Pith. sign in

REVIEW 3 major objections 5 minor 56 references

GraphMinNet: Learning Dependencies in Graphs with Light Complexity Minimal Architecture

T0 review · 3 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read GraphMinNet is a graph neural network that claims to capture long-range dependencies with linear computational complexity and provable stability, reporting the best average rank across ten benchmark datasets.

desk verdict GraphMinNet has a reasonable new architecture and strong empirical results, but its advertised theoretical guarantees—non-decaying gradients, stability, and >1-WL expressiveness—are not actually proven, and the LRD proof rests on an invalid identification of the Laplacian with the normalized adjacency. read the letter →

arxiv 2502.00282 v1 pith:52CF7UDH submitted 2025-02-01 cs.LG

Md Atik Ahamed , Andrew Cheng , Qiang Ye , Qiang Cheng This is my paper · ORCID
classification cs.LG
keywords graphneuralnetworkslong-rangedependenciesminimalGRUpositionalencodinglinearcomplexitypermutationequivarianceexpressivenessWeisfeiler-Lehmantest
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

GraphMinNet is a graph neural network built by transplanting the minimal-GRU update from sequences to graphs. The paper's central claim is that this design keeps gradients between far-away nodes from decaying, so a single lightweight model can learn long-range dependencies without stacking many message-passing layers, and it does so in time linear in the number of nodes. The architecture combines Laplacian positional encodings with feature encodings, and the paper proves permutation equivariance, Lipschitz stability, expressiveness beyond the 1-WL test, and non-decaying long-range gradients. On ten standard datasets it reports the top result on six and the second-best result on three, giving the lowest average rank among all compared methods.

What carries the argument

The load-bearing object is the explicit state-free minGRU expansion, which rewrites the recurrent state as a weighted sum over all past positions, with cumulative gate products $c_i = \prod_{j=1}^i (1-z_j)$ as weights. GraphMinNet replaces sequence positions with graph nodes and converts the element-wise products into matrix encodings $A_u$ (features modulated by learned functions of Laplacian eigenvalues) and $C_u$ (positional eigenvectors modulated by the same functions), then combines them as $a_u = B(A_u \oplus_1 C_u)$ and forms the hidden state by an inner product against the global sum $\bar{a}$. The spectral choice $\phi(\lambda)=\sum_k b_k\lambda^k$ is what makes powers of the normalized adjacency matrix appear in the gradient, so the proof uses shortest-path walk counts to keep gradients bounded away from zero.

What would settle it

Take a small non-regular graph for which direct computation shows that the Laplacian $L$ and the normalized adjacency matrix $\tilde{A}$ do not commute, construct GraphMinNet exactly as in the proof, and compute the gradient norm $\|\partial h_u/\partial x_v\|$ for pairs of nodes with increasing shortest-path distance; if the spectral identification fails, the lower bound $b_k\gamma>0$ no longer follows and the gradient should decay with distance. Alternatively, measure the empirical gradient norm on the Peptides-func task on such a graph and check whether it stays bounded away from zero as the node distance grows.

Watch

Extended reading notes

Core claim

The central discovery is a recurrence-free, state-free expression for the minimal GRU hidden state, $h_t = \sum_{i=1}^t (\prod_{j=i+1}^t (1-z_j)) \odot z_i \odot \tilde{h}_i$, which lets sequence positions be replaced by all graph nodes. GraphMinNet encodes each node through a feature matrix built from Laplacian eigenvalues and a positional matrix built from Laplacian eigenvectors, combines them into a shared embedding, and computes $h_u = \langle a_u, \bar{a}\rangle \odot z_u \odot \tilde{h}_u$. The paper shows that for a particular spectral function $\phi(\lambda)=\sum_k b_k \lambda^k$ with $b_k>0$, the gradient $\partial h_u/\partial x_v$ contains a term proportional to $(\tilde{A}^{\,k})_{u,v}$ with $k=\mathrm{spd}(u,v)$, which is positive because it counts walks, so the gradient norm is bounded below independently of distance. Together with the $O(nmdl)$ complexity estimate, this yields a GNN with provable long-range propagation, permutation equivariance, stability, and expressiveness strictly above 1-WL and no higher than 3-WL.

