Pith. sign in

REVIEW 4 major objections 6 minor 74 references

Using Random Noise Equivariantly to Boost Graph Neural Networks Universally

T0 review · 4 major / 6 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 The paper argues that random noise can make graph neural networks more expressive without hurting generalization, provided the network is invariant to permutations of noise channels.

desk verdict A promising noise-equivariant GNN architecture with strong empirical gains, but the theoretical expressivity and sample-complexity claims are not yet proven. read the letter →

arxiv 2502.02479 v1 pith:4DNPPLJF submitted 2025-02-04 cs.LG

Xiyuan Wang , Muhan Zhang This is my paper · ORCID
classification cs.LG
keywords graphneuralnetworksrandomnoisefeaturesequivariancechannelpermutationsamplecomplexityuniversalexpressivitymessagepassinggeneralization
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 takes on a puzzle in graph learning: adding random noise to node features can in principle make message-passing networks as expressive as any graph function, but in practice it usually hurts. The authors trace the harm to sample complexity—a naive network sees the same graph with different noise as different inputs and must relearn it many times—and show that enforcing symmetry on the noise removes most of that cost. Their proposed architecture, ENGNN, keeps a channel-permutation-equivariant representation of noise through message passing and pooling, so it keeps the expressivity gain while generalizing. On node, link, subgraph, and graph-level benchmarks, ENGNN consistently beats plain MPNN and noise-fed MPNN, and it matches task-specific noise-based link predictors. If the claim holds, random noise becomes a cheap, task-agnostic expressivity booster usable across graph tasks.

What carries the argument

The load-bearing object is the channel-permutation-equivariant aggregator AGGR, which processes pairs of invariant node features $X\in\mathbb{R}^{k\times d}$ and equivariant noise features $Z\in\mathbb{R}^{k\times L\times C}$. Step 1 gives each noise channel a stable identifier by applying a node-permutation-invariant DeepSet $\psi$ to each channel's entries; step 2 encodes each node's noise with a channel-permutation-invariant DeepSet $\phi$; step 3 aggregates the node set with DeepSet $\varphi$; step 4 uses MLPs $g,h$ to output updated invariant and equivariant features. This construction is equivariant to node and channel permutations, universal under mild conditions, and runs in $\Theta(k)$ per aggregator and $O(|E|)$ per message-passing layer. The generalization theory is carried by the quotient pseudo-metric $\rho_{Z,T}(Z_1,Z_2)=\inf_{t_1,t_2\in T}\rho_Z(t_1(Z_1),t_2(Z_2))$, which makes the effective noise space smaller when $T$ is the group of channel permutations.

What would settle it

Take a pair of non-isomorphic graphs that the standard color-refinement test cannot tell apart, give every node the same features and fresh random noise, and train a single-layer ENGNN to classify which graph is which; if it cannot reach perfect separation under the paper's noise conditions, the claimed 'for all non-isomorphic graphs' separation fails.

Watch

Extended reading notes

Core claim

The central claim is that random noise is not inherently bad for graph neural networks; bad design is. Ordinary MPNNs treat each noise realization as a distinct input, so covering the noise space multiplies the number of examples needed. The paper proves a PAC bound in which the sample-complexity term scales with the covering number of the noise space under a quotient metric, and making the hypothesis class invariant to channel permutations shrinks that covering number by a factor of $1/C!$ for noise in $[0,1]^{n\times C}$. The proposed ENGNN realizes this symmetry with an equivariant aggregator built from DeepSet modules, and the authors prove it is still universal: for any non-isomorphic graphs there is a parameterization under which the outputs differ for all noise realizations, while the expectation over noise agrees on isomorphic pairs. Experiments across node, link, subgraph, and graph tasks show consistent gains over both MPNN and noise-fed MPNN, with performance comparable to task-specific noise methods.

Load-bearing premise

The proof that ENGNN can separate all non-isomorphic graphs assumes that after one message-passing layer and pooling, the representation pins down the full graph—nodes, edges, and subset—for every possible noise sample, an injectivity property the paper asserts rather than derives from the approximation theorems.

