Pith. sign in

REVIEW 4 major objections 5 minor 44 references

On the Expressive Power of Subgraph Graph Neural Networks for Graphs with Bounded Cycles

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

Pith's one-line read k-hop subgraph GNNs can approximate any continuous invariant or equivariant function on graphs whose cycles are no longer than 2k+1.

desk verdict The k=1 separation theorem is a genuine contribution; the k≥2 extension has a real proof gap in Theorem 3.5, so the main universal approximation claim for k≥2 is not established as written. read the letter →

arxiv 2502.03703 v1 pith:KWV3SORY submitted 2025-02-06 cs.LG

classification cs.LG MSC 68T0705C6068R10
keywords subgraphgraphneuralnetworksWeisfeiler-Lehmantestuniversalapproximationisomorphismboundedcyclesk-hopaggregationpermutationequivariance
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

Subgraph GNNs that read each vertex's k-hop neighborhood are known to beat ordinary message-passing GNNs, but exactly what they can represent has been unclear. This paper claims a precise threshold: under a technical k-separability condition, a k-hop subgraph GNN can approximate any continuous, permutation-invariant or permutation-equivariant function on connected graphs with no cycle longer than 2k+1, with the k=1 case holding unconditionally. If correct, the aggregation radius k becomes a design parameter tied to the data's longest cycle, and it explains why performance on molecular benchmarks saturates once k reaches about half the longest cycle. The paper also shows that k-hop GNNs that skip the subgraph structure retain the guarantee on graphs with no cycle longer than 2k-1, at the price of a stronger separability assumption.

What carries the argument

The machinery is the k-hop subgraph Weisfeiler-Lehman test, which recolors each vertex by the isomorphism type of the rooted subgraph spanned by all vertices within distance k of it. Theorem 2.5 equates its separation power with that of k-hop subgraph GNNs, so proving universal approximation reduces to proving that the test distinguishes all graphs in the class. The proof of that separation statement works by induction: start with two matched vertices of equal stabilized color, extend the partial isomorphism one vertex at a time using the rooted subgraphs of matched boundary vertices, and use Lemma B.1 to rule out hidden edges connecting the new k-hop neighborhood to the previously matched part. The k-separability condition is what makes boundary vertices uniquely identifiable by color during this extension.

What would settle it

Enumerate small connected graphs and look for two non-isomorphic graphs that both have no cycle longer than 2k+1 and are k-separable but receive identical color multisets from the k-hop subgraph WL test; the theorem asserts no such pair exists, so finding one would refute it directly. A softer empirical check is to train a k-hop subgraph GNN on graphs whose longest cycle is L and see whether accuracy on a cycle-sensitive function saturates exactly once 2k+1 >= L.

Watch

Extended reading notes

Core claim

The paper's central claim is that k-hop subgraph GNNs are universal approximators on connected graphs with bounded cycle length: under the k-separability condition, for any continuous permutation-invariant (graph-level) or permutation-equivariant (vertex-level) function and any tolerances epsilon and delta, there is a k-hop subgraph GNN whose output differs from the target by more than delta only on a set of probability less than epsilon, provided the graphs have no cycles longer than 2k+1. For k=1 the result is unconditional and covers every connected graph whose only cycles are triangles. The separating statement underneath is Theorem 3.5: two connected, k-separable graphs with no cycles longer than 2k+1 that are indistinguishable by the k-hop subgraph WL test must be isomorphic. An analogous theorem for plain k-hop GNNs without subgraph structure gives the same approximation guarantee under k-strong separability for graphs with no cycles longer than 2k-1.

Load-bearing premise

The argument depends on k-separability, meaning that every vertex at distance exactly k from a common root receives a distinct stabilized color in the k-hop subgraph WL test, which fails for symmetric graphs like paths; the authors state they do not know whether this condition can be removed, and the unconditional k=1 result does not generalize automatically.

Editorial extensions

If this is right

  • Choose k as the smallest integer with 2k+1 at least the longest cycle in the data: a k-hop subgraph GNN then has, in principle, enough expressive power to approximate any continuous invariant or equivariant property of those graphs.
  • A 1-hop subgraph GNN is already universal on connected graphs whose only cycles are triangles, with no k-separability assumption.
  • Dropping subgraph structure costs one hop: a plain k-hop GNN reaches the same universal approximation guarantee only for graphs whose cycles are at most 2k-1, and needs a stronger strong-separability condition.
  • Graph-level and vertex-level functions are covered together: both the permutation-invariant and permutation-equivariant approximation statements hold under the same cycle bound.
  • In the ZINC experiments, performance rises as k increases to about half the dominant cycle length and saturates beyond it, with k=infinity slightly worse, matching the theoretical prediction that extra aggregation distance adds no expressive benefit once the cycle bound is met.

