Pith. sign in

REVIEW 4 major objections 5 minor 95 references

Resolving Oversmoothing with Opinion Dissensus

T0 review · 4 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

Pith's one-line read Deep graph networks can avoid oversmoothing by turning node features into opinions that provably reach dissensus rather than consensus.

desk verdict A strong empirical paper with a novel architecture, but the headline no-oversmoothing theorem is not proved for the actual BIMP model. read the letter →

arxiv 2501.19089 v2 pith:67P4XAPR submitted 2025-01-31 cs.LG

classification cs.LG
keywords oversmoothinggraphneuralnetworksopiniondynamicsdissensuscontinuous-depthGNNsbifurcationDirichletenergymessagepassing
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 establishes an analogy between oversmoothing in graph neural networks and consensus in opinion dynamics: treat node features as opinions, message passing as opinion exchange, and the collapse of node features as agents agreeing. It proves that any GNN whose aggregation is a linear opinion-dynamics step—including SGC/DGC-style linear layers and Laplacian-based continuous-depth models—will oversmooth for every graph and every input. Using this diagnosis, it designs BIMP, a continuous-depth GNN whose message passing is the nonlinear opinion dynamics model, and proves that with a time-independent input term whose entries are all distinct, the equilibrium cannot be a consensus, so oversmoothing is avoided. The paper reports that BIMP keeps its Dirichlet energy stable over 1,000 integration steps, holds accuracy out to 128 layers, and beats amended baselines on homophilic, heterophilic, and large-graph benchmarks.

What carries the argument

The load-bearing mechanism is the nonlinear opinion dynamics equation, written in BIMP as Ẋ = −dX + tanh[u(αX + AᵃX + X(Aᵒ)ᵀ + AᵃX(Aᵒ)ᵀ)] + X(0). The communication matrix Aᵃ and option matrix Aᵒ are learned via multi-head attention, and the vectorized dynamics reduce to a single effective adjacency matrix à = (Aᵒ + I) ⊗ (Aᵃ + I) with leading eigenvalue 4. Setting the attention parameter u = d/(α + 3) places the system at the pitchfork bifurcation point, where it is ultrasensitive to the input term, and choosing the input B = X(0) with unique entries guarantees that the equilibrium cannot be a consensus.

What would settle it

Run BIMP on any connected graph with a time-independent input B that has unique entries and integrate the dynamics for a very long horizon (e.g., 10,000 Euler steps with step size smaller than 1/d), measuring the Dirichlet energy; if the energy decays to machine precision for some initial condition while the trajectory remains bounded, the dissensus claim fails. More directly, search numerically for a bounded non-convergent trajectory of equation (11) that does not approach any equilibrium, which would break Lemma 5.5 and void the equilibrium-based argument of Theorem 5.6.

Watch

Extended reading notes

Core claim

The central claim is that oversmoothing in GNNs is the same phenomenon as opinion consensus in opinion dynamics, and that the remedy is to replace the linear aggregation rule with the nonlinear opinion dynamics model. For BIMP, the paper proves Theorem 5.6: when the input parameter B is time-independent and has unique entries, the dynamical system cannot converge to a state where all node features are identical, because at a consensus point the distinct input values force the dynamics to be nonzero. Combined with the convergence guarantee of Lemma 5.5, this rules out oversmoothing: the equilibrium reached by BIMP is a dissensus, so node features stay distinguishable with depth. The paper also proves that BIMP's gradients are bounded and do not vanish exponentially, and that its dynamics act as a tunable high-pass filter when the self-reinforcement parameter exceeds one, which explains its performance on heterophilic graphs.

Load-bearing premise

The dissensus guarantee assumes every trajectory of BIMP's dynamics actually converges to an equilibrium, because the proof rules out consensus equilibria but does not prove convergence for all initial conditions; the cited theorem gives almost-everywhere convergence, and the multi-option case is only sketched.

Editorial extensions

If this is right

  • Any linear message-passing GNN, discrete or continuous, is guaranteed to oversmooth on all graphs and inputs, so depth improvements require nonlinear or explicitly repulsive dynamics.
  • Architectures that add source terms to linear Laplacian dynamics (such as GRAND++-ℓ and GREAD variants) can still oversmooth because the dynamics remain linear; nonlinearity is what breaks consensus.
  • BIMP is the first continuous-depth GNN with a proof that oversmoothing is avoided when the input term has unique entries, not merely mitigated by normalization or residual connections.
  • The tunable filter view of BIMP—high-pass when α > 1, low-pass when α ≤ 1—gives a single hyperparameter to adapt a deep GNN to homophilic and heterophilic graphs without changing the architecture.
  • BIMP's gradient bounds imply that very deep versions of the model can be trained without the vanishing-gradient failure that accompanies oversmoothing in other architectures.