Editorial extensions

If this is right

  • If correct, any graph task can use random noise without paying the usual generalization penalty, and ENGNN should keep outperforming plain MPNNs across node, link, subgraph, and graph-level benchmarks.
  • The $1/C!$ covering-number reduction gives a quantitative reason that channel-permutation symmetry is the right inductive bias for noise features.
  • Universal expressivity, in the sense of separating all non-isomorphic graphs given unique per-node noise, is achieved while keeping the same asymptotic $O(|E|)$ complexity as ordinary message passing.
  • Prior noise-based link prediction methods, whose invariances to orthogonal or rotation groups are special cases, fit inside this framework and can be replaced by the more flexible ENGNN design.
  • On saturated molecule benchmarks ENGNN does not beat specialized models, but the paper's ablation shows it still improves over plain MPNN and noise-fed MPNN with substantially less memory and time.

Reading between the lines

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

  • A testable extension is to compare ENGNN with naive noise-feeding as the number of noise channels $C$ grows: the theory predicts the generalization gap should widen roughly by a factor of $C!$, and that trend could be measured directly.
  • The symmetry group could be enlarged beyond channel permutations to, say, orthogonal transformations, which would shrink the effective noise space further but would require higher-order representations; whether an intermediate group gives a practical trade-off is left open by the paper.
  • Because the separation result relies on fresh unique noise at each evaluation, test-time predictions are stochastic; averaging over several noise samples would reduce variance but adds an overhead the paper does not quantify.
  • The injectivity assumption behind the universality proof could be probed empirically with finite-width models on adversarial non-isomorphic graph pairs that are hard to separate by color refinement.
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 / 6 minor

Summary. The paper proposes a theoretical and architectural framework for using random noise as an auxiliary input in graph neural networks. It first gives PAC-style sample-complexity bounds for GNNs with noise inputs and argues that enforcing invariance of the model to noise-channel permutations reduces the covering number of the noise space. It then introduces ENGNN, a message-passing architecture whose aggregator is equivariant to node permutations and to noise-channel permutations, and states theorems claiming that ENGNN preserves universal expressivity for graph, node, link, and subgraph tasks. The experimental section reports comparisons on node classification, link prediction, subgraph classification, substructure counting, and molecule property prediction, together with an ablation against MPNN and a noise-fed MPNN (NMPNN).

Significance. If fully established, the paper's central principle would be valuable: it would show a simple, task-agnostic way to keep the expressivity benefit of random noise while mitigating the generalization penalty, and the proposed architecture is efficient and broadly applicable. The empirical ablation is genuinely informative: NMPNN degrades sharply relative to MPNN, while ENGNN consistently outperforms both, which provides direct evidence that equivariant treatment of noise helps. The paper also ships substantial experiments across many task families. However, the significance is currently conditional because the two load-bearing theoretical pillars—the exact covering-number ratio in Proposition 3.3 and the universal-expressivity separation in Theorem 4.3—are not rigorously established. In particular, the expressivity proof relies on an injectivity assertion that goes beyond the stated universal-approximation result, and the quantitative sample-complexity claim appears to be an inequality rather than the exact equality stated.