Reading between the lines

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

  • The open k-separability question likely has a softer resolution: instead of all distance-k vertices having distinct colors, a condition on the stabilizer of the color partition (for example, a regular color-class structure) may suffice for the inductive extension, which would bring paths and other symmetric graphs back under the theorem.
  • The cycle bound suggests a cheap practical diagnostic for any graph dataset: compute the longest cycle, set k to the smallest value satisfying 2k+1 >= L (or 2k-1 >= L for plain k-hop GNNs), and stop increasing k there; the ZINC experiments are consistent with this plateau behavior.
  • The same radius-versus-diameter logic should transfer to distance-aware transformers and other architectures that mask attention to a k-hop neighborhood, predicting a similar performance cliff when the mask radius falls below half the longest cycle.
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. The paper studies k-hop subgraph GNNs, which update a vertex representation from the subgraph induced by all vertices within distance k, together with an associated k-hop subgraph WL test. The main theoretical claim is that, under a k-separability condition (Definition 3.3), k-hop subgraph GNNs can approximate any permutation-invariant or permutation-equivariant continuous function on connected graphs whose cycles all have length at most 2k+1 (Theorems 3.4 and 3.5), with an unconditional version for k=1 (Theorems 3.1 and 3.2). An analogous result is stated for k-hop GNNs that do not use the subgraph structure, under a stronger k-strong separability condition and the bound 2k-1 on cycle length (Theorems 3.8 and 3.9). The proofs proceed by Stone-Weierstrass separation arguments and by inductively constructing isomorphisms from WL color equivalence. The paper also reports ZINC experiments with a k-hop Graphormer, observing improved MAE as k increases to about 3, which the authors interpret as confirming the predicted relation between aggregation distance and cycle length.

Significance. If the main theorems were fully established, the paper would give a clean design heuristic: the aggregation radius k controls the longest cycle that a subgraph GNN can resolve. This would generalize the known tree result of Bamberger (2022) to bounded-cycle graphs and would provide a nontrivial universal approximation theorem for a widely used class of architectures. The paper is careful to state assumptions and to acknowledge that k-separability might not be removable. The inductive separation proofs in the appendices do real work beyond a direct citation of the tree result. However, the significance is substantially tempered by three issues: the k-separability and k-strong separability assumptions are ad hoc and exclude many graphs that satisfy the cycle bound; the central equivalence theorem (Theorem 2.5) is stated without proof; and the proof of the key separation theorem for k>=2 (Theorem 3.5) has a gap in the edge-preservation argument. The ZINC experiments are suggestive but do not test the k-separability assumptions, so the abstract's claim that the experiments validate the theory is stronger than what the evidence supports.