Reading between the lines

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

  • The dissensus guarantee is sensitive to the uniqueness of the input entries: if two nodes receive the same constant input, the proof's contradiction no longer applies, so a version of BIMP with shared inputs could plausibly exhibit oversmoothing and would be a direct test of the mechanism.
  • The bifurcation point choice makes BIMP ultrasensitive to the input, which suggests the model could be adapted as an uncertainty or out-of-distribution detector by examining how quickly node opinions diverge.
  • The option graph Aᵒ performs feature mixing across dimensions, an operation linear GNNs lack; this could be exploited for tasks where feature channels carry semantically distinct information, such as multi-modal or multi-relational graphs.
  • Because the proof only needs uniqueness of the input and convergence to equilibrium, other nonlinear opinion-dynamics variants—for example with stubborn agents or time-varying attention—could yield new families of provably non-oversmoothing GNNs.
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 / 5 minor

Summary. This paper draws an analogy between oversmoothing in GNNs and consensus in opinion dynamics, proves that several existing linear discrete and continuous GNN models oversmooth, and introduces BIMP, a continuous-depth GNN whose message passing is a nonlinear opinion dynamics model. The central theoretical claims are that BIMP does not oversmooth when the input parameter B has unique entries (Theorem 5.6), that its gradients are bounded and do not vanish exponentially (Theorem 5.7), and that it can be tuned between low-pass and high-pass behavior. The paper reports competitive or state-of-the-art accuracy on six homophilic and three heterophilic datasets, stability of Dirichlet energy at 1000 steps, and ablations over nonlinearities and oversmoothing mitigation techniques.

Significance. If the dissensus theorem were fully proved, the paper would make a valuable conceptual and architectural contribution: it gives a principled reason to import nonlinear opinion dynamics into GNN design, and it provides an architecture that is empirically stable at depths where standard baselines collapse. The theoretical analysis is ambitious and mostly transparent, with proofs collected in an appendix, and the experimental evaluation is unusually thorough, including amended baselines and depth sweeps. However, the key guarantee is currently supported by a proof that upgrades an almost-everywhere convergence theorem to all-trajectory convergence without justification, and the multi-option case that corresponds to the actual model is asserted rather than proved. These gaps are load-bearing for the paper's headline claims.

major comments (4)
  1. [Appendix A.2.4, Lemma 5.5] The proof of Lemma 5.5 invokes Hirsch's theorem [92], which guarantees convergence to equilibrium for almost every point with compact forward orbit, and then concludes that the dynamical system converges to an equilibrium set and that all trajectories tend to equilibrium. This is an invalid upgrade: Hirsch's theorem leaves open a measure-zero set of exceptional non-convergent or non-equilibrium orbits, and no additional argument (Lyapunov function, omega-limit set analysis, or monotone convergence for every initial condition) is supplied. Because Theorem 5.6 proves only that consensus cannot be an equilibrium when B has unique entries, it does not rule out oversmoothing along an exceptional trajectory whose omega-limit behavior is not an equilibrium. The theorem and the abstract's unconditional no-oversmoothing claim therefore need either a complete all-trajectory convergence proof or a qualified statement such as 'almost-everywhere' or 'for generic initial conditions'.
  2. [Appendix A.2.5, Theorem 5.6] The proof begins 'Without loss of generality, consider the case where the graph is undirected and the system has only one option (i.e., Ao=0)', but the actual BIMP model always learns an option graph Ao, and the reduction changes the effective adjacency matrix from (Ao+I)⊗(Aa+I) to Aa+I, altering both the eigenvalue bound used in Lemma 5.4 and the structure of the equilibrium equations. The multi-option case is dismissed in one sentence ('can be shown analogously') with no analogue of the block-decomposition, irreducibility, or contradiction argument. Since the trained model uses Ao≠0, the paper's central no-oversmoothing guarantee is not proved for the architecture that is evaluated; this needs a concrete proof for the full vectorized system of Equation (36).
  3. [Section 5.2 and Theorem 5.6] Theorem 5.6 requires B to be time-independent with unique entries, and the paper sets B=X(0)=ϕ(X_in). No property of the encoder ϕ guarantees that the resulting feature matrix has pairwise distinct entries; for example, a constant input or a linear encoder with a zero row can produce repeated entries. If entries coincide, the contradiction argument in Equations (49)–(53) does not apply, and the dissensus claim is not established for those inputs. The authors should either prove genericity of unique entries, add a small perturbation or tie-breaking mechanism, or state the theorem conditionally on the encoder output being in general position.
  4. [Appendix A.2.6, Theorem 5.7 and Lemma A.4] The proof that gradients do not vanish exponentially is based on expanding the product ∏(I+ΔtE_t) as I + Δt∑E_t + O(Δt^2) and calling this a 'recursive sum'. This expansion is only first order in Δt and, for M layers with MΔt=T, the omitted term contains O(M^2Δt^2) contributions that need not be negligible. More importantly, expressing the gradient as a sum shows only that the product is not obviously exponential; it does not provide a lower bound, and the gradients may still decay polynomially or become very small (as Lemma A.4 itself concedes: 'gradients may still become very small'). The abstract's 'well behaved gradients' claim needs a quantitative statement (e.g., a lower bound that decays at most polynomially in T) or a corrected proof with a controlled remainder.
minor comments (5)
  1. [Appendix B, Lemma B.3] Equation (112) writes 'tanh[u(α + 3y)]' where the intended expression appears to be 'tanh[u((α+3)y)]'; the missing multiplication makes the reduced equation hard to parse, and 'Lyapunov-Schimit' should be 'Lyapunov-Schmidt'.
  2. [Theorem 5.6 and Appendix A.2.5] The theorem statement does not mention the undirectedness or block-diagonalizability assumptions that the proof uses; these assumptions should be stated in the theorem or the proof should explain why they are without loss of generality.
  3. [Appendix A.1.3, Lemma A.2] The GREAD-F proof argues that divergence to −∞ produces oversmoothing because components with the fastest decay dominate; this is not the standard Dirichlet-energy definition of oversmoothing and should be clarified or replaced with a direct argument about feature differences.
  4. [Section 5.2] The paper sets B=X(0), which means the forcing term enters with the same sign as the initial condition; the paper should note that this choice also affects the effective dynamics through the ∂X1/∂X0 term in Equation (85), which is used in the gradient upper bound.
  5. [Appendix A.2.2 and Theorem 5.6 proof] There are small inconsistencies in notation, such as 'belief adjacency matrix' instead of 'option adjacency matrix' in Lemma 5.3 and the use of both ˜α and α in the proof of Theorem 5.6; these should be cleaned up for readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the dissensus theorem is conditional, proved from external dynamical-systems results, and B=X(0) is a design choice rather than a fitted prediction.

