Pith. sign in

REVIEW 3 major objections 5 minor 44 references

Persistent Gaussian noise after every message-passing step keeps recurrent graph neural networks from asymptotic feature collapse, with a proven positive floor on stationary Dirichlet energy.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-31 15:16 UTC pith:RJH63BIW

load-bearing objection Clean IRF+spectral theorem: under global contraction, persistent Gaussian noise gives an explicit stationary Dirichlet-energy floor proportional to σ²λ₂; useful scoped theory, not a practical fix for ordinary GCNs. the 3 major comments →

arxiv 2607.28185 v1 pith:RJH63BIW submitted 2026-07-30 cs.LG cs.AIcs.ITmath.IT

Persistent Gaussian Perturbations Prevent Oversmoothing in Recurrent Graph Neural Networks

Mostafa Haghir Chehreghani This is my paper
classification cs.LG cs.AIcs.ITmath.IT
keywords Graph Neural NetworksOversmoothingStochastic Dynamical SystemsDirichlet EnergyGeometric ErgodicitySpectral Graph TheoryMarkov chainsInvariant measures
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Deep message-passing on graphs tends to make node features more and more alike until they collapse; this paper shows that simply adding independent Gaussian noise after every recurrent step stops that collapse. Under a global contraction assumption on the deterministic update, the hidden states form a geometrically ergodic Markov chain with a unique stationary distribution. The expected Dirichlet energy of that distribution is bounded below by a positive quantity proportional to noise variance times the graph’s spectral gap, so representations cannot settle onto the constant manifold. The result supplies a stochastic mechanism for preventing oversmoothing that sits alongside residual connections, normalization, and rewiring. Linear and nonlinear simulations match the predicted stationary energy and its quadratic scaling with noise intensity.

Core claim

Under global contraction of the projected deterministic map, additive Gaussian noise of intensity σ turns a recurrent graph update into a geometrically ergodic Markov chain on the zero-mean subspace. Its unique invariant measure has projected second moment at least σ²(N−1)d/(1−α²)>0, and therefore expected stationary Dirichlet energy at least (λ₂(L)/m)·σ²(N−1)d/(1−α²)>0. The dynamics therefore escape oversmoothing in the precise sense that limiting expected Dirichlet energy cannot vanish.

What carries the argument

The stationary second-moment identity E[∥H̄∥²_F]=E[∥Δ(H̄)∥²_F]+σ²(N−1)d, combined with the contractive bound E[∥Δ(H̄)∥²_F]≤α² E[∥H̄∥²_F] and the spectral-gap inequality E(H)≥(λ₂(L)/m)∥H̄∥²_F. Together they yield the explicit positive lower bound on stationary Dirichlet energy that proves escape from oversmoothing.

Load-bearing premise

The deterministic update must shrink distances by a uniform factor strictly less than one everywhere on the zero-mean representation space and must leave the all-equal representation fixed.