Load-bearing premise

The load-bearing premise of the long-range-gradient proof is that the graph Laplacian and the normalized adjacency matrix can be interchanged inside the learned spectral function $\phi$, so that walk counts of $\tilde{A}$ appear in the gradient; that interchange is not generally valid on arbitrary graphs, and the proof also assumes without a bound that the Laplacian positional eigenvectors are stable under graph perturbations.

Editorial extensions

If this is right

  • A single GraphMinNet layer can propagate information across long distances without vanishing gradients, so deep message-passing stacks are not required for long-range tasks.
  • After a one-time eigendecomposition, training and inference cost $O(nmdl)$ in the number of nodes, making the method applicable to graphs with tens of thousands of nodes with linearly growing memory.
  • Permutation equivariance and Lipschitz stability mean the model's output does not depend on node ordering and is provably robust to small perturbations of features and eigenvalues.
  • The architecture is strictly more expressive than the 1-WL test and no more expressive than the 3-WL test, placing its discrimination power in a known part of the Weisfeiler-Lehman hierarchy.
  • The reported experiments cover diverse graph types, including molecular graphs, superpixel image graphs, synthetic community graphs, and Android function-call graphs, with top results on 6 of 10 datasets.

Reading between the lines

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

  • If the spectral identification used in the long-range proof can be justified or extended, the same walk-counting mechanism would suggest that any position-aware GNN whose gradient expands into powers of a normalized adjacency matrix can inherit a non-decaying long-range property, not just GraphMinNet.
  • A natural testable extension is to treat the spectral function $\phi$ as a learnable diffusion kernel: varying the weights $b_k$ interpolates between local and global mixing, and this could be probed on long-range benchmark tasks without changing the model's linear complexity.
  • The paper establishes eigenvalue stability but does not bound changes in the Laplacian eigenvectors themselves; an experiment measuring robustness on graphs with repeated or nearly repeated Laplacian eigenvalues would directly test whether the stability guarantee holds in practice.
  • The paper states that edge features are not explicitly accounted for; incorporating edge features into the gating or inner-product terms is a direct extension that the current formulation does not support.
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

3 major / 5 minor

Summary. The paper introduces GraphMinNet, a GNN that adapts the minGRU recurrent cell to graphs by replacing sequence positions with graph nodes and combining feature encodings with Laplacian positional encodings. The authors claim that GraphMinNet is permutation equivariant, Lipschitz stable, achieves non-decaying gradient norms over long shortest-path distances, has O(nmdl) complexity, and is strictly more expressive than 1-WL but not more than 3-WL. These properties are stated in Section 3.3 and proved in Appendix 7. The empirical section compares against nine baselines on ten datasets, reporting the best average rank and state-of-the-art results on six datasets, plus scaling and robustness experiments.

Significance. The model is simple, efficient, and achieves competitive empirical results, which would be valuable if the theoretical guarantees were established. The linear complexity and the non-decaying-gradient property would distinguish GraphMinNet from quadratic attention models and from message-passing GNNs that suffer from over-squashing. However, the central theoretical results currently have serious gaps: the long-range-dependency proof relies on an invalid identification of the graph Laplacian with the normalized adjacency matrix, the stability proof omits eigenvector perturbation bounds, and the expressiveness lower bound is asserted rather than demonstrated. Since these are the paper's headline guarantees, the contribution is not yet established.