full rationale

The central theoretical claim, Theorem 5.6, states that BIMP will not oversmooth when the constant input B has unique entries. The proof is a genuine equilibrium analysis: assuming a consensus equilibrium leads to a contradiction because distinct entries b_m and b_n cannot both make the corresponding right-hand sides vanish simultaneously. This does not reduce to the assumptions by construction; it is a conditional mathematical statement. Setting B = X(0) is presented as a way to try to satisfy that condition, not as a fitted parameter that later reappears as a prediction. The claim is independent of the paper's own empirical benchmarks. The convergence prerequisite, Lemma 5.5, is the weakest point of the derivation: it invokes Hirsch's cooperative-systems theorem, which gives convergence almost everywhere, and then upgrades this to all trajectories without proof, and it dismisses the multi-option (Ao != 0) case as analogous. This is a proof gap and a correctness risk, but it is not circular reasoning, because the cited theorem is external and the conclusion is not contained in the hypotheses by definition. Similarly, the bifurcation-point choice u = d/(alpha+3) follows from the Jacobian analysis of the nonlinear opinion dynamics model from Refs. [35,36], not from fitting BIMP's outputs. Self-citations such as Ref. [65] appear only in related-work context and are not load-bearing for the oversmoothing guarantee. The empirical evaluation is against standard external benchmarks and does not feed fitted values back into the theorem. Overall, no step in the paper's derivation chain equates the claimed result with its own input by construction.

Assumptions & free parameters 5 free parameters · 5 assumptions · 2 invented entities

The central theoretical claims rest on two constructed matrices (A^a, A^o), two tuned hyperparameters (d, α), the chosen initial-condition input B=X(0), and the Hirsch convergence theorem. The no-oversmoothing guarantee is conditional on unique entries of B and on all trajectories reaching equilibrium; these conditions are not verified empirically in the paper.

free parameters (5)
  • d (damping) = 0.0086-1.0970 per dataset (Tables 7-8)
    Hyperparameter controlling decay; selected by Ray Tune on validation, affects dynamics and gradient bounds.
  • alpha (self-reinforcement) = 1 (homophilic), 1.5-2 (heterophilic)
    Hyperparameter setting the balance between smoothing and sharpening; chosen per dataset via hyperparameter search.
  • terminal time T = 2.03-12.27 (homophilic); 0.01 (heterophilic)
    Integration horizon treated as network depth; tuned per dataset, not derived from the no-oversmoothing theorem.
  • opinion dimension N_o = 16-256 depending on dataset
    Feature dimension of the opinion space; tuned by search, changes complexity of A^o.
  • integration step size Δt = 1.0 in reported experiments
    Numerical integration step; the gradient bounds assume Δt<<1, but experiments use Δt=1, creating a mismatch.
assumptions (5)
  • domain assumption BIMP trajectories converge to an equilibrium (Hirsch's theorem applied to cooperative irreducible systems)
    Invoked in Lemma 5.5 to analyze oversmoothing via equilibria; the source theorem gives convergence almost everywhere, not necessarily for all initial conditions.
  • ad hoc to paper Learned communication and option graphs A^a, A^o are right stochastic with positive entries
    Softmax attention guarantees row-stochastic matrices; positivity and irreducibility are needed for the monotone-systems argument and for lambda_max=4.
  • ad hoc to paper B = X(0) has pairwise unique entries
    Theorem 5.6 requires unique entries, but the paper does not verify that the encoder output satisfies this for the datasets used.
  • ad hoc to paper First-order expansion of the gradient product ∏(I+ΔtE_t) is valid for large depth
    The non-vanishing gradient proof uses I + ΔtΣE_t + O(Δt^2) as the exact gradient for arbitrary M, which fails when MΔt is not small.
  • standard math Bifurcation analysis by linearizing at xe=0
    Standard linear stability analysis from [35], used to set u=d/(α+3).
invented entities (2)
  • Option graph A^o
    purpose: Couples feature dimensions during message passing, enabling multi-option dissensus and feature mixing.
    A learned adjacency matrix over feature dimensions; it has no external falsifiable prediction and is validated only through the proposed model's performance.
  • Effective adjacency matrix ~A = (A^o+I)⊗(A^a+I)
    purpose: Vectorizes the BIMP dynamics and yields a constant leading eigenvalue lambda_max=4 used to place the attention parameter at the bifurcation point.
    A mathematical construction specific to this paper; its eigenvalue property follows by construction and is not independently testable outside the model.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Resolving Oversmoothing with Opinion Dissensus." pith.science (2026). https://pith.science/paper/67P4XAPR

@misc{pith2026250119089,
  author       = {Pith},
  title        = {Pith review of: Resolving Oversmoothing with Opinion Dissensus},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/67P4XAPR}},
  note         = {Machine review of arXiv:2501.19089}
}
read the original abstract