major comments (4)
  1. [Appendix B, proof of Theorem 3.5] The inductive extension of f is not established. The proof asserts that f takes Nk(S1) ∩ Nk(v1) to Nk(S2) ∩ Nk(v2), and in Case 2 it claims that equality of the multisets of neighbor colors of w1 and w2, combined with k-separability, implies that u1w1 is an edge iff u2w2 is an edge. This inference is invalid as written: multiset equality only gives a neighbor x2 of w2 whose color equals that of u1, and x2 may have distance less than k from v2, where k-separability imposes no uniqueness. Nothing in the paragraph rules out x2 ≠ u2. The assertion about the image of the intersection is also in need of proof, because the rooted isomorphism Nk(v1) ≅ Nk(v2) need not agree with the color-based extension f on vertices whose colors repeat inside the ball. Since Theorem 3.5 is the load-bearing step for Theorem 3.4, this gap must be repaired, either by strengthening the induction hypothesis or by supplying the missing uniqueness argument.
  2. [Section 2.3, Theorem 2.5] Theorem 2.5, which equates the separation power of k-hop subgraph GNNs with that of the k-hop subgraph WL test, is stated without proof and is used at every Stone-Weierstrass step, including the proofs of Theorems 3.1(i), 3.1(ii), and, through Theorem 3.5, Theorem 3.4. The statement that the proof follows the lines of Theorem 4.2 of Chen et al. (2023) is not a substitute for a proof or a precise derivation, since the update in (4) is a rooted-subgraph WL refinement rather than the setting of that reference. This is a central equivalence: it is exactly what converts WL indistinguishability into GNN separation for the Stone-Weierstrass argument. Please include a full proof in an appendix or provide an exact statement of the cited theorem together with a detailed adaptation.
  3. [Definitions 3.3 and 3.7; Theorems 3.4 and 3.8] The k-separability and k-strong separability conditions are not implied by the bounded-cycle hypothesis and exclude many graphs that satisfy the cycle bound. For example, a path with constant initial features is not k-separable for any k>=2, and a symmetric cycle is not k-strongly separable. The limitation paragraph in Section 5 does acknowledge that removal of k-separability is open, but the abstract and introduction phrase the result as applying to graphs with bounded cycles. The paper should state the scope precisely in the abstract and should quantify how restrictive k-separability is; otherwise the phrase 'under appropriate assumptions' does not convey that the main theorem may have very limited applicability. The k-strong separability condition for Theorem 3.8 is even more restrictive, since all pairs of vertices within distance 2k must have distinct stabilized colors.
  4. [Section 4, experiments] The experiments are presented as validating the theory, but the ZINC graphs are not shown to satisfy k-separability or k-strong separability, and the k-hop Graphormer is not shown to realize the separation power of the k-hop subgraph WL test used in Theorems 3.4 and 3.5. The observed MAE improvement around k=3 is consistent with the heuristic that the receptive field should cover the longest observed cycles, but it does not test the theorems' assumptions or their quantitative predictions. Please temper the claim of validation in the abstract and Section 4.3, or add synthetic experiments designed to satisfy and to violate k-separability under controlled cycle lengths.
minor comments (5)
  1. [Abstract and Section 1] The phrase 'without cycles of length greater than 2k+1' is correct but can be read as excluding the value 2k+1; consider writing 'all cycles of length at most 2k+1' for readability.
  2. [Algorithm 1] The loop is written 'while l = 1,2,...,L' but should be 'for l = 1,2,...,L'; also, the iteration limit L in Algorithm 1 is not explicitly identified with the quantifier 'for any L>0' in Definition 2.4.
  3. [Definition 3.3] The definition of k-separability refers to 'when the k-hop subgraph WL test terminates with stabilized colors and without hash collisions'; this should be formalized, since termination and stabilization depend on the hash function and the initial features.
  4. [Table 1 and Section 4.3] The text says the 1-hop Graphormer has MAE 'similar in scale' to traditional 1-hop GNNs, but Table 1 reports 0.459 for the former versus 0.088 for GIN; the difference is a factor of about five, so the wording should be adjusted.
  5. [Appendix B, notation] The notation Nk(S) is defined in Appendix B but is used in the main text only informally; moving this definition to Section 2 would improve readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the bounded-cycle approximation theorems are derived from WL separation, not from fitted inputs or self-referential definitions.

full rationale

The central claim is derived rather than fitted. Theorem 3.4 follows from a Stone-Weierstrass separation argument: Theorems 3.2 and 3.5 show that the k-hop subgraph WL test is complete on the stated bounded-cycle graph classes, and Theorem 2.5 converts WL separation into GNN approximation power. No parameter is fitted to the ZINC data and then renamed a prediction; the experiments are post hoc empirical validation, not evidence that the theorem is an input. The k-separability assumption in Definition 3.3 is a domain restriction stated in terms of WL colors, but it is not the target predicate 'all non-isomorphic graphs in the class are distinguished by the k-hop WL test'; assuming it does not make Theorem 3.5's conclusion an input by construction. The only self-citation is the omitted proof of Theorem 2.5, which refers to Theorem 4.2 of Chen et al. (2023) for the standard GNN/WL equivalence; that prior result is parameter-free, its assumptions do not include the present bounded-cycle theorem, and it is a standard equivalence rather than the paper's central claim, so it constitutes independent support rather than circularity. A possible gap in the edge-preservation step of the Theorem 3.5 proof would be a correctness concern, not a circularity, and for that reason it is not counted here. Overall, the paper is self-contained against external benchmarks and does not exhibit a circular derivation chain.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The theoretical results introduce no fitted constants; the theorems are proved from explicit color-refinement and Stone-Weierstrass arguments. The main extra assumptions are the k-separability and k-strong separability conditions, plus the abstract strength of the internal subgraph readout g^(l). The experimental k sweep is a hand-chosen hyperparameter, not part of the proof.