major comments (3)
  1. [Appendix 7.2, Proposition 7.2 (proof of Proposition 3.5)] The proof sets 'Let ϕ1^2(L) = ϕ(Ã)' after defining L as the combinatorial graph Laplacian and à = D^{-1/2}AD^{-1/2} as the normalized adjacency matrix. These matrices do not generally share eigenvectors, so a polynomial in L cannot equal a polynomial in à for arbitrary graphs. The subsequent lower bound relies on (Ã^k)_{u,v} > 0 for k = spd(u,v), which is only available through this invalid identification. Therefore the non-decaying gradient claim of Proposition 3.5 is not established for general graphs. This is a load-bearing step because LRD is a central advertised contribution.
  2. [Appendix 7.1, Proposition 7.1 (stability)] The stability proof bounds eigenvalue perturbations using Weyl's inequality, but the positional component C_u in Eq. (5) depends on the eigenvectors p_u. The proof bounds ∥ΔA_u∥ but does not bound ∥Δp_u∥ under graph perturbation. Consequently, the claimed Lipschitz stability with respect to eigenvalues, and the generalization result of Corollary 3.4 that relies on it, is incomplete.
  3. [Appendix 7.4, Proposition 7.4 (expressiveness)] The strict improvement over 1-WL is asserted rather than proven; no concrete pair of non-isomorphic graphs that are indistinguishable by 1-WL but distinguishable by GraphMinNet is given. The derivation also contains a typo (⟨Au ⊙ Cu, Au ⊙ Cu⟩ should be ⟨Au⊙Cu, Av⊙Cv⟩), and with the truncated positional encoding ⟨Cu,Cv⟩ equals the Laplacian entry only when d=n. The upper-bound citation to (Zhang et al., 2024) is not justified in the text. Thus the claimed 'between 1-WL and 3-WL' expressiveness is not proven.
minor comments (5)
  1. [Section 3.2, Eq. (11)] The sentence after Eq. (11) says 'This formulation generalizes Eq. (7), which can be recovered as a special case when β = 1.' This is incorrect because with β = 1 the self-term (2−β)⟨Ws1Cu, Ws2Cu⟩1l does not vanish; the correct value is β = 2 (or the sentence should be reworded).
  2. [Appendix 7.2, Proposition 7.2] The claim that the proof 'holds for other types by appropriately choosing B' is not demonstrated; please provide the construction for each inner-product type or state explicitly that only Type 4 is used.
  3. [Section 4, Table 1] The 'Avg. Rank' column should be accompanied by a description of how ranks are computed, including how ties and missing entries are handled.
  4. [Section 4, Figure 3] The x-axis tick labels (777, 5814, 12059, 15129, 17536) do not match the text's stated range of '1,000 to 20,000'; please align the description with the actual experiment.
  5. [Section 5.1, Table 3] Please clarify which configuration (with or without self-term) is used for each dataset in Table 1, since the ablation shows dataset-dependent defaults.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found; the LRD proof's key equality is a correctness gap, not a self-referential reduction.

full rationale

GraphMinNet's central empirical claims are measured on predefined test splits against external baselines (Table 1), and its expressiveness bounds are imported from independent sources (Xu et al. 2018 for the 1-WL equivalence; Zhang et al. 2024 for the 3-WL upper bound), so the main results are not fitted values renamed as predictions. The only self-citations (Ahamed & Cheng 2024; Wu & Cheng 2022) are contextual remarks and carry no logical weight in the proofs. The heaviest theoretical step, Proposition 7.2, contains the assertion 'Let φ_1^2(L) = φ(Ã)', which is not generally valid because L and the normalized adjacency matrix need not share eigenvectors; this is a genuine correctness gap, but the proposition is existential and the proof attempts a construction of φ, so it is incomplete rather than a definitional equivalence. Likewise, the spectral-stability proof bounds eigenvalue changes but not eigenvector changes; that is an omitted control, not a circular reduction. Accordingly, no step in the paper's derivation chain is equivalent to its own input by construction.

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

The central theoretical claims rely on several unstated or ad hoc premises. The LRD proof relies on identifying the Laplacian with the normalized adjacency, the stability proof implicitly assumes eigenvector Lipschitzness, the 3-WL upper bound is imported from a cited theorem, and the Lipschitzness of the permutation equivariant functions is assumed for typical choices. The learnable parameters W, B, and the β weight are standard trained parameters; the proof constants b_k and the per-dataset hyperparameters are ad hoc choices needed to reproduce the results.

free parameters (3)
  • b_k (positive constants in LRD proof)
    Introduced ad hoc in Appendix 7.2 to make the gradient lower bound positive; the proof does not specify values or a construction independent of the graph.
  • β (self-term weight, Eq. 11) = learned in [0,2]
    A learnable scalar controlling the self-loop contribution; tuned per dataset during training.
  • Hyperparameters (hidden dim, layers, Lap dim, dropout rates) = per-dataset values in Table 9
    Hand-chosen per dataset, as is standard, but essential to reproduce the reported results.
assumptions (4)
  • ad hoc to paper The graph Laplacian L and normalized adjacency \tilde A can be identified in the LRD proof.
    The proof sets ϕ_1^2(L)=ϕ(\tilde A), requiring L and \tilde A to share eigenvectors, which holds only for special graphs; this is not stated as an assumption.
  • domain assumption Laplacian positional eigenvectors p_u are Lipschitz continuous under graph perturbations.
    The stability proof bounds ΔA_u but not ΔC_u, so it implicitly assumes eigenvector stability; eigenvectors can change discontinuously under perturbations, so this is an unproven premise.
  • standard math The basis-invariant GNN bound from Zhang et al. (2024) applies to GraphMinNet.
    Used to cap expressiveness at 3-WL; accepted as a cited theorem, not re-derived.
  • standard math Permutation equivariant functions ϕ_i are Lipschitz.
    Invoked to obtain stability; plausible for typical choices but stated rather than proven in the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GraphMinNet: Learning Dependencies in Graphs with Light Complexity Minimal Architecture." pith.science (2026). https://pith.science/paper/52CF7UDH

