

    and Sarkar, R

    Rozemberczki, B. and Sarkar, R. Characteristic functions on graphs: Birds of a feather, from statistical descriptors to parametric models. In CIKM '20: The 29th ACM International Conference on Information and Knowledge Management, Virtual Event, Ireland, October 19-23, 2020 . ...

    2020

  31. [39]

    Multi-scale attributed node embedding

    Rozemberczki, B., Allen, C., and Sarkar, R. Multi-scale attributed node embedding. CoRR, abs/1909.13021, 2019

    1909 arXiv

  32. [40]

    K., Chamberlain, B., Rowbottom, J., Mishra, S., and Bronstein, M

    Rusch, T. K., Chamberlain, B., Rowbottom, J., Mishra, S., and Bronstein, M. M. Graph-coupled oscillator networks. In Chaudhuri, K., Jegelka, S., Song, L., Szepesv \' a ri, C., Niu, G., and Sabato, S. (eds.), International Conference on Machine Learning, ICML 2022, 17-23 July 2...

    2022

  33. [41]

    Pitfalls of graph neural network evaluation

    Shchur, O., Mumme, M., Bojchevski, A., and G \" u nnemann, S. Pitfalls of graph neural network evaluation. CoRR, abs/1811.05868, 2018

    2018 arXiv

  34. [42]

    W., and Vlimant, J

    Shlomi, J., Battaglia, P. W., and Vlimant, J. Graph neural networks in particle physics. Mach. Learn. Sci. Technol., 2 0 (2): 0 21001, 2021. doi:10.1088/2632-2153/ABBF9A

    2021 doi

  35. [43]

    I., Narang, S

    Shuman, D. I., Narang, S. K., Frossard, P., Ortega, A., and Vandergheynst, P. The emerging field of signal processing on graphs: Extending high-dimensional data analysis to networks and other irregular domains. IEEE Signal Process. Mag. , 30 0 (3): 0 83--98, 2013. doi:10.1109/...

    2013

  36. [44]

    Stone, M. H. Applications of the theory of boolean rings to general topology. Transactions of the American Mathematical Society, 41 0 (3): 0 375--481, 1937

    1937

  37. [45]

    M., Xia, H., Strohmer, T., Bertozzi, A

    Thorpe, M., Nguyen, T. M., Xia, H., Strohmer, T., Bertozzi, A. L., Osher, S. J., and Wang, B. GRAND++: graph neural diffusion with A source term. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022 . OpenReview.net, 2022

    2022

  38. [46]

    Graph attention networks

    Velickovic, P., Cucurull, G., Casanova, A., Romero, A., Li \` o , P., and Bengio, Y. Graph attention networks. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings . OpenReview.net, 2018

    2018

  39. [47]

    and Zhang, M

    Wang, X. and Zhang, M. How powerful are spectral graph neural networks. In Chaudhuri, K., Jegelka, S., Song, L., Szepesv \' a ri, C., Niu, G., and Sabato, S. (eds.), International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA , volume 162...

    2022

  40. [48]

    Dissecting the diffusion process in linear graph convolutional networks

    Wang, Y., Wang, Y., Yang, J., and Lin, Z. Dissecting the diffusion process in linear graph convolutional networks. In Ranzato, M., Beygelzimer, A., Dauphin, Y. N., Liang, P., and Vaughan, J. W. (eds.), Advances in Neural Information Processing Systems 34: Annual Conference on ...

    2021

  41. [49]

    Wu, F., Jr., A. H. S., Zhang, T., Fifty, C., Yu, T., and Weinberger, K. Q. Simplifying graph convolutional networks. In Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA . PMLR , 2019

    2019

  42. [50]

    Graph neural networks in recommender systems: A survey

    Wu, S., Sun, F., Zhang, W., Xie, X., and Cui, B. Graph neural networks in recommender systems: A survey. ACM Comput. Surv. , 55 0 (5): 0 97:1--97:37, 2023. doi:10.1145/3535101

    2023 doi

  43. [51]

    Wu, Z., Pan, S., Chen, F., Long, G., Zhang, C., and Yu, P. S. A comprehensive survey on graph neural networks. IEEE Trans. Neural Networks Learn. Syst. , 32 0 (1): 0 4--24, 2021. doi:10.1109/TNNLS.2020.2978386

    2021

  44. [52]

    Xhonneux, L. A. C., Qu, M., and Tang, J. Continuous graph neural networks. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event , volume 119 of Proceedings of Machine Learning Research, pp.\ 10432--10441. PMLR , 2020

    2020

  45. [53]

    W., and Salakhutdinov, R

    Yang, Z., Cohen, W. W., and Salakhutdinov, R. Revisiting semi-supervised learning with graph embeddings. In Proceedings of the 33nd International Conference on Machine Learning, ICML 2016, New York City, NY, USA, June 19-24, 2016 . JMLR.org, 2016

    2016

  46. [54]

    Graph wave networks

    Yue, J., Li, H., Sheng, J., Guo, Y., Zhang, X., Zhou, C., Liu, T., and Guo, L. Graph wave networks. In Long, G., Blumestein, M., Chang, Y., Lewin - Eytan, L., Huang, Z. H., and Yom - Tov, E. (eds.), Proceedings of the ACM on Web Conference 2025, WWW 2025, Sydney, NSW, Australi...

    2025

  47. [55]

    Adaptive diffusion in graph neural networks

    Zhao, J., Dong, Y., Ding, M., Kharlamov, E., and Tang, J. Adaptive diffusion in graph neural networks. In Ranzato, M., Beygelzimer, A., Dauphin, Y. N., Liang, P., and Vaughan, J. W. (eds.), Advances in Neural Information Processing Systems 34: Annual Conference on Neural Infor...

    2021

  48. [56]

    Node-oriented spectral filtering for graph neural networks

    Zheng, S., Zhu, Z., Liu, Z., Li, Y., and Zhao, Y. Node-oriented spectral filtering for graph neural networks. IEEE Trans. Pattern Anal. Mach. Intell. , 46 0 (1): 0 388--402, 2024. doi:10.1109/TPAMI.2023.3324937

    2024

  49. [57]

    Interpreting and unifying graph neural networks with an optimization framework

    Zhu, M., Wang, X., Shi, C., Ji, H., and Cui, P. Interpreting and unifying graph neural networks with an optimization framework. In Leskovec, J., Grobelnik, M., Najork, M., Tang, J., and Zia, L. (eds.), WWW '21: The Web Conference 2021, Virtual Event / Ljubljana, Slovenia, Apri...

    2021

  50. [58]

    Wingnn: Dynamic graph neural networks with random gradient aggregation window

    Zhu, Y., Cong, F., Zhang, D., Gong, W., Lin, Q., Feng, W., Dong, Y., and Tang, J. Wingnn: Dynamic graph neural networks with random gradient aggregation window. In Singh, A. K., Sun, Y., Akoglu, L., Gunopulos, D., Yan, X., Kumar, R., Ozcan, F., and Ye, J. (eds.), Proceedings o...

    2023

  51. [59]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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