free parameters (1)
  • k (aggregation distance in ZINC experiments) = k=3 to 4 chosen from test MAE sweep
    The paper sweeps k over 1,2,3,4,5,7,10,infinity on the ZINC test set and interprets the plateau after k=3 as validation. This is a hyperparameter selected on test data, not a fitted constant in the theorem.
assumptions (5)
  • standard math Stone-Weierstrass theorem, including the equivariant version of Azizian and Lelarge
    Used in the proofs of Theorems 3.1, 3.4, and 3.8 to convert point separation into uniform approximation on a compact set. Stated in Appendix A as Theorem A.2.
  • domain assumption k-hop subgraph WL stabilization and collision-free hashing
    All separation proofs assume the color refinement stabilizes and the hash is injective, which is the standard idealization in the WL literature.
  • ad hoc to paper k-separability for k at least 2
    Definition 3.3 is needed for the inductive extension in Theorem 3.5. The authors state that it is unclear whether it can be removed, and it fails for many symmetric bounded-cycle graphs such as paths.
  • ad hoc to paper k-strong separability for k-hop GNNs
    Definition 3.7 is needed for Theorem 3.8 and Theorem 3.9. It is stronger than k-separability and is not guaranteed by cycle-length bounds alone.
  • domain assumption g^(l) has enough separation power on rooted k-hop subgraphs
    The model class F_k is defined abstractly with arbitrary continuous g^(l). In practice the subnetwork parameterizing g^(l) may be too weak, and the authors concede this is difficult for dense graphs in Section 2.1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of On the Expressive Power of Subgraph Graph Neural Networks for Graphs with Bounded Cycles." pith.science (2026). https://pith.science/paper/KWV3SORY

@misc{pith2026250203703,
  author       = {Pith},
  title        = {Pith review of: On the Expressive Power of Subgraph Graph Neural Networks for Graphs with Bounded Cycles},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KWV3SORY}},
  note         = {Machine review of arXiv:2502.03703}
}
abstract

Graph neural networks (GNNs) have been widely used in graph-related contexts. It is known that the separation power of GNNs is equivalent to that of the Weisfeiler-Lehman (WL) test; hence, GNNs are imperfect at identifying all non-isomorphic graphs, which severely limits their expressive power. This work investigates $k$-hop subgraph GNNs that aggregate information from neighbors with distances up to $k$ and incorporate the subgraph structure. We prove that under appropriate assumptions, the $k$-hop subgraph GNNs can approximate any permutation-invariant/equivariant continuous function over graphs without cycles of length greater than $2k+1$ within any error tolerance. We also provide an extension to $k$-hop GNNs without incorporating the subgraph structure. Our numerical experiments on established benchmarks and novel architectures validate our theory on the relationship between the information aggregation distance and the cycle size.

Figures

Figures reproduced from arXiv: 2502.03703 by the authors.

Figure 1
Figure 1. Two non-isomorphic graphs that cannot be distinguished by MP-GNNs or the WL test. The separation and expressive power of MP-GNNs are closely tied to the Weisfeiler-Lehman (WL) test (Weisfeiler & Leman, 1968), a classic algorithm designed to tackle the graph isomorphism problem. At its core, the WL test oper￾ates as a color refinement procedure: initially, each vertex vi is assigned a color C (0)(vi) based on its ini… view at source ↗
Figure 2
Figure 2. 2-hop subgraphs rooted at v1 for graphs in [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Two non-isomorphic 3-separable graphs indistinguish￾able by the classic WL test, but distinguishable by the 3-hop sub￾graph WL test. cycle with more than 2k + 1 = 7 vertices. Furthermore, for any vertex u in either graph, any distinct vertices v1 and v2 with distance exactly 3 from u are of different colors. Thus, our results imply that these two graphs can be distinguished by the 3-hop subgraph WL test. However, we… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: The k-strong separability assumption is necessary in The￾orem 3.8 0 5 10 15 20 Longest Cycle 0 1 2 3 Occurrence (×1000) [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Statistics on the longest cycle length across all testing molecules in the ZINC dataset indicate that most molecules have a longest cycle of 6, aligning with the chemical intuition that 6- membered rings are particularly stable. The peaks observed at 9 and 10 also supp…
Figure 6
Figure 6. Figure 6: Mean absolute error (MAE) on ZINC testing dataset w.r.t. different values of k for our k-hop Graphormer. Aligned with our theoretical results and dataset statistics in [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 28 canonical work pages

  1. [1]

    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 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    and Lelarge, M

    Azizian, W. and Lelarge, M. Expressive power of invariant and equivariant graph neural networks. In International Conference on Learning Representations, 2021

  3. [3]