While graph neural networks (GNNs) have allowed researchers to successfully apply neural networks to non-Euclidean domains, deep GNNs often exhibit lower predictive performance than their shallow counterparts. This phenomena has been attributed in part to oversmoothing, the tendency of node representations to become increasingly similar with network depth. In this paper we introduce an analogy between oversmoothing in GNNs and consensus (i.e., perfect agreement) in the opinion dynamics literature. We show that the message passing algorithms of several GNN models are equivalent to linear opinion dynamics models which have been shown to converge to consensus for all inputs regardless of the graph structure. This new perspective on oversmoothing motivates the use of nonlinear opinion dynamics as an inductive bias in GNN models. In our Behavior-Inspired Message Passing (BIMP) GNN, we leverage the nonlinear opinion dynamics model which is more general than the linear opinion dynamics model, and can be designed to converge to dissensus for general inputs. Through extensive experiments we show that BIMP resists oversmoothing beyond 100 time steps and consistently outperforms existing architectures even when those architectures are amended with oversmoothing mitigation techniques. We also show that BIMP has several desirable properties including well behaved gradients and adaptability to homophilic and heterophilic datasets.

Figures

Figures reproduced from arXiv: 2501.19089 by the authors.

Figure 1
Figure 1. Nonlinear opinion dynamics and dissensus. (Left) The pitchfork bifurcation diagram illustrates a change in the number and stability of opinion states with the attention parameter u (stable equilibria are illustrated with a solid line and unstable equilibria are illustrated with a dotted line). In the diagram, z represents the weighted average of agent opinions, and u ∗ represents the bifurcation point. When the inpu… view at source ↗
Figure 2
Figure 2. Classification accuracy and Dirichlet energy. BIMP is designed to learn node represen￾tations that resist oversmoothing even for very large depths. (Left) We compare the classification accuracy of BIMP to baseline models for architectures with 1, 2, 4, 8, 16, 32, 64 and 128 timesteps. Our BIMP model is stable out to 128 timesteps, while baseline performance deteriorates after 32 timesteps. (Right) We compare the Dir… view at source ↗
Figure 3
Figure 3. Classification accuracy. BIMP is designed to learn node representations that resist oversmoothing even for very large depths. We compare the classification accuracy of BIMP to baseline models for architectures with 1, 2, 4, 8, 16, 32, 64 and 128 timesteps. Our BIMP model and its variants are stable out to 128 timesteps, while baseline performance deteriorates after 32 timesteps. 10 0 10 1 10 2 10 3 Layer Number 10 −… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Dirichlet energy. BIMP is designed to learn node representations that resist oversmoothing even for very large depths. We compare the Dirichlet energy of node features over a range of network depths. The Dirichlet energy of BIMP remains stable even at very deep layers,…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

95 extracted references · 46 canonical work pages

  1. [92]

    Systems of differential equations that are competitive or cooperative ii: Convergence almost everywhere

    Morris W Hirsch. Systems of differential equations that are competitive or cooperative ii: Convergence almost everywhere. SIAM Journal on Mathematical Analysis, 16(3):423–439, 1985

  2. [1]

    Geometry-enhanced molecular representation learning for property prediction

    Xiaomin Fang, Lihang Liu, Jieqiong Lei, Donglong He, Shanzhuo Zhang, Jingbo Zhou, Fan Wang, Hua Wu, and Haifeng Wang. Geometry-enhanced molecular representation learning for property prediction. Nature Machine Intelligence, 4(2):127–134, 2022

  3. [2]

    Starling flock networks manage uncertainty in consensus at low cost

    George F Young, Luca Scardovi, Andrea Cavagna, Irene Giardina, and Naomi E Leonard. Starling flock networks manage uncertainty in consensus at low cost. PLoS computational biology, 9(1):e1002894, 2013

  4. [3]

    A latency-defined edge node placement scheme for opportunistic smart cities

    Oluwashina Madamori, Esther Max-Onakpoya, Gregory D Erhardt, and Corey E Baker. A latency-defined edge node placement scheme for opportunistic smart cities. In 2021 IEEE International Conference on Pervasive Computing and Communications Workshops and other Affiliated Events (PerCom Workshops), pages 142–147. IEEE, 2021

  5. [4]

    Graph convolutional neural networks for web-scale recommender systems

    Rex Ying, Ruining He, Kaifeng Chen, Pong Eksombatchai, William L Hamilton, and Jure Leskovec. Graph convolutional neural networks for web-scale recommender systems. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pages 974–983, 2018

  6. [5]

    Molecular contrastive learning of representations via graph neural networks

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

  7. [6]

    Prediction of protein–protein interaction using graph neural networks

    Kanchan Jha, Sriparna Saha, and Hiteshi Singh. Prediction of protein–protein interaction using graph neural networks. Scientific Reports, 12(1):8360, 2022

  8. [7]

    Deeprank-gnn: a graph neural network framework to learn patterns in protein–protein interfaces

    Manon Réau, Nicolas Renaud, Li C Xue, and Alexandre MJJ Bonvin. Deeprank-gnn: a graph neural network framework to learn patterns in protein–protein interfaces. Bioinformatics, 39(1), 2023

Show all 95 references
  1. [8]

    A graph placement methodology for fast chip design

    Azalia Mirhoseini, Anna Goldie, Mustafa Yazgan, Joe Wenjie Jiang, Ebrahim Songhori, Shen Wang, Young-Joon Lee, Eric Johnson, Omkar Pathak, Azade Nova, et al. A graph placement methodology for fast chip design. Nature, 594(7862):207–212, 2021

  2. [9]

    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. In Proceedings of the AAAI conference on artificial intelligence, 2018

  3. [10]

    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. arXiv preprint arXiv:1905.10947, 2019

  4. [11]