major comments (4)
  1. [Section 3, Proposition 3.3] The claim that N(Z, ρ_Z,T, r) / N(Z, ρ_Z, r) = 1/C! for the channel-permutation quotient of [0,1]^{n×C} is stated without proof and is not correct as an exact equality for a general finite covering radius r. For a free action of a finite group of isometries, the covering number of the quotient is only bounded between N_orig/C! and N_orig; the upper bound can be attained when a single ball covers the space, in which case the ratio is 1. The paper needs either a proof of the equality under specific choices of ρ_Z and r, or a corrected statement such as N_quot ≥ N_orig/C! with matching upper bounds. As written, Eq. (5) is a quantitative pillar of the sample-complexity argument and is unsupported.
  2. [Section 4.3, Theorem 4.3 and Appendix B.3] The universal-expressivity half of the central claim is not established. The proof in Appendix B.3 asserts the injective mapping u_i, v_i → X_i, Z_i, {{(X_j,Z_j)|(i,j)∈E}} and concludes that non-isomorphic graphs produce different representations for all noise realizations. Proposition 4.2 provides universal approximation on compact sets satisfying a distinct-channel-multiset condition; approximation does not imply the exact injectivity needed here, and the theorem requires exact separation over the full noise space simultaneously for a single parameterization. In particular, the channel-wise update Z2_i,:,j = h(X1||X0_i||Z1_i,:,j) requires h to be injective in its last argument on an unbounded domain, which is not guaranteed by any stated width or activation condition. The same issue propagates to Theorem 4.4. Please supply a rigorous injectivity construction with explicit width conditions, or restate the theorem in a weaker form that the current proofs actually support.
  3. [Appendix B.2, proof of Proposition 4.2] The proof says 'Let all DeepSet and MLPs in our aggregator be injective' and then derives that equal aggregated representations force a common node permutation and a common channel permutation. This is not a proof: universal approximation theorems do not imply injectivity, finite-width MLPs are injective only under specific dimensional constraints, and the step that forces a single P2 across rows from 'unique column labels' is asserted rather than demonstrated. Since Proposition 4.2 is the stated basis for the expressivity theorem, the proof must be made precise, for example by invoking a known injective set-encoding theorem and giving width bounds, or the proposition should be restated with explicit hypotheses under which injectivity holds.
  4. [Section 3 and Appendix A, Eq. (47)-(54)] The covering-number proof for the hypothesis class H constructs a cover F of piecewise-constant functions on a product grid of G and Z. These functions are not T-invariant, so the resulting bound is for a larger class than the T-invariant hypothesis class that appears in Theorem 3.1. The proof should either build a cover of the invariant class directly, using the quotient metric ρ_Z,T, or explain why the non-invariant cover can be converted into one for the invariant class without changing the stated bound. As written, the theorem's proof does not match its hypothesis-class restriction.
minor comments (6)
  1. [Table 4] The table contains two rows labeled 'MPNN' with different numerical values; the second such row appears to be a different baseline and should be relabeled.
  2. [Appendix D, Table 10] The caption says 'statistics of link datasets' but the table lists node classification datasets; please correct the caption.
  3. [Appendix A, Eq. (37) and (55)] The proof uses Ω(...) in the final sample-complexity expressions while Theorem 3.1 states O(...); align the notation.
  4. [Section 6.1] The sentence 'NMPNN decrease accuracy by 32% compared with MPNN on average' should be 'decreases', and the averaging procedure across datasets should be specified.
  5. [Section 7] The conclusion contains the typo 'auxliary'; it should be 'auxiliary'.
  6. [Section 4.1, Eq. (12)] The simultaneous node-and-channel permutation action P2(P1(Z')) is easy to misread; a brief statement of the group action would improve clarity.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the central generalization and expressivity arguments are derived from PAC/covering-number definitions and a purpose-built invariant architecture; the main proof gap is an unproven injectivity claim, which is a correctness risk rather than a circular reduction.

full rationale

The paper's derivation chain is essentially self-contained. Theorem 3.1 is obtained from standard PAC/covering-number arguments in Appendix A and does not presuppose the paper's empirical claims. Proposition 3.2 and Proposition 3.3 follow directly from the definition of the quotient pseudometric in Eq. (2), not from the target generalization result. The ENGNN architecture is then explicitly constructed to realize the S_C invariance whose sample-complexity benefit Eq. (5) quantifies; this is normal theory-to-design reasoning rather than a self-definitional loop. The expressivity proof of Theorem 4.3 does contain an important gap: Appendix B.3 asserts an injective encoding 'u_i, v_i → X_i, Z_i, {(X_j,Z_j)|(i,j)∈E}' and concludes that pooling recovers edges and node features, while Proposition 4.2 only guarantees universal approximation on compact sets and the proof in Appendix B.2 assumes 'Let all DeepSet and MLPs in our aggregator be injective.' This is an unsupported injectivity/universality assertion, so it is a genuine correctness risk, but it is not circular: the theorem's conclusion is not used as an input, nor is any fitted parameter renamed as a prediction. Self-citations to Wang & Zhang or Wang et al. appear only as background for message-passing, spectral GNNs, high-order GNNs, and link-prediction features; none of the central claims (sample-complexity bound, invariance benefit, ENGNN expressivity) is justified solely by those citations. Experiments use external benchmarks with standard splits and previously reported baselines, so no fitted-input-called-prediction pattern is present. Overall, the paper is self-contained and the score reflects only the presence of non-load-bearing self-citations plus the non-circular proof gap noted above.