    A topological characterisation of Weisfeiler-Leman equivalence classes

    Bamberger, J. A topological characterisation of Weisfeiler-Leman equivalence classes. In Topological, Algebraic and Geometric Learning Workshops 2022, pp.\ 17--27. PMLR, 2022

  4. [4]

    Equivariant Subgraph Aggregation Networks

    Bevilacqua, B., Frasca, F., Lim, D., Srinivasan, B., Cai, C., Balamurugan, G., Bronstein, M. M., and Maron, H. Equivariant subgraph aggregation networks. arXiv preprint arXiv:2110.02910, 2021

  5. [5]

    An optimal lower bound on the number of variables for graph identification

    Cai, J.-Y., F \"u rer, M., and Immerman, N. An optimal lower bound on the number of variables for graph identification. Combinatorica, 12 0 (4): 0 389--410, 1992

  6. [6]

    On representing linear programs by graph neural networks

    Chen, Z., Liu, J., Wang, X., Lu, J., and Yin, W. On representing linear programs by graph neural networks. In The Eleventh International Conference on Learning Representations , 2023

  7. [7]

    W., Jin, W., Rogers, L., Jamison, T

    Coley, C. W., Jin, W., Rogers, L., Jamison, T. F., Jaakkola, T., Green, W. H., Barzilay, R., and Jensen, K. F. A graph-convolutional neural network model for the prediction of chemical reactivity. Chemical Science, 10: 0 370 -- 377, 2018. URL https://api.semanticscholar.org/CorpusID:59943695

  8. [8]

    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

Show all 44 references
  1. [9]

    How powerful are k-hop message passing graph neural networks

    Feng, J., Chen, Y., Li, F., Sarkar, A., and Zhang, M. How powerful are k-hop message passing graph neural networks. Advances in Neural Information Processing Systems, 35: 0 4776--4790, 2022

  2. [10]

    Understanding and extending subgraph GNN s by rethinking their symmetries

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

  3. [11]

    Fung, V., Zhang, J., Juarez, E., and Sumpter, B. G. Benchmarking graph neural networks for materials chemistry. npj Computational Materials, 7: 0 1--8, 2021. URL https://api.semanticscholar.org/CorpusID:234207025

  4. [12]

    Exact combinatorial optimization with graph convolutional neural networks

    Gasse, M., Ch \'e telat, D., Ferroni, N., Charlin, L., and Lodi, A. Exact combinatorial optimization with graph convolutional neural networks. Advances in neural information processing systems, 32, 2019

  5. [13]

    The expressive power of k th-order invariant graph networks

    Geerts, F. The expressive power of k th-order invariant graph networks. arXiv preprint arXiv:2007.12035, 2020 a

  6. [14]

    Walk message passing neural networks and second-order graph neural networks

    Geerts, F. Walk message passing neural networks and second-order graph neural networks. arXiv preprint arXiv:2006.09499, 2020 b

  7. [15]

    and Reutter, J

    Geerts, F. and Reutter, J. L. Expressiveness and approximation properties of graph neural networks. In International Conference on Learning Representations, 2022

  8. [16]

    S., Riley, P

    Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O., and Dahl, G. E. Neural message passing for quantum chemistry. In International Conference on Machine Learning, 2017. URL https://api.semanticscholar.org/CorpusID:9665943

  9. [17]

    N., Duvenaud, D., Hern \'a ndez-Lobato, J

    G \'o mez-Bombarelli, R., Wei, J. N., Duvenaud, D., Hern \'a ndez-Lobato, J. M., S \'a nchez-Lengeling, B., Sheberla, D., Aguilera-Iparraguirre, J., Hirzel, T. D., Adams, R. P., and Aspuru-Guzik, A. Automatic chemical design using a data-driven continuous representation of mol...

  10. [18]

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

    Hamilton, W. L., Ying, R., and Leskovec, J. Inductive representation learning on large graphs. In Proceedings of the 31st International Conference on Neural Information Processing Systems, volume 30, pp.\ 1025--1035, 2017

  11. [19]

    An overview on the application of graph neural networks in wireless networks

    He, S., Xiong, S., Ou, Y., Zhang, J., Wang, J., Huang, Y., and Zhang, Y. An overview on the application of graph neural networks in wireless networks. IEEE Open Journal of the Communications Society, 2: 0 2547--2565, 2021

  12. [20]

    Boosting the cycle counting power of graph neural networks with I ^2 - GNN s

    Huang, Y., Peng, X., Ma, J., and Zhang, M. Boosting the cycle counting power of graph neural networks with I ^2 - GNN s. In The