What would settle it

Run a recurrent graph network that meets the stated contraction condition, keep constant-variance Gaussian noise on after every step, and measure long-run average Dirichlet energy; if it still decays to zero for σ>0, or fails to scale proportionally to σ² and to the graph spectral gap, the central claim is false.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Recurrent, equilibrium, and continuous-time GNN layers can keep representation diversity by leaving additive noise active at inference time, not only during training.
  • Stationary Dirichlet energy grows with both noise variance and algebraic connectivity, giving an explicit rule of thumb for choosing σ.
  • Oversmoothing defined as vanishing expected Dirichlet energy is impossible under the paper’s assumptions, independently of depth.
  • Stochastic perturbations act as a dynamical mechanism complementary to residual connections, normalization, and graph rewiring.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If a positive energy floor can be proved under only local or average contraction, the same noise mechanism would apply directly to ordinary GCN and attention stacks used in practice.
  • The lower bound invites an optimal noise schedule that trades representation diversity against signal fidelity on downstream tasks—left open by the paper.
  • The same projected-Markov analysis could be repeated for multiplicative or state-dependent noise, which may produce different stationary geometries.
  • Methods that already inject noise (feature dropout, edge dropping) may be capturing a weak version of this stationary-energy effect whenever noise remains active at test time.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper studies recurrent graph neural networks with independent additive Gaussian noise injected after every message-passing step, interpreting the architecture as a stochastic dynamical system on the zero-mean representation subspace. Under graph connectivity, projection invariance, global Lipschitz contraction of the projected map Δ with rate α<1 and Δ(0)=0, and i.i.d. additive noise (Assumptions 4.1–4.4), it proves that the projected hidden states form a geometrically ergodic Markov chain with a unique invariant measure (Theorem 5.1). An exact stationary second-moment identity (Proposition 5.2) and a contractive variance bound then yield a strictly positive lower bound on stationary projected variance (Theorem 5.4). Combined with the spectral inequality for the combinatorial Laplacian (Lemma 3.3), this produces an explicit positive lower bound on expected stationary Dirichlet energy proportional to (λ₂(L)/m)·σ²(N−1)d/(1−α²) (Theorem 5.6), so asymptotic oversmoothing in the energy sense cannot occur. Three numerical experiments on linear contractive propagation and a recurrent GCN illustrate non-vanishing stationary energy and the predicted quadratic scaling in σ².

Significance. If the result holds under the stated assumptions, it cleanly isolates persistent additive noise as a distinct, depth-independent mechanism that prevents asymptotic feature collapse, complementary to residual connections, normalization, and rewiring. The derivation uses standard tools (contractive iterated random functions, Wasserstein ergodicity, spectral graph theory) in a transparent way and supplies an explicit, falsifiable scaling law linking noise variance, spectral gap, and stationary Dirichlet energy. The linear experiment and the quadratic-scaling regression (R²≈1) give direct numerical support for the theory. The main limitation on impact is that the load-bearing global-contraction assumption is stronger than what typical trained GNNs satisfy; the paper itself flags this in §7. Within that scope the contribution is a solid, self-contained theoretical guarantee rather than an immediately deployable architectural recipe.

major comments (3)
  1. [§4.4 Assumption 4.3; §7; Theorem 5.6] Assumption 4.3 (global Lipschitz contraction of Δ on the whole zero-mean subspace X with α<1 and Δ(0)=0) is load-bearing for Theorems 5.1, 5.4 and 5.6. Section 7 correctly notes that practical GNNs typically exhibit only local or average contraction. The manuscript should either (i) supply concrete sufficient conditions or examples of common layers (linear GCN, residual GCNII-style maps, GraphCON, etc.) under which global contraction holds with an explicit α, or (ii) state clearly that the guarantee is conditional on this strong premise and does not automatically transfer to ordinary trained GCNs. Without that clarification the practical reading of Theorem 5.6 is overstated.
  2. [§6.2 Experiment 2] Experiment 2 instantiates a nonlinear recurrent GCNConv+ReLU without an explicit contraction coefficient and without reporting any estimate of a local or empirical Lipschitz constant of the realized map. The qualitative match to the theory is suggestive, but it does not verify that the dynamics remain inside the regime of Assumptions 4.2–4.3. A short diagnostic (e.g., empirical contraction ratios along trajectories, or confirmation that biases are absent so Δ(0)=0) would make the experiment a genuine stress test rather than an illustration under unknown stability.
  3. [Abstract; §1; Definitions 3.2–3.3] The escape-from-oversmoothing claim is defined solely via non-vanishing expected Dirichlet energy (Definitions 3.2–3.3, Theorem 5.6). The paper is explicit that it does not address optimization or downstream accuracy (§1, §7). For a journal audience this is acceptable only if the abstract and introduction consistently frame the result as a geometric/dynamical guarantee, not as a proven remedy for oversmoothing in trained predictive models. A brief tightening of the abstract’s final claim and of the “why this theorem matters for practice” paragraph would avoid over-reading.
minor comments (5)
  1. [Figure 1] Figure 1 is conceptual only; a short caption note that it is schematic (not a simulated trajectory) would help.
  2. [§6.1] In Experiment 1 the normalized adjacency  is used for propagation while the combinatorial Laplacian L is used for energy; this is consistent with the theory but should be stated once in the experimental protocol to avoid confusion.
  3. [§5.2] Lemma 3.1 and Proposition 5.2 use the projected noise covariance correctly; a one-line remark that the same second-moment identity holds for any centered noise with E∥Ξ̄∥_F²=(N−1)d (Gaussianity needed only for the explicit covariance) would align the text with Remark 5.1.
  4. [§2] Typos / style: “tooversmoothing” (p.1), “dynamicalsystems” and similar concatenated words appear in the related-work section; a pass for spacing and hyphenation is needed.
  5. [§6.3] The OLS slope 137.86 in Experiment 3 is left without an order-of-magnitude comparison to the theoretical prefactor λ₂(N−1)d/(m(1−α²)); even a rough plug-in with a proxy α would strengthen the quantitative claim.