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

The central derivations use standard PAC covering-number machinery and DeepSet universal approximation. The main nonstandard inputs are the unproved exact covering-number ratio in Prop 3.3, the distinct-channel condition on the input domain, and the node-permutation-invariance of the noise distribution. No constants are fitted to data in the theory; hyperparameters in experiments are standard tuning.

assumptions (4)
  • standard math PAC setting with Lipschitz loss, bounded loss, and covering-number bounds on input and output spaces (Thm 3.1, Appendix A).
    Standard concentration and covering-number arguments; the proof is given in Appendix A.
  • ad hoc to paper The covering-number ratio for channel-permutation-invariant noise on [0,1]^{n×C} is exactly 1/C! (Prop 3.3).
    Stated without proof; exact equality is metric- and radius-dependent, so this is an unsupported quantitative premise.
  • domain assumption AGGR universal approximation holds only on the domain U where each noise channel has a distinct multiset of elements (Prop 4.2).
    The paper says this can be achieved by resampling noise, so in practice it holds with probability 1 for continuous noise, but it is a restriction on the input space.
  • domain assumption Noise is drawn from a node-permutation-invariant distribution and is unique per node and per channel (Thm 4.3 and 4.4).
    Standard iid continuous noise satisfies this; it is needed for the isomorphism separation argument.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Using Random Noise Equivariantly to Boost Graph Neural Networks Universally." pith.science (2026). https://pith.science/paper/4DNPPLJF

@misc{pith2026250202479,
  author       = {Pith},
  title        = {Pith review of: Using Random Noise Equivariantly to Boost Graph Neural Networks Universally},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4DNPPLJF}},
  note         = {Machine review of arXiv:2502.02479}
}
read the original abstract

Recent advances in Graph Neural Networks (GNNs) have explored the potential of random noise as an input feature to enhance expressivity across diverse tasks. However, naively incorporating noise can degrade performance, while architectures tailored to exploit noise for specific tasks excel yet lack broad applicability. This paper tackles these issues by laying down a theoretical framework that elucidates the increased sample complexity when introducing random noise into GNNs without careful design. We further propose Equivariant Noise GNN (ENGNN), a novel architecture that harnesses the symmetrical properties of noise to mitigate sample complexity and bolster generalization. Our experiments demonstrate that using noise equivariantly significantly enhances performance on node-level, link-level, subgraph, and graph-level tasks and achieves comparable performance to models designed for specific tasks, thereby offering a general method to boost expressivity across various graph tasks.

Figures

Figures reproduced from arXiv: 2502.02479 by the authors.

Figure 1
Figure 1. Example of GNNs with noise as auxiliary input. Each node has C channel noise. architectures. This includes improving message aggregation methods (Xu et al., 2019; Corso et al., 2020; Wang & Zhang, 2022b), leveraging high-order tensors to capture complex structures (Wang & Zhang, 2023; Maron et al., 2019a;b; Zhang & Li, 2021; Huang et al., 2023b; Bevilacqua et al., 2022; Qian et al., 2022; Frasca et al., 2022; Zhao e… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

74 extracted references · 63 canonical work pages

  1. [1]