@misc{pith2026250200282,
  author       = {Pith},
  title        = {Pith review of: GraphMinNet: Learning Dependencies in Graphs with Light Complexity Minimal Architecture},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/52CF7UDH}},
  note         = {Machine review of arXiv:2502.00282}
}
read the original abstract

Graph Neural Networks (GNNs) have demonstrated remarkable success in various applications, yet they often struggle to capture long-range dependencies (LRD) effectively. This paper introduces GraphMinNet, a novel GNN architecture that generalizes the idea of minimal Gated Recurrent Units to graph-structured data. Our approach achieves efficient LRD modeling with linear computational complexity while maintaining permutation equivariance and stability. The model incorporates both structural and positional information through a unique combination of feature and positional encodings, leading to provably stronger expressiveness than the 1-WL test. Theoretical analysis establishes that GraphMinNet maintains non-decaying gradients over long distances, ensuring effective long-range information propagation. Extensive experiments on ten diverse datasets, including molecular graphs, image graphs, and synthetic networks, demonstrate that GraphMinNet achieves state-of-the-art performance while being computationally efficient. Our results show superior performance on 6 out of 10 datasets and competitive results on the others, validating the effectiveness of our approach in capturing both local and global graph structures.

Figures

Figures reproduced from arXiv: 2502.00282 by the authors.

Figure 1
Figure 1. Schematic diagram of our proposed method. Here [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Run time comparison per epochs including train, [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Scalability analysis of GraphMinNet. (a) shows [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Robustness analysis of our model under varying [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Effectiveness of dropouts. 5.3. Effectiveness of Dropout Regularization We analyze the impact of dropout regularization on model performance [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

56 extracted references · 25 canonical work pages

  1. [1]

    Slic superpixels compared to state-of-the-art superpixel methods

    Achanta, R., Shaji, A., Smith, K., Lucchi, A., Fua, P., and S \"u sstrunk, S. Slic superpixels compared to state-of-the-art superpixel methods. IEEE transactions on pattern analysis and machine intelligence, 34 0 (11): 0 2274--2282, 2012

    work page 2012

  2. [2]

    Ahamed, M. A. and Cheng, Q. Timemachine: A time series is worth 4 mambas for long-term forecasting. arXiv preprint arXiv:2403.09898, 2024

    arXiv 2024

  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, 2022

    work page 2022

  4. [4]

    and Elisseeff, A

    Bousquet, O. and Elisseeff, A. Stability and generalization. The Journal of Machine Learning Research, 2: 0 499--526, 2002

    2002

  5. [5]

    and Laurent, T

    Bresson, X. and Laurent, T. Residual gated graph convnets. arXiv preprint arXiv:1711.07553, 2017

    arXiv 2017

  6. [6]

    I., Bronstein, M., Webb, S., and Rossi, E

    Chamberlain, B., Rowbottom, J., Gorinova, M. I., Bronstein, M., Webb, S., and Rossi, E. Grand: Graph neural diffusion. In International conference on machine learning, pp.\ 1407--1418. PMLR, 2021

    work page 2021

  7. [7]

    Principal neighbourhood aggregation for graph nets

    Corso, G., Cavalleri, L., Beaini, D., Li \`o , P., and Veli c kovi \'c , P. Principal neighbourhood aggregation for graph nets. Advances in Neural Information Processing Systems, 33: 0 13260--13271, 2020

    work page 2020

  8. [8]

    Recurrent distance filtering for graph representation learning

    Ding, Y., Orvieto, A., He, B., and Hofmann, T. Recurrent distance filtering for graph representation learning. In Forty-first International Conference on Machine Learning, 2024

    work page 2024

Show all 56 references
  1. [9]