Circularity Check

0 steps flagged

No circularity: the stationary Dirichlet-energy lower bound is a forward consequence of stated contraction, noise, and spectral assumptions plus classical IRF facts.

full rationale

The load-bearing chain is Assumptions 4.1–4.4 → geometric ergodicity of the projected Markov chain (Thm 5.1, via classical contractive iterated-random-function results) → exact stationary second-moment identity (Prop 5.2) → contractive variance estimate (Lem 5.3) → strictly positive stationary projected variance (Thm 5.4) → spectral lower bound (Lem 3.3) → positive stationary Dirichlet energy (Thm 5.6). None of these steps defines the conclusion into the premises: oversmoothing/escape are geometric definitions (Defs 3.2–3.3), the noise second moment is computed from the projected Gaussian (Lem 3.1), and α, σ, λ₂ enter as free parameters of the bound rather than quantities fitted to the target energy. Experiment 3’s OLS fit of E∞ vs σ² is a post-hoc numerical check of the already-proved quadratic scaling, not an input to any theorem. Author self-citations appear only in Related Work (prior oversmoothing/oversquashing heuristics) and are not invoked inside the proofs. The derivation is therefore self-contained against its stated axioms; weakness of the global-contraction premise is a scope/correctness issue, not circularity.

Axiom & Free-Parameter Ledger

3 free parameters · 7 axioms · 0 invented entities

The central theorem is conditional mathematics: once global contraction, projection invariance, connectivity, and i.i.d. additive noise with finite second moments are granted, the energy floor follows from textbook IRF ergodicity and Laplacian spectral geometry. No new physical entities. Free parameters (α, σ, graph spectrum, N, d, m) appear in the bound as inputs, not as quantities fitted to force the claim. The strongest modeling cost is the global Lipschitz assumption, which the paper flags but which practical message-passing maps rarely satisfy globally.

free parameters (3)
  • contraction constant α = 0.95 in Experiment 1; unspecified/uncontrolled in nonlinear Exp 2
    Chosen in (0,1) as part of Assumption 4.3; in Exp 1 fixed at 0.95 by hand. Enters the denominator 1−α² of every quantitative bound. Not fitted to data to prove the theorem, but is a free modeling choice that controls the bound’s magnitude.
  • noise intensity σ = experimental grid 0 to 0.20
    Free design parameter of the architecture; theorem states the energy floor scales with σ². Grid {0,…,0.20} chosen for experiments. Not fitted to produce the existence of a positive floor.
  • Exp 3 OLS slope (137.86) = 137.86 (R²=1.0000)
    Post-hoc regression coefficient relating empirical stationary energy to σ² on the nonlinear GCN; used only to illustrate quadratic scaling, not to establish Theorem 5.6.
axioms (7)
  • domain assumption Graph G is finite, connected, undirected (Assumption 4.1), so λ₂(L)>0.
    Standard spectral-graph hypothesis; needed for any positive energy lower bound via Lemma 3.3.
  • domain assumption Projection invariance: P F(H) = P F(P H) for all H (Assumption 4.2).
    Makes projected dynamics autonomous on Im(P). Plausible for relative message-passing without global bias; structural rather than generic for all GNNs.
  • ad hoc to paper Global Lipschitz contraction of Δ on X with rate α<1 and Δ(0)=0 (Assumption 4.3).
    Load-bearing stability hypothesis enabling classical IRF ergodicity and the variance floor. Stronger than typical practical GNN behavior; paper acknowledges this in §7.
  • domain assumption i.i.d. centered additive noise with finite second moments; Gaussian for explicit covariance (Assumption 4.4).
    Standard stochastic-excitation hypothesis; Gaussianity used for E∥Ξ̄∥_F²=(N−1)d (Lemma 3.1).
  • standard math Contractive iterated random functions on a complete metric space admit a unique invariant probability and geometric W₂ ergodicity (Diaconis–Freedman / Hairer–Mattingly style).
    Invoked for Theorem 5.1; cited as classical, not re-proved in full.
  • standard math Courant–Fischer: Tr(H̄ᵀ L H̄) ≥ λ₂(L)∥H̄∥_F² for H̄ ⊥ ker(L).
    Lemma 3.3; converts projected variance into Dirichlet energy.
  • standard math W₂ convergence plus quadratic growth of E implies convergence of expected Dirichlet energy.
    Used in proof of Theorem 5.6 citing Villani; standard optimal-transport fact.