    I., Grohe, M., and Lukasiewicz, T

    Abboud, R., Ceylan, I. I., Grohe, M., and Lukasiewicz, T. The surprising power of graph neural networks with random node initialization. In IJCAI, pp.\ 2112--2118, 2021 a

    work page 2021

  2. [2]

    \.I ., Grohe, M., and Lukasiewicz, T

    Abboud, R., Ceylan, \.I . \.I ., Grohe, M., and Lukasiewicz, T. The surprising power of graph neural networks with random node initialization. In IJCAI, pp.\ 2112--2118, 2021 b

    work page 2021

  3. [3]

    Adamic, L. A. and Adar, E. Friends and neighbors on the web. Social networks,, 2003

    work page 2003

  4. [4]

    Adhikari, B., Zhang, Y., Ramakrishnan, N., and Prakash, B. A. Sub2vec: Feature learning for subgraphs. In PAKDD, pp.\ 170--182, 2018

    work page 2018

  5. [5]

    and Yahav, E

    Alon, U. and Yahav, E. On the bottleneck of graph neural networks and its practical implications. In ICLR, 2021

    work page 2021

  6. [6]

    G., Li, M

    Alsentzer, E., Finlayson, S. G., Li, M. M., and Zitnik, M. Subgraph neural networks. NeurIPS, 2020

    work page 2020

  7. [7]

    Boosting Graph Neural Network Expressivity with Learnable Lanczos Constraints

    Azizi, N., Kriege, N., and Bischof, H. Enhanced expressivity in graph neural networks with lanczos-based linear constraints. CoRR, abs/2408.12334, 2024

    work page Pith review arXiv 2024

  8. [8]

    and Albert, R

    Barab \'a si, A.-L. and Albert, R. Emergence of scaling in random networks. science, 286 0 (5439): 0 509--512, 1999

    1999

Show all 74 references
  1. [9]

    M., and Maron, H

    Bevilacqua, B., Frasca, F., Lim, D., Srinivasan, B., Cai, C., Balamurugan, G., Bronstein, M. M., and Maron, H. Equivariant subgraph aggregation networks. In ICLR, 2022

    2022

  2. [10]

    P., Shirobokov, S., Rossi, E., Frasca, F., Markovich, T., Hammerla, N., Bronstein, M

    Chamberlain, B. P., Shirobokov, S., Rossi, E., Frasca, F., Markovich, T., Hammerla, N., Bronstein, M. M., and Hansmire, M. Graph neural networks for link prediction with subgraph sketching. ICLR, 2023 a

    2023

  3. [11]

    P., Shirobokov, S., Rossi, E., Frasca, F., Markovich, T., Hammerla, N., Bronstein, M

    Chamberlain, B. P., Shirobokov, S., Rossi, E., Frasca, F., Markovich, T., Hammerla, N., Bronstein, M. M., and Hansmire, M. Graph neural networks for link prediction with subgraph sketching. ICLR, 2023 b

    2023

  4. [12]

    Fastgcn: Fast learning with graph convolutional networks via importance sampling

    Chen, J., Ma, T., and Xiao, C. Fastgcn: Fast learning with graph convolutional networks via importance sampling. In ICLR, 2018

    2018

  5. [13]

    Adaptive universal generalized pagerank graph neural network

    Chien, E., Peng, J., Li, P., and Milenkovic, O. Adaptive universal generalized pagerank graph neural network. In ICLR, 2021

    2021

  6. [14]

    Principal neighbourhood aggregation for graph nets

    Corso, G., Cavalleri, L., Beaini, D., Li \` o , P., and Velickovic, P. Principal neighbourhood aggregation for graph nets. In NeurIPS, 2020

    2020

  7. [15]

    Convolutional neural networks on graphs with fast localized spectral filtering

    Defferrard, M., Bresson, X., and Vandergheynst, P. Convolutional neural networks on graphs with fast localized spectral filtering. In NeurIPS, 2016

    2016

  8. [16]

    Dong, K., Guo, Z., and Chawla, N. V. Pure message passing can estimate common neighbor for link prediction. arXiv:2309.00976, 2023