    S., Hou, K., Salakhutdinov, R

    Du, S. S., Hou, K., Salakhutdinov, R. R., Poczos, B., Wang, R., and Xu, K. Graph neural tangent kernel: Fusing graph neural networks with graph kernels. Advances in neural information processing systems, 32, 2019

    2019

  2. [10]

    P., Ramp \'a s ek, L., Galkin, M., Parviz, A., Wolf, G., Luu, A

    Dwivedi, V. P., Ramp \'a s ek, L., Galkin, M., Parviz, A., Wolf, G., Luu, A. T., and Beaini, D. Long range graph benchmark. Advances in Neural Information Processing Systems, 35: 0 22326--22340, 2022

    2022

  3. [11]

    P., Joshi, C

    Dwivedi, V. P., Joshi, C. K., Luu, A. T., Laurent, T., Bengio, Y., and Bresson, X. Benchmarking graph neural networks. Journal of Machine Learning Research, 24 0 (43): 0 1--48, 2023

    2023

  4. [12]

    K., Winn, J., and Zisserman, A

    Everingham, M., Van Gool, L., Williams, C. K., Winn, J., and Zisserman, A. The pascal visual object classes (voc) challenge. International journal of computer vision, 88: 0 303--338, 2010

    2010

  5. [13]

    O., Bengio, Y., and Hajimirsadegh, H

    Feng, L., Tung, F., Ahmed, M. O., Bengio, Y., and Hajimirsadegh, H. Were rnns all we needed? arXiv preprint arXiv:2410.01201, 2024

    2024 arXiv

  6. [14]

    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

  7. [15]

    Understanding and extending subgraph gnns by rethinking their symmetries

    Frasca, F., Bevilacqua, B., Bronstein, M., and Maron, H. Understanding and extending subgraph gnns by rethinking their symmetries. Advances in Neural Information Processing Systems, 35: 0 31376--31390, 2022

    2022

  8. [16]

    Freitas, S., Dong, Y., Neil, J., and Chau, D. H. A large-scale database for graph representation learning. arXiv preprint arXiv:2011.07682, 2020

    2011 arXiv

  9. [17]

    and Leskovec, J

    Grover, A. and Leskovec, J. node2vec: Scalable feature learning for networks. In Proceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining, pp.\ 855--864, 2016

    2016

  10. [18]

    and Dao, T

    Gu, A. and Dao, T. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023

    2023 arXiv

  11. [19]

    Hippo: Recurrent memory with optimal polynomial projections

    Gu, A., Dao, T., Ermon, S., Rudra, A., and R \'e , C. Hippo: Recurrent memory with optimal polynomial projections. Advances in neural information processing systems, 33: 0 1474--1487, 2020

    2020

  12. [20]

    Combining recurrent, convolutional, and continuous-time models with linear state space layers

    Gu, A., Johnson, I., Goel, K., Saab, K., Dao, T., Rudra, A., and R \'e , C. Combining recurrent, convolutional, and continuous-time models with linear state space layers. Advances in neural information processing systems, 34: 0 572--585, 2021

    2021

  13. [21]

    Inductive representation learning on large graphs

    Hamilton, W., Ying, Z., and Leskovec, J. Inductive representation learning on large graphs. Advances in neural information processing systems, 30, 2017

    2017

  14. [22]

    Open graph benchmark: Datasets for machine learning on graphs

    Hu, W., Fey, M., Zitnik, M., Dong, Y., Ren, H., Liu, B., Catasta, M., and Leskovec, J. Open graph benchmark: Datasets for machine learning on graphs. Advances in neural information processing systems, 33: 0 22118--22133, 2020

    2020

  15. [23]

    Boosting the cycle counting power of graph neural networks with i ^ 2 -gnns

    Huang, Y., Peng, X., Ma, J., and Zhang, M. Boosting the cycle counting power of graph neural networks with i ^ 2 -gnns. In The Eleventh International Conference on Learning Representations, 2022

    2022

  16. [24]

    On the stability of expressive positional encodings for graph neural networks

    Huang, Y., Lu, W., Robinson, J., Yang, Y., Zhang, M., Jegelka, S., and Li, P. On the stability of expressive positional encodings for graph neural networks. arXiv preprint arXiv:2310.02579, 2023

    2023 arXiv

  17. [25]