pith-pipeline@v1.2.0-daily-grok45 · 22539 in / 3903 out tokens · 72134 ms · 2026-07-31T15:16:27.518525+00:00 · methodology

0 comments
read the original abstract

Oversmoothing is a fundamental limitation of deep graph neural networks (GNNs), where repeated message passing causes node representations to become increasingly similar, eventually collapsing toward a low-dimensional subspace. This phenomenon limits the effective depth of message-passing architectures and motivates the search for mechanisms that preserve representation diversity. In this paper, we study a recurrent graph neural network in which independent Gaussian noise is injected after every propagation step and analyze the resulting architecture as a stochastic dynamical system. Under a standard global contraction assumption on the deterministic update, we prove that the hidden representations form a geometrically ergodic Markov chain admitting a unique invariant probability measure. Our main theoretical result establishes an explicit positive lower bound on the expected stationary Dirichlet energy, proportional to both the noise variance and the spectral gap of the underlying graph. Consequently, the stationary representations cannot collapse onto the constant manifold, providing a rigorous guarantee that asymptotic oversmoothing is prevented in the sense of non-vanishing Dirichlet energy. Our analysis reveals persistent stochastic perturbations as a fundamentally different mechanism for combating oversmoothing, complementing existing deterministic approaches based on residual connections, normalization, and graph rewiring. Finally, numerical experiments on both linear and nonlinear recurrent graph neural networks closely match the theoretical predictions, illustrating the emergence of a stationary distribution and the predicted dependence of the limiting Dirichlet energy on the noise intensity.

Figures

Figures reproduced from arXiv: 2607.28185 by Mostafa Haghir Chehreghani.

Figure 1
Figure 1. Figure 1: Conceptual illustration of the long-term behavior of recurrent graph neural networks. Left: with [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Logical dependency of the theoretical development. Theorem [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Evolution of the Dirichlet energy for the linear stochastic propagation model. The deterministic [PITH_FULL_IMAGE:figures/full_fig_p019_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Evolution of the projected feature energy [PITH_FULL_IMAGE:figures/full_fig_p019_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Stationary Dirichlet energy as a function of the noise level. The equilibrium energy increases [PITH_FULL_IMAGE:figures/full_fig_p020_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Stationary projected feature energy as a function of the noise level. The stationary projected [PITH_FULL_IMAGE:figures/full_fig_p020_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Evolution of the Dirichlet energy in the nonlinear recurrent graph convolutional network. The [PITH_FULL_IMAGE:figures/full_fig_p022_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Stationary Dirichlet energy of the nonlinear recurrent GCN as a function of the noise level. [PITH_FULL_IMAGE:figures/full_fig_p023_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Verification of the quadratic scaling law. The stationary Dirichlet energy obtained from repeated [PITH_FULL_IMAGE:figures/full_fig_p024_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

44 extracted references

  1. [1]