    2023 arXiv

  9. [17]

    Dwivedi, V. P. and Bresson, X. A generalization of transformer networks to graphs, 2020

    2020

  10. [18]

    P., Luu, A

    Dwivedi, V. P., Luu, A. T., Laurent, T., Bengio, Y., and Bresson, X. Graph neural networks with learnable structural and positional representations. In ICLR, 2022 a

    2022

  11. [19]

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

    Dwivedi, V. P., Ramp \' a sek, L., Galkin, M., Parviz, A., Wolf, G., Luu, A. T., and Beaini, D. Long range graph benchmark. In NeurIPS, 2022 b

    2022

  12. [20]

    and Maron, H

    Dym, N. and Maron, H. On the universality of rotation equivariant point cloud networks. In ICLR, 2021

    2021

  13. [21]

    Protein interface prediction using graph convolutional networks

    Fout, A., Byrd, J., Shariat, B., and Ben-Hur, A. Protein interface prediction using graph convolutional networks. In NeurIPS, pp.\ 6533–6542, 2017

    2017

  14. [22]

    M., and Maron, H

    Frasca, F., Bevilacqua, B., Bronstein, M. M., and Maron, H. Understanding and extending subgraph gnns by rethinking their symmetries. In NeurIPS, 2022

    2022

  15. [23]

    S., Riley, P

    Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O., and Dahl, G. E. Neural message passing for quantum chemistry. In ICML, 2017

    2017

  16. [24]

    M., Aguilera - Iparraguirre, J., Hirzel, T

    G \' o mez - Bombarelli, R., Duvenaud, D., Hern \' a ndez - Lobato, J. M., Aguilera - Iparraguirre, J., Hirzel, T. D., Adams, R. P., and Aspuru - Guzik, A. Automatic chemical design using a data-driven continuous representation of molecules, 2016

    2016

  17. [25]

    L., Ying, R., and Leskovec, J

    Hamilton, W. L., Ying, R., and Leskovec, J. Inductive representation learning on large graphs. NeurIPS, pp.\ 1025--1035, 2017

    2017

  18. [26]

    Bernnet: Learning arbitrary graph spectral filters via bernstein approximation

    He, M., Wei, Z., Huang, Z., and Xu, H. Bernnet: Learning arbitrary graph spectral filters via bernstein approximation. NeurIPS, 2021

    2021

  19. [27]

    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. In NeurIPS, 2020

    2020

  20. [28]

    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 a

    2023 arXiv

  21. [29]

    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 ICLR, 2023 b

    2023

  22. [30]

    Transformers generalize deepsets and can be extended to graphs & hypergraphs

    Kim, J., Oh, S., and Hong, S. Transformers generalize deepsets and can be extended to graphs & hypergraphs. In NeurIPS, 2021

    2021

  23. [31]

    Kipf, T. N. and Welling, M. Semi-supervised classification with graph convolutional networks. CoRR, abs/1609.02907, 2016

    2016 arXiv

  24. [32]

    Predict then propagate: Graph neural networks meet personalized pagerank

    Klicpera, J., Bojchevski, A., and G \" u nnemann, S. Predict then propagate: Graph neural networks meet personalized pagerank. In ICLR, 2019

    2019

  25. [33]

    L., L \' e tourneau, V., and Tossou, P

    Kreuzer, D., Beaini, D., Hamilton, W. L., L \' e tourneau, V., and Tossou, P. Rethinking graph transformers with spectral attention. In NeurIPS, 2021

    2021

  26. [34]

    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. In NeurIPS, 2020

    2020

  27. [35]

    Deeper insights into graph convolutional networks for semi-supervised learning

    Li, Q., Han, Z., and Wu, X. Deeper insights into graph convolutional networks for semi-supervised learning. In AAAI, 2018

    2018

  28. [36]

    D., Zhao, L., Smidt, T

    Lim, D., Robinson, J. D., Zhao, L., Smidt, T. E., Sra, S., Maron, H., and Jegelka, S. Sign and basis invariant networks for spectral graph representation learning. In ICLR, 2023