    Revisiting graph neural networks: All we have is low-pass filters

    Hoang Nt and Takanori Maehara. Revisiting graph neural networks: All we have is low-pass filters. arXiv preprint arXiv:1905.09550, 2019

  5. [12]

    Simple and deep graph convolutional networks

    Ming Chen, Zhewei Wei, Zengfeng Huang, Bolin Ding, and Yaliang Li. Simple and deep graph convolutional networks. In International conference on machine learning, pages 1725–1735. PMLR, 2020

  6. [13]

    Towards deeper graph neural networks

    Meng Liu, Hongyang Gao, and Shuiwang Ji. Towards deeper graph neural networks. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 338–348, 2020

  7. [14]

    InInternational conference on machine learning, pages 6878–6917

    Guoji Fu, Peilin Zhao, and Yatao Bian.p-laplacian based graph neural networks. InInternational conference on machine learning, pages 6878–6917. PMLR, 2022

  8. [15]

    Pairnorm: Tackling oversmoothing in gnns

    Lingxiao Zhao and Leman Akoglu. Pairnorm: Tackling oversmoothing in gnns. arXiv preprint arXiv:1909.12223, 2019

  9. [16]

    Towards deeper graph neural networks with differentiable group normalization

    Kaixiong Zhou, Xiao Huang, Yuening Li, Daochen Zha, Rui Chen, and Xia Hu. Towards deeper graph neural networks with differentiable group normalization. Advances in neural information processing systems, 33:4917–4928, 2020

  10. [17]

    Understanding and resolving performance degradation in deep graph convolutional networks

    Kuangqi Zhou, Yanfei Dong, Kaixin Wang, Wee Sun Lee, Bryan Hooi, Huan Xu, and Jiashi Feng. Understanding and resolving performance degradation in deep graph convolutional networks. In Proceedings of the 30th ACM international conference on information & knowledge management, p...

  11. [18]

    Cluster-gcn: An efficient algorithm for training deep and large graph convolutional networks

    Wei-Lin Chiang, Xuanqing Liu, Si Si, Yang Li, Samy Bengio, and Cho-Jui Hsieh. Cluster-gcn: An efficient algorithm for training deep and large graph convolutional networks. In Proceedings 10 of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, p...

  12. [19]

    Mixhop: Higher-order graph convolutional architectures via sparsified neighborhood mixing

    Sami Abu-El-Haija, Bryan Perozzi, Amol Kapoor, Nazanin Alipourfard, Kristina Lerman, Hrayr Harutyunyan, Greg Ver Steeg, and Aram Galstyan. Mixhop: Higher-order graph convolutional architectures via sparsified neighborhood mixing. In international conference on machine learning...

  13. [20]

    Graphsaint: Graph sampling based inductive learning method

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

  14. [21]

    Adaptive universal generalized pagerank graph neural network

    Eli Chien, Jianhao Peng, Pan Li, and Olgica Milenkovic. Adaptive universal generalized pagerank graph neural network. arXiv preprint arXiv:2006.07988, 2020

  15. [22]

    Beltrami flow and neural diffusion on graphs

    Benjamin Chamberlain, James Rowbottom, Davide Eynard, Francesco Di Giovanni, Xiaowen Dong, and Michael Bronstein. Beltrami flow and neural diffusion on graphs. Advances in Neural Information Processing Systems, 34:1594–1609, 2021

  16. [23]

    Clifford group equivariant simplicial message passing networks

    Cong Liu, David Ruhe, Floor Eijkelboom, and Patrick Forré. Clifford group equivariant simplicial message passing networks. arXiv preprint arXiv:2402.10011, 2024

  17. [24]

    A generalized neural diffusion framework on graphs

    Yibo Li, Xiao Wang, Hongrui Liu, and Chuan Shi. A generalized neural diffusion framework on graphs. In Proceedings of the AAAI Conference on Artificial Intelligence, 2024

  18. [25]

    Graph neural ordinary differential equations

    Michael Poli, Stefano Massaroli, Junyoung Park, Atsushi Yamashita, Hajime Asama, and Jinkyoo Park. Graph neural ordinary differential equations. arXiv preprint arXiv:1911.07532, 2019

  19. [26]

    Grand: Graph neural diffusion

    Ben Chamberlain, James Rowbottom, Maria I Gorinova, Michael Bronstein, Stefan Webb, and Emanuele Rossi. Grand: Graph neural diffusion. In International conference on machine learning, pages 1407–1418. PMLR, 2021

  20. [27]

    Pde-gcn: Novel architectures for graph neural networks motivated by partial differential equations

    Moshe Eliasof, Eldad Haber, and Eran Treister. Pde-gcn: Novel architectures for graph neural networks motivated by partial differential equations. Advances in neural information processing systems, 34:3836–3849, 2021

  21. [28]

    Nonlinear dynamics and chaos

    John Michael Tutill Thompson and H Bruce Stewart. Nonlinear dynamics and chaos. John Wiley & Sons, 2002

  22. [29]

    Discovering governing equations from data by sparse identification of nonlinear dynamical systems

    Steven L Brunton, Joshua L Proctor, and J Nathan Kutz. Discovering governing equations from data by sparse identification of nonlinear dynamical systems. Proceedings of the national academy of sciences, 113(15):3932–3937, 2016

  23. [30]

    Output-only identification of self- excited systems using discrete-time lur’e models with application to a gas-turbine combustor

    Juan A Paredes, Yulong Yang, and Dennis S Bernstein. Output-only identification of self- excited systems using discrete-time lur’e models with application to a gas-turbine combustor. International Journal of Control, 97(2):187–212, 2024

  24. [31]