    On the stability of expressive positional encodings for graph neural networks

    Huang, Y., Lu, W., Robinson, J., Yang, Y., Zhang, M., Jegelka, S., and Li, P. On the stability of expressive positional encodings for graph neural networks. In The Twelfth International Conference on Learning Representations, 2024 a

    2024

  18. [26]

    What can we learn from state space models for machine learning on graphs? arXiv preprint arXiv:2406.05815, 2024 b

    Huang, Y., Miao, S., and Li, P. What can we learn from state space models for machine learning on graphs? arXiv preprint arXiv:2406.05815, 2024 b

    2024 arXiv

  19. [27]

    Neural tangent kernel: Convergence and generalization in neural networks

    Jacot, A., Gabriel, F., and Hongler, C. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018

    2018

  20. [28]

    R., Savarese, S., and Saxena, A

    Jain, A., Zamir, A. R., Savarese, S., and Saxena, A. Structural-rnn: Deep learning on spatio-temporal graphs. In Proceedings of the ieee conference on computer vision and pattern recognition, pp.\ 5308--5317, 2016

    2016

  21. [29]

    Kipf, T. N. and Welling, M. Semi-supervised classification with graph convolutional networks. In International Conference on Learning Representations, 2016

    2016

  22. [30]

    Rethinking graph transformers with spectral attention

    Kreuzer, D., Beaini, D., Hamilton, W., L \'e tourneau, V., and Tossou, P. Rethinking graph transformers with spectral attention. Advances in Neural Information Processing Systems, 34: 0 21618--21629, 2021

    2021

  23. [31]

    Distance encoding: Design provably more powerful neural networks for graph representation learning

    Li, P., Wang, Y., Wang, H., and Leskovec, J. Distance encoding: Design provably more powerful neural networks for graph representation learning. Advances in Neural Information Processing Systems, 33: 0 4465--4478, 2020

    2020

  24. [32]

    D., Zhao, L., Smidt, T., Sra, S., Maron, H., and Jegelka, S

    Lim, D., Robinson, J. D., Zhao, L., Smidt, T., Sra, S., Maron, H., and Jegelka, S. Sign and basis invariant networks for spectral graph representation learning. In The Eleventh International Conference on Learning Representations, 2022

    2022

  25. [33]

    K., Coates, M., Torr, P., and Lim, S.-N

    Ma, L., Lin, C., Lim, D., Romero-Soriano, A., Dokania, P. K., Coates, M., Torr, P., and Lim, S.-N. Graph inductive biases in transformers without message passing. In International Conference on Machine Learning, pp.\ 23321--23337. PMLR, 2023

    2023

  26. [34]

    L., Lenssen, J

    Morris, C., Ritzert, M., Fey, M., Hamilton, W. L., Lenssen, J. E., Rattan, G., and Grohe, M. Weisfeiler and leman go neural: Higher-order graph neural networks. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pp.\ 4602--4609, 2019

    2019

  27. [35]

    Deepwalk: Online learning of social representations

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

    2014

  28. [36]

    Graph neural ordinary differential equations

    Poli, M., Massaroli, S., Park, J., Yamashita, A., Asama, H., and Park, J. Graph neural ordinary differential equations. arXiv preprint arXiv:1911.07532, 2019

    1911 arXiv

  29. [37]

    P., Luu, A

    Ramp \'a s ek, L., Galkin, M., Dwivedi, V. P., Luu, A. T., Wolf, G., and Beaini, D. Recipe for a general, powerful, scalable graph transformer. Advances in Neural Information Processing Systems, 35: 0 14501--14515, 2022

    2022

  30. [38]

    Learnability, stability and uniform convergence

    Shalev-Shwartz, S., Shamir, O., Srebro, N., and Sridharan, K. Learnability, stability and uniform convergence. The Journal of Machine Learning Research, 11: 0 2635--2670, 2010

    2010

  31. [39]

    J., and Sinop, A

    Shirzad, H., Velingker, A., Venkatachalam, B., Sutherland, D. J., and Sinop, A. K. Exphormer: Sparse transformers for graphs. In International Conference on Machine Learning, pp.\ 31613--31632. PMLR, 2023

    2023

  32. [40]

    K., Bhalla, S., Usmani, S

    Singh, S., Chaudhary, K., Dhanda, S. K., Bhalla, S., Usmani, S. S., Gautam, A., Tuknait, A., Agrawal, P., Mathur, D., and Raghava, G. P. Satpdb: a database of structurally annotated therapeutic peptides. Nucleic acids research, 44 0 (D1): 0 D1119--D1126, 2016