    On the bottleneck of graph neural networks and its practical implications

    Uri Alon and Eran Yahav. On the bottleneck of graph neural networks and its practical implications. InInternational Conference on Learning Representations, 2021. 25

    2021

  2. [2]

    Springer Monographs in Mathematics

    Ludwig Arnold.Random Dynamical Systems. Springer Monographs in Mathematics. Springer, Berlin, 1998

    1998

  3. [3]

    Zico Kolter, and Vladlen Koltun

    Shaojie Bai, J. Zico Kolter, and Vladlen Koltun. Deep equilibrium models. InAdvances in Neural Information Processing Systems, 2019

    2019

  4. [4]

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

    Muhammet Balcilar, Guillaume Renton, Pierre Héroux, Benoît Gaüzère, Sébastien Adam, and Paul Honeine. Analyzing the expressive power of graph neural networks in a spectral perspective. InInter- national Conference on Learning Representations, 2021

    2021

  5. [5]

    GraphNorm: A princi- pled approach to accelerating graph neural network training

    Tianle Cai, Shengjie Luo, Keyulu Xu, Di He, Tie-Yan Liu, and Liwei Wang. GraphNorm: A princi- pled approach to accelerating graph neural network training. InInternational Conference on Machine Learning, 2021

    2021

  6. [6]

    Gorinova, Stefan Webb, Emanuele Rossi, and Michael M

    Benjamin Paul Chamberlain, James Rowbottom, Maria I. Gorinova, Stefan Webb, Emanuele Rossi, and Michael M. Bronstein. GRAND: Graph neural diffusion. InInternational Conference on Machine Learning, 2021

    2021

  7. [7]

    Half a decade of graph convolutional networks.Nature Machine Intelli- gence, 4(3):192–193, 2022

    Mostafa Haghir Chehreghani. Half a decade of graph convolutional networks.Nature Machine Intelli- gence, 4(3):192–193, 2022

    2022

  8. [8]

    Simple and deep graph convolutional networks

    Ming Chen, Zhe Wei, Zengfeng Huang, Bolin Ding, and Yaliang Li. Simple and deep graph convolutional networks. InInternational Conference on Machine Learning, 2020

    2020

  9. [9]

    Neuralordinarydifferential equations

    RickyT.Q.Chen, YuliaRubanova, JesseBettencourt, andDavidDuvenaud. Neuralordinarydifferential equations. InAdvances in Neural Information Processing Systems, 2018

    2018

  10. [10]

    Fan R. K. Chung.Spectral Graph Theory, volume 92 ofCBMS Regional Conference Series in Mathe- matics. American Mathematical Society, 1997

    1997

  11. [11]

    Iterated random functions.SIAM Review, 41(1):45–76, 1999

    Persi Diaconis and David Freedman. Iterated random functions.SIAM Review, 41(1):45–76, 1999

    1999

  12. [12]

    Revisiting over-smoothing in graph neural networks

    Panos Elinas and Edwin Bonilla. Revisiting over-smoothing in graph neural networks. InInternational Conference on Learning Representations, 2023

    2023

  13. [13]

    Bronstein

    Francesco Di Giovanni, Lorenzo Giusti, Federico Barbero, Giulia Luise, Pietro Liò, and Michael M. Bronstein. How does over-squashing affect graph neural networks?Transactions on Machine Learning Research, 2023

    2023

  14. [14]

    Simple GNN regularisation for 3D molecular property prediction and beyond

    Jonathan Godwin, Michael Schaarschmidt, Alexander Gaunt, Alvaro Sanchez-Gonzalez, Yulia Rubanova, Petar Veličković, James Kirkpatrick, and Peter Battaglia. Simple GNN regularisation for 3D molecular property prediction and beyond. InInternational Conference on Learning Representations, 2022

    2022

  15. [15]

    Implicit graph neural networks

    Fangda Gu, Heng Chang, Wenwu Zhu, Somayeh Sojoudi, and Laurent El Ghaoui. Implicit graph neural networks. InAdvances in Neural Information Processing Systems, volume 33, pages 11984–11995. Curran Associates, Inc., 2020