    Output-only identi- fication of lur’e systems with hysteretic feedback nonlinearities

    Riley J Richards, Yulong Yang, Juan A Paredes, and Dennis S Bernstein. Output-only identi- fication of lur’e systems with hysteretic feedback nonlinearities. In 2024 American Control Conference (ACC), pages 2891–2896. IEEE, 2024

  25. [32]

    Graph-coupled oscillator networks

    T Konstantin Rusch, Ben Chamberlain, James Rowbottom, Siddhartha Mishra, and Michael Bronstein. Graph-coupled oscillator networks. In International Conference on Machine Learning, pages 18888–18909. PMLR, 2022

  26. [33]

    Acmp: Allen-cahn message passing for graph neural networks with particle phase transition

    Yuelin Wang, Kai Yi, Xinliang Liu, Yu Guang Wang, and Shi Jin. Acmp: Allen-cahn message passing for graph neural networks with particle phase transition. arXiv preprint arXiv:2206.05437, 2022

  27. [34]

    From continuous dynamics to graph neural networks: Neural diffusion and beyond

    Andi Han, Dai Shi, Lequan Lin, and Junbin Gao. From continuous dynamics to graph neural networks: Neural diffusion and beyond. arXiv preprint arXiv:2310.10121, 2023

  28. [35]

    Fast and flexible multiagent decision-making

    Naomi Ehrich Leonard, Anastasia Bizyaeva, and Alessio Franci. Fast and flexible multiagent decision-making. Annual Review of Control, Robotics, and Autonomous Systems, 7, 2024

  29. [36]

    Nonlinear opinion dynamics with tunable sensitivity

    Anastasia Bizyaeva, Alessio Franci, and Naomi Ehrich Leonard. Nonlinear opinion dynamics with tunable sensitivity. IEEE Transactions on Automatic Control, 68(3):1415–1430, 2022

  30. [37]

    On the number of linear regions of deep neural networks

    Guido F Montufar, Razvan Pascanu, Kyunghyun Cho, and Yoshua Bengio. On the number of linear regions of deep neural networks. Advances in neural information processing systems, 27, 2014. 11

  31. [38]

    Deep learning

    Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature, 521(7553):436–444, 2015

  32. [39]

    A non-asymptotic analysis of oversmoothing in graph neural networks

    Xinyi Wu, Zhengdao Chen, William Wang, and Ali Jadbabaie. A non-asymptotic analysis of oversmoothing in graph neural networks. arXiv preprint arXiv:2212.10701, 2022

  33. [40]

    Semi-supervised classification with graph convolutional networks

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

  34. [41]

    Revisiting over-smoothing in deep gcns

    Chaoqi Yang, Ruijie Wang, Shuochao Yao, Shengzhong Liu, and Tarek Abdelzaher. Revisiting over-smoothing in deep gcns. arXiv preprint arXiv:2003.13663, 2020

  35. [42]

    Graph attention networks

    Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. arXiv preprint arXiv:1710.10903, 2017

  36. [43]

    Demystifying oversmoothing in attention-based graph neural networks

    Xinyi Wu, Amir Ajorlou, Zihui Wu, and Ali Jadbabaie. Demystifying oversmoothing in attention-based graph neural networks. Advances in Neural Information Processing Systems, 36:35084–35106, 2023

  37. [44]

    Understanding convolution on graphs via energies

    Francesco Di Giovanni, James Rowbottom, Benjamin P Chamberlain, Thomas Markovich, and Michael M Bronstein. Understanding convolution on graphs via energies. arXiv preprint arXiv:2206.10991, 2022

  38. [45]

    Preventing representational rank collapse in mpnns by splitting the computational graph

    Andreas Roth, Franka Bause, Nils M Kriege, and Thomas Liebig. Preventing representational rank collapse in mpnns by splitting the computational graph. arXiv preprint arXiv:2409.11504, 2024

  39. [46]

    Not too little, not too much: a theoretical analysis of graph (over) smoothing

    Nicolas Keriven. Not too little, not too much: a theoretical analysis of graph (over) smoothing. Advances in Neural Information Processing Systems, 35:2268–2281, 2022

  40. [47]

    Neural ordinary differential equations

    Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. Advances in neural information processing systems, 31, 2018

  41. [48]

    Grand++: Graph neural diffusion with a source term

    Matthew Thorpe, Tan Nguyen, Hedi Xia, Thomas Strohmer, Andrea Bertozzi, Stanley Osher, and Bao Wang. Grand++: Graph neural diffusion with a source term. ICLR, 2022

  42. [49]

    Gread: Graph neural reaction-diffusion networks

    Jeongwhan Choi, Seoyoung Hong, Noseong Park, and Sung-Bae Cho. Gread: Graph neural reaction-diffusion networks. In International Conference on Machine Learning, pages 5722–

  43. [50]

    From coupled oscillators to graph neural networks: Reducing over-smoothing via a kuramoto model-based approach

    Tuan Nguyen, Hirotada Honda, Takashi Sano, Vinh Nguyen, Shugo Nakamura, and Tan Minh Nguyen. From coupled oscillators to graph neural networks: Reducing over-smoothing via a kuramoto model-based approach. In International Conference on Artificial Intelligence and Statistics, p...