    2016

  33. [41]

    Graph attention networks

    Veli c kovi \'c , P., Cucurull, G., Casanova, A., Romero, A., Li \`o , P., and Bengio, Y. Graph attention networks. In International Conference on Learning Representations, 2018

    2018

  34. [42]

    Equivariant and stable positional encoding for more powerful graph neural networks

    Wang, H., Yin, H., Zhang, M., and Li, P. Equivariant and stable positional encoding for more powerful graph neural networks. In International Conference on Learning Representations, 2022

    2022

  35. [43]

    State space model for new-generation network alternative to transformers: A survey

    Wang, X., Wang, S., Ding, Y., Li, Y., Wu, W., Rong, Y., Kong, W., Huang, J., Li, S., Yang, H., et al. State space model for new-generation network alternative to transformers: A survey. arXiv preprint arXiv:2404.09516, 2024

    2024 arXiv

  36. [44]

    and Cheng, Q

    Wu, X. and Cheng, Q. Stabilizing and enhancing link prediction through deepened graph auto-encoders. In IJCAI: proceedings of the conference, volume 2022, pp.\ 3587. NIH Public Access, 2022

    2022

  37. [45]

    N., Gomes, J., Geniesse, C., Pappu, A

    Wu, Z., Ramsundar, B., Feinberg, E. N., Gomes, J., Geniesse, C., Pappu, A. S., Leswing, K., and Pande, V. Moleculenet: a benchmark for molecular machine learning. Chemical science, 9 0 (2): 0 513--530, 2018

    2018

  38. [46]

    Continuous graph neural networks

    Xhonneux, L.-P., Qu, M., and Tang, J. Continuous graph neural networks. In International conference on machine learning, pp.\ 10432--10441. PMLR, 2020

    2020

  39. [47]

    How powerful are graph neural networks? In International Conference on Learning Representations, 2018

    Xu, K., Hu, W., Leskovec, J., and Jegelka, S. How powerful are graph neural networks? In International Conference on Learning Representations, 2018

    2018

  40. [48]

    Do transformers really perform badly for graph representation? Advances in neural information processing systems, 34: 0 28877--28888, 2021

    Ying, C., Cai, T., Luo, S., Zheng, S., Ke, G., He, D., Shen, Y., and Liu, T.-Y. Do transformers really perform badly for graph representation? Advances in neural information processing systems, 34: 0 28877--28888, 2021

    2021

  41. [49]

    Hierarchical graph representation learning with differentiable pooling

    Ying, Z., You, J., Morris, C., Ren, X., Hamilton, W., and Leskovec, J. Hierarchical graph representation learning with differentiable pooling. Advances in neural information processing systems, 31, 2018

    2018

  42. [50]

    Position-aware graph neural networks

    You, J., Ying, R., and Leskovec, J. Position-aware graph neural networks. In International conference on machine learning, pp.\ 7134--7143. PMLR, 2019

    2019

  43. [51]

    M., Ying, R., and Leskovec, J

    You, J., Gomes-Selman, J. M., Ying, R., and Leskovec, J. Identity-aware graph neural networks. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pp.\ 10737--10745, 2021

    2021

  44. [52]

    On the expressive power of spectral invariant graph neural networks

    Zhang, B., Zhao, L., and Maron, H. On the expressive power of spectral invariant graph neural networks. arXiv preprint arXiv:2406.04336, 2024

    2024 arXiv

  45. [53]

    and Li, P

    Zhang, M. and Li, P. Nested graph neural networks. Advances in Neural Information Processing Systems, 34: 0 15734--15747, 2021

    2021

  46. [54]

    From stars to subgraphs: Uplifting any gnn with local structure awareness

    Zhao, L., Jin, W., Akoglu, L., and Shah, N. From stars to subgraphs: Uplifting any gnn with local structure awareness. In International Conference on Learning Representations, 2021

    2021

  47. [55]

    C., and Dong, X

    Zhi, Y.-C., Ng, Y. C., and Dong, X. Gaussian processes on graphs via spectral kernel learning. IEEE Transactions on Signal and Information Processing over Networks, 9: 0 304--314, 2023

    2023

  48. [56]

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