    2023

  29. [37]

    Laplacian canonization: A minimalist approach to sign and basis invariant spectral embedding

    Ma, G., Wang, Y., and Wang, Y. Laplacian canonization: A minimalist approach to sign and basis invariant spectral embedding. In NeurIPS, 2023

    2023

  30. [38]

    Provably powerful graph networks

    Maron, H., Ben - Hamu, H., Serviansky, H., and Lipman, Y. Provably powerful graph networks. In NeurIPS, 2019 a

    2019

  31. [39]

    Invariant and equivariant graph networks

    Maron, H., Ben - Hamu, H., Shamir, N., and Lipman, Y. Invariant and equivariant graph networks. In IGN, 2019 b

    2019

  32. [40]

    On learning sets of symmetric elements

    Maron, H., Litany, O., Chechik, G., and Fetaya, E. On learning sets of symmetric elements. In ICML, volume 119, 2020

    2020

  33. [41]

    Graphit: Encoding graph structure in transformers, 2021

    Mialon, G., Chen, D., Selosse, M., and Mairal, J. Graphit: Encoding graph structure in transformers, 2021

    2021

  34. [42]

    A., Martinkus, K., Faber, L., and Wattenhofer, R

    Papp, P. A., Martinkus, K., Faber, L., and Wattenhofer, R. Dropgnn: Random dropouts increase the expressiveness of graph neural networks. In NeurIPS, pp.\ 21997--22009, 2021

    2021

  35. [43]

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

    Pei, H., Wei, B., Chang, K. C., Lei, Y., and Yang, B. Geom-gcn: Geometric graph convolutional networks. In ICLR, 2020

    2020

  36. [44]

    Ordered subgraph aggregation networks

    Qian, C., Rattan, G., Geerts, F., Niepert, M., and Morris, C. Ordered subgraph aggregation networks. In NeurIPS, 2022

    2022

  37. [45]

    P., Luu, A

    Ramp \' a sek, L., Galkin, M., Dwivedi, V. P., Luu, A. T., Wolf, G., and Beaini, D. Recipe for a general, powerful, scalable graph transformer. In NeurIPS, 2022

    2022

  38. [46]

    Multi-scale attributed node embedding

    Rozemberczki, B., Allen, C., and Sarkar, R. Multi-scale attributed node embedding. J. Complex Networks, 2021

    2021

  39. [47]

    Random features strengthen graph neural networks

    Sato, R., Yamada, M., and Kashima, H. Random features strengthen graph neural networks. SDM, pp.\ 333--341, 2021

    2021

  40. [48]

    and Lipman, Y

    Segol, N. and Lipman, Y. On universal equivariant set networks. In ICLR, 2020

    2020

  41. [49]

    Pitfalls of graph neural network evaluation

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

    2018 arXiv

  42. [50]

    J., and Sinop, A

    Shirzad, H., Velingker, A., Venkatachalam, B., Sutherland, D. J., and Sinop, A. K. Exphormer: Sparse transformers for graphs. In ICML, 2023

    2023

  43. [51]

    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 ICLR, 2022

    2022

  44. [52]

    and Zhang, M

    Wang, X. and Zhang, M. GLASS: GNN with labeling tricks for subgraph representation learning. In ICLR, 2022 a

    2022

  45. [53]

    and Zhang, M

    Wang, X. and Zhang, M. How powerful are spectral graph neural networks. In ICML, 2022 b

    2022

  46. [54]

    and Zhang, M

    Wang, X. and Zhang, M. Pytorch geometric high order: A unified library for high order graph neural network. CoRR, abs/2311.16670, 2023

    2023 arXiv

  47. [55]

    Graph as point set

    Wang, X., Li, P., and Zhang, M. Graph as point set. In ICML, 2024 a

    2024

  48. [56]

    Neural common neighbor with completion for link prediction

    Wang, X., Yang, H., and Zhang, M. Neural common neighbor with completion for link prediction. In ICLR, 2024 b