  44. [51]

    Graph neural ricci flow: Evolving feature from a curvature perspective

    Jialong Chen, Bowen Deng, Chuan Chen, Zibin Zheng, et al. Graph neural ricci flow: Evolving feature from a curvature perspective. In The Thirteenth International Conference on Learning Representations, 2025

  45. [52]

    Lectures on network systems, volume 1

    Francesco Bullo. Lectures on network systems, volume 1. CreateSpace, 2018

  46. [53]

    Natural frames and interacting particles in three dimensions

    Eric W Justh and PS Krishnaprasad. Natural frames and interacting particles in three dimensions. In Proceedings of the 44th IEEE Conference on Decision and Control, pages 2841–2846. IEEE, 2005

  47. [54]

    Coordinated control of an underwater glider fleet in an adaptive ocean sampling field experiment in monterey bay

    Naomi E Leonard, Derek A Paley, Russ E Davis, David M Fratantoni, Francois Lekien, and Fumin Zhang. Coordinated control of an underwater glider fleet in an adaptive ocean sampling field experiment in monterey bay. Journal of Field Robotics, 27(6):718–740, 2010

  48. [55]

    Collective motion, sensor networks, and ocean sampling

    Naomi Ehrich Leonard, Derek A Paley, Francois Lekien, Rodolphe Sepulchre, David M Fratantoni, and Russ E Davis. Collective motion, sensor networks, and ocean sampling. Proceedings of the IEEE, 95(1):48–74, 2007

  49. [56]

    Interaction ruling animal collective behavior depends on topological rather than metric distance: Evidence from a field study

    Michele Ballerini, Nicola Cabibbo, Raphael Candelier, Andrea Cavagna, Evaristo Cisbani, Irene Giardina, Vivien Lecomte, Alberto Orlandi, Giorgio Parisi, Andrea Procaccini, et al. Interaction ruling animal collective behavior depends on topological rather than metric distance: ...

  50. [57]

    Odnet: Opinion dynamics-inspired neural message passing for graphs and hypergraphs.Transactions on Machine Learning Research, 2024

    Bingxin Zhou, Outongyi Lv, Jing Wang, Xiang Xiao, and Weishu Zhao. Odnet: Opinion dynamics-inspired neural message passing for graphs and hypergraphs.Transactions on Machine Learning Research, 2024. 12

  51. [58]

    Consensus problems on networks with antagonistic interactions

    Claudio Altafini. Consensus problems on networks with antagonistic interactions. IEEE transactions on automatic control, 58(4):935–946, 2012

  52. [59]

    Biased assimilation, homophily, and the dynamics of polarization

    Pranav Dandekar, Ashish Goel, and David T Lee. Biased assimilation, homophily, and the dynamics of polarization. Proceedings of the National Academy of Sciences, 110(15):5791– 5796, 2013

  53. [60]

    Springer Science & Business Media, 2012

    Martin Golubitsky, Ian Stewart, and David G Schaeffer.Singularities and Groups in Bifurcation Theory: Volume II, volume 69. Springer Science & Business Media, 2012

  54. [61]

    The nonlinear feedback dynamics of asymmetric political polarization

    Naomi Ehrich Leonard, Keena Lipsitz, Anastasia Bizyaeva, Alessio Franci, and Yphtach Lelkes. The nonlinear feedback dynamics of asymmetric political polarization. Proceedings of the National Academy of Sciences, 118(50):e2102149118, 2021

  55. [62]

    Active risk aversion in sis epidemics on networks

    Anastasia Bizyaeva, Marcela Ordorica Arango, Yunxiu Zhou, Simon Levin, and Naomi Ehrich Leonard. Active risk aversion in sis epidemics on networks. In 2024 American Control Conference (ACC), pages 4428–4433. IEEE, 2024

  56. [63]

    Opinion-driven risk perception and reaction in sis epidemics

    Marcela Ordorica Arango, Anastasia Bizyaeva, Simon A Levin, and Naomi Ehrich Leonard. Opinion-driven risk perception and reaction in sis epidemics. arXiv preprint arXiv:2410.12993, 2024

  57. [64]

    Opinion dynamics for decentralized decision-making in a robot swarm

    Marco A Montes de Oca, Eliseo Ferrante, Nithin Mathews, Mauro Birattari, and Marco Dorigo. Opinion dynamics for decentralized decision-making in a robot swarm. In Swarm Intelli- gence: 7th International Conference, ANTS 2010, Brussels, Belgium, September 8-10, 2010. Proceeding...

  58. [65]

    Behavior-inspired neural networks for relational inference

    Yulong Yang, Bowen Feng, Keqin Wang, Naomi Leonard, Adji Bousso Dieng, and Christine Allen-Blanchette. Behavior-inspired neural networks for relational inference. arXiv preprint arXiv:2406.14746, 2024

  59. [66]

    Neural message passing for quantum chemistry

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

  60. [67]

    A survey on oversmoothing in graph neural networks

    T Konstantin Rusch, Michael M Bronstein, and Siddhartha Mishra. A survey on oversmoothing in graph neural networks. arXiv preprint arXiv:2303.10993, 2023

  61. [68]

    Continuous graph neural networks

    Louis-Pascal Xhonneux, Meng Qu, and Jian Tang. Continuous graph neural networks. In International conference on machine learning, pages 10432–10441. PMLR, 2020

  62. [69]

    A note on over-smoothing for graph neural networks

    Chen Cai and Yusu Wang. A note on over-smoothing for graph neural networks. arXiv preprint arXiv:2006.13318, 2020

  63. [70]

    Reaching a consensus

    Morris H DeGroot. Reaching a consensus. Journal of the American Statistical association , 69(345):118–121, 1974

  64. [71]

    L. Moreau. Stability of multiagent systems with time-dependent communication links. IEEE Transactions on Automatic Control, 50(2):169–182, 2005

  65. [72]