    2020

  16. [16]

    Yetanotherlookatharris’ergodictheoremformarkovchains

    MartinHairerandJonathanC.Mattingly. Yetanotherlookatharris’ergodictheoremformarkovchains. In Robert C. Dalang, Marco Dozzi, and Francesco Russo, editors,Seminar on Stochastic Analysis, Random Fields and Applications VI, volume 63 ofProgress in Probability, pages 109–117. Birkhäuser, 2011

    2011

  17. [17]

    Inductive representation learning on large graphs

    William Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. InAdvances in Neural Information Processing Systems, 2017

    2017

  18. [18]

    Mitigating over-smoothing in graph neural networks for node classification through adaptive early embedding and biased DropEdge procedures.Knowledge-Based Systems, 320:113615, 2025

    Fateme Hoseinnia, Mehdi Ghatee, and Mostafa Haghir Chehreghani. Mitigating over-smoothing in graph neural networks for node classification through adaptive early embedding and biased DropEdge procedures.Knowledge-Based Systems, 320:113615, 2025. 26

    2025

  19. [19]

    Kipf and Max Welling

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

    2017

  20. [20]

    Predict then propagate: Graph neural networks meet personalized pagerank

    Johannes Klicpera, Stefan Weißenberger, and Stephan Günnemann. Predict then propagate: Graph neural networks meet personalized pagerank. InInternational Conference on Learning Representations, 2019

    2019

  21. [21]

    DeeperGCN: All you need to train deeper GCNs, 2020

    Guohao Li, Chenxin Xiong, Ali Thabet, and Bernard Ghanem. DeeperGCN: All you need to train deeper GCNs, 2020

    2020

  22. [22]

    Deeper insights into graph convolutional networks for semi-supervised learning

    Qimai Li, Zhichao Han, and Xiao-Ming Wu. Deeper insights into graph convolutional networks for semi-supervised learning. InProceedings of the AAAI Conference on Artificial Intelligence, 2018

    2018

  23. [23]

    Meyn and Richard L

    Sean P. Meyn and Richard L. Tweedie.Markov Chains and Stochastic Stability. Cambridge University Press, 2 edition, 2009

    2009

  24. [24]

    Mitigating over-squashing in graph few-shot learn- ing by leveraging local and global similarities.Applied Soft Computing, 184:113863, 2025

    Yassin Mohamadi and Mostafa Haghir Chehreghani. Mitigating over-squashing in graph few-shot learn- ing by leveraging local and global similarities.Applied Soft Computing, 184:113863, 2025

    2025

  25. [25]

    Content augmented graph neural networks.ACM Transactions on the Web, 19(4):40:1– 40:19, 2025

    Fatemeh Gholamzadeh Nasrabadi, AmirHossein Kashani, Pegah Zahedi, and Mostafa Haghir Chehreghani. Content augmented graph neural networks.ACM Transactions on the Web, 19(4):40:1– 40:19, 2025

    2025

  26. [26]

    Revisiting over-smoothing in graph neural networks, 2020

    Hoang NT and Tyler Maehara. Revisiting over-smoothing in graph neural networks, 2020

    2020

  27. [27]

    Graph neural networks exponentially lose expressive power for node classification

    Kenta Oono and Taiji Suzuki. Graph neural networks exponentially lose expressive power for node classification. InInternational Conference on Learning Representations, 2020

    2020

  28. [28]

    Graph neural ordinary differential equations, 2019

    Michael Poli, Stefano Massaroli, Junyoung Park, Atsushi Yamashita, Hajime Asama, and Jinkyoo Park. Graph neural ordinary differential equations, 2019

    2019

  29. [29]

    Recipe for a general, powerful, scalable graph transformer

    Ladislav Rampášek, Michael Galkin, Vijay Prakash Dwivedi, Anh Tuan Luu, Guy Wolf, and Dominique Beaini. Recipe for a general, powerful, scalable graph transformer. InAdvances in Neural Information Processing Systems, 2022

    2022

  30. [30]