    2024

  49. [57]

    A., Mirhoseini, A., Gonzalez, J

    Wu, Z., Jain, P., Wright, M. A., Mirhoseini, A., Gonzalez, J. E., and Stoica, I. Representing long-range context for graph neural networks with global attention. In NeurIPS, 2021

    2021

  50. [58]

    How powerful are graph neural networks? In ICLR, 2019

    Xu, K., Hu, W., Leskovec, J., and Jegelka, S. How powerful are graph neural networks? In ICLR, 2019

    2019

  51. [59]

    W., and Salakhutdinov, R

    Yang, Z., Cohen, W. W., and Salakhutdinov, R. Revisiting semi-supervised learning with graph embeddings. In ICML, volume 48, pp.\ 40--48, 2016

    2016

  52. [60]

    Graph convolutional networks for text classification

    Yao, L., Mao, C., and Luo, Y. Graph convolutional networks for text classification. AAAI, 33 0 (01): 0 7370--7377, 2019

    2019

  53. [61]

    Do transformers really perform badly for graph representation? In NeurIPS, 2021

    Ying, C., Cai, T., Luo, S., Zheng, S., Ke, G., He, D., Shen, Y., and Liu, T. Do transformers really perform badly for graph representation? In NeurIPS, 2021

    2021

  54. [62]

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

    You, J., Gomes - Selman, J. M., Ying, R., and Leskovec, J. Identity-aware graph neural networks. AAAI, pp.\ 10737--10745, 2021

    2021

  55. [63]

    Yun, S., Kim, S., Lee, J., Kang, J., and Kim, H. J. Neo-gnns: Neighborhood overlap-aware graph neural networks for link prediction. In NeurIPS, pp.\ 13683--13694, 2021 a

    2021

  56. [64]

    Yun, S., Kim, S., Lee, J., Kang, J., and Kim, H. J. Neo-gnns: Neighborhood overlap-aware graph neural networks for link prediction. In NeurIPS, pp.\ 13683--13694, 2021 b

    2021

  57. [65]

    Zaheer, M., Kottur, S., Ravanbakhsh, S., P \' o czos, B., Salakhutdinov, R., and Smola, A. J. Deep sets. In NeurIPS, pp.\ 3391--3401, 2017

    2017

  58. [66]

    A complete expressiveness hierarchy for subgraph gnns via subgraph weisfeiler-lehman tests

    Zhang, B., Feng, G., Du, Y., He, D., and Wang, L. A complete expressiveness hierarchy for subgraph gnns via subgraph weisfeiler-lehman tests. In ICML, 2023

    2023

  59. [67]

    and Chen, Y

    Zhang, M. and Chen, Y. Link prediction based on graph neural networks. NeurIPS, 31: 0 5165--5175, 2018

    2018

  60. [68]

    and Li, P

    Zhang, M. and Li, P. Nested graph neural networks. In NeurIPS, 2021

    2021

  61. [69]

    Labeling trick: A theory of using graph neural networks for multi-node representation learning

    Zhang, M., Li, P., Xia, Y., Wang, K., and Jin, L. Labeling trick: A theory of using graph neural networks for multi-node representation learning. NeurIPS, 34: 0 9061--9073, 2021

    2021

  62. [70]

    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 ICLR, 2022

    2022

  63. [71]

    Distance-restricted folklore weisfeiler-leman gnns with provable cycle counting power, 2023

    Zhou, J., Feng, J., Wang, X., and Zhang, M. Distance-restricted folklore weisfeiler-leman gnns with provable cycle counting power, 2023

    2023

  64. [72]

    Predicting missing links via local information

    Zhou, T., L u , L., and Zhang, Y.-C. Predicting missing links via local information. Predicting missing links via local information, 2009

    2009

  65. [73]

    Zhu, Z., Zhang, Z., Xhonneux, L. A. C., and Tang, J. Neural bellman-ford networks: A general graph neural network framework for link prediction. In NeurIPS, 2021

    2021

  66. [74]

    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.