    On convergence rate of weighted-averaging dynamics for consensus problems

    Angelia Nedi´c and Ji Liu. On convergence rate of weighted-averaging dynamics for consensus problems. IEEE Transactions on Automatic Control, 62(2):766–781, 2017

  66. [73]

    Information flow and cooperative control of vehicle formations

    J Alexander Fax and Richard M Murray. Information flow and cooperative control of vehicle formations. IEEE transactions on automatic control, 49(9):1465–1476, 2004

  67. [74]

    Convergence in multiagent coordination, consensus, and flocking

    Vincent D Blondel, Julien M Hendrickx, Alex Olshevsky, and John N Tsitsiklis. Convergence in multiagent coordination, consensus, and flocking. In Proceedings of the 44th IEEE Conference on Decision and Control, pages 2996–3000. IEEE, 2005

  68. [75]

    Proactive opinion- driven robot navigation around human movers

    Charlotte Cathcart, María Santos, Shinkyu Park, and Naomi Ehrich Leonard. Proactive opinion- driven robot navigation around human movers. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4052–4058, 2023

  69. [76]

    Simplifying graph convolutional networks

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

  70. [77]

    Dissecting the diffusion process in linear graph convolutional networks

    Yifei Wang, Yisen Wang, Jiansheng Yang, and Zhouchen Lin. Dissecting the diffusion process in linear graph convolutional networks. Advances in Neural Information Processing Systems, 34:5758–5769, 2021. 13

  71. [78]

    Graph neural convection-diffusion with heterophily

    Kai Zhao, Qiyu Kang, Yang Song, Rui She, Sijie Wang, and Wee Peng Tay. Graph neural convection-diffusion with heterophily. arXiv preprint arXiv:2305.16780, 2023

  72. [79]

    Adr-gnn: advection-diffusion-reaction graph neural networks

    Moshe Eliasof, Eldad Haber, and Eran Treister. Adr-gnn: advection-diffusion-reaction graph neural networks. arXiv preprint arXiv:2307.16092, 108, 2023

  73. [80]

    On the difficulty of training recurrent neural networks

    Razvan Pascanu, Tomas Mikolov, and Yoshua Bengio. On the difficulty of training recurrent neural networks. In International conference on machine learning, pages 1310–1318. Pmlr, 2013

  74. [81]

    A convergence analysis of gradient descent on graph neural networks

    Pranjal Awasthi, Abhimanyu Das, and Sreenivas Gollapudi. A convergence analysis of gradient descent on graph neural networks. Advances in Neural Information Processing Systems , 34:20385–20397, 2021

  75. [82]

    On vanishing gradients, over- smoothing, and over-squashing in gnns: Bridging recurrent and graph learning

    Álvaro Arroyo, Alessio Gravina, Benjamin Gutteridge, Federico Barbero, Claudio Gallicchio, Xiaowen Dong, Michael Bronstein, and Pierre Vandergheynst. On vanishing gradients, over- smoothing, and over-squashing in gnns: Bridging recurrent and graph learning. arXiv preprint arXi...

  76. [83]

    Beyond low-frequency information in graph convolutional networks

    Deyu Bo, Xiao Wang, Chuan Shi, and Huawei Shen. Beyond low-frequency information in graph convolutional networks. In Proceedings of the AAAI conference on artificial intelligence, 2021

  77. [84]

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

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

  78. [85]

    Automating the construction of internet portals with machine learning

    Andrew Kachites McCallum, Kamal Nigam, Jason Rennie, and Kristie Seymore. Automating the construction of internet portals with machine learning. Information Retrieval, 3:127–163, 2000

  79. [86]

    Collective classification in network data

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

  80. [87]

    Query-driven active surveying for collective classification

    Galileo Namata, Ben London, Lise Getoor, Bert Huang, and U Edu. Query-driven active surveying for collective classification. In 10th international workshop on mining and learning with graphs, volume 8, page 1, 2012

  81. [88]

    Pitfalls of graph neural network evaluation

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

  82. [89]

    Learning to extract symbolic knowledge from the world wide web.AAAI/IAAI, 3(3.6):2, 1998

    Mark Craven, Dan DiPasquo, Dayne Freitag, Andrew McCallum, Tom Mitchell, Kamal Nigam, and Seán Slattery. Learning to extract symbolic knowledge from the world wide web.AAAI/IAAI, 3(3.6):2, 1998

  83. [90]

    Open graph benchmark: Datasets for machine learning on graphs

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

  84. [91]

    Matrix differential calculus with applications in statistics and econometrics

    Jan R Magnus and Heinz Neudecker. Matrix differential calculus with applications in statistics and econometrics. John Wiley & Sons, 2019

  85. [93]

    Image-based recommendations on styles and substitutes

    Julian McAuley, Christopher Targett, Qinfeng Shi, and Anton Van Den Hengel. Image-based recommendations on styles and substitutes. In Proceedings of the 38th international ACM SIGIR conference on research and development in information retrieval, pages 43–52, 2015

  86. [94]

    Diffusion-convolutional neural networks

    James Atwood and Don Towsley. Diffusion-convolutional neural networks. Advances in neural information processing systems, 29, 2016

  87. [95]

    Tune: A research platform for distributed model selection and training

    Richard Liaw, Eric Liang, Robert Nishihara, Philipp Moritz, Joseph E Gonzalez, and Ion Stoica. Tune: A research platform for distributed model selection and training. arXiv preprint arXiv:1807.05118, 2018. 14 A Proofs A.1 Oversmoothing and Opinion Consensus A.1.1 Lemma 4.1: Li...

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.