    DropEdge: Towards deep graph con- volutional networks on node classification

    Yu Rong, Wenbing Huang, Tingyang Xu, and Junzhou Huang. DropEdge: Towards deep graph con- volutional networks on node classification. InInternational Conference on Learning Representations, 2020

    2020

  31. [31]

    Konstantin Rusch, Michael M

    T. Konstantin Rusch, Michael M. Bronstein, and Siddhartha Mishra. Graph-coupled oscillator networks. InInternational Conference on Machine Learning, 2023

    2023

  32. [32]

    Random features strengthen graph neural net- works

    Ryoma Sato, Makoto Yamada, and Hisashi Kashima. Random features strengthen graph neural net- works. InInternational Conference on Machine Learning, 2021

    2021

  33. [33]

    Residual connections and normal- ization can provably prevent oversmoothing in graph neural networks

    Michael Scholkemper, Xinyi Wu, Ali Jadbabaie, and Soledad Villar. Residual connections and normal- ization can provably prevent oversmoothing in graph neural networks. InInternational Conference on Learning Representations, 2025

    2025

  34. [34]

    Shuman, Sunil Narang, Pascal Frossard, Antonio Ortega, and Pierre Vandergheynst

    David I. Shuman, Sunil Narang, Pascal Frossard, Antonio Ortega, and Pierre Vandergheynst. The emerging field of signal processing on graphs.IEEE Signal Processing Magazine, 30(3):83–98, 2013

    2013

  35. [35]

    Bronstein

    Jake Topping, Francesco Di Giovanni, Benjamin Paul Chamberlain, Xiaowen Dong, and Michael M. Bronstein. Understanding over-squashing and bottlenecks on graphs via curvature. InInternational Conference on Learning Representations, 2022

    2022

  36. [36]

    Graph Attention Networks

    PetarVeličković, GuillemCucurull, ArantxaCasanova, AdrianaRomero, PietroLiò, andYoshuaBengio. Graph Attention Networks. InInternational Conference on Learning Representations, 2018

    2018

  37. [37]

    American Mathematical Society, Providence, RI, 2003

    Cédric Villani.Topics in Optimal Transportation, volume 58 ofGraduate Studies in Mathematics. American Mathematical Society, Providence, RI, 2003. 27

    2003

  38. [38]

    Springer, Berlin, 2009

    Cédric Villani.Optimal Transport: Old and New, volume 338 ofGrundlehren der mathematischen Wissenschaften. Springer, Berlin, 2009

    2009

  39. [39]

    How powerful are graph neural networks? InInternational Conference on Learning Representations, 2019

    Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? InInternational Conference on Learning Representations, 2019

    2019

  40. [40]

    Representation learning on graphs with jumping knowledge networks

    Keyulu Xu, Chengtao Li, Yonglong Tian, Tomohiro Sonobe, Ken ichi Kawarabayashi, and Stefanie Jegelka. Representation learning on graphs with jumping knowledge networks. InInternational Con- ference on Machine Learning, 2018

    2018

  41. [41]

    Do transformers really perform bad for graph representation? InAdvances in Neural Information Processing Systems, 2021

    Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, and Tie-Yan Liu. Do transformers really perform bad for graph representation? InAdvances in Neural Information Processing Systems, 2021

    2021

  42. [42]

    PairNorm: Tackling oversmoothing in GNNs

    Lingxiao Zhao and Leman Akoglu. PairNorm: Tackling oversmoothing in GNNs. InInternational Conference on Learning Representations, 2020

    2020

  43. [43]

    Dvornek, Xiaoxiao Li, Junzhou Yang, and James S

    Juntang Zhuang, Nicha C. Dvornek, Xiaoxiao Li, Junzhou Yang, and James S. Duncan. SDGNN: Learning stochastic graph neural networks, 2020

    2020

  44. [44]

    Zohrabi, S

    M. Zohrabi, S. Saravani, and Mostafa Haghir Chehreghani. Centrality-based and similarity-based neigh- borhood extension in graph neural networks.The Journal of Supercomputing, 80(16):24638–24663, 2024. 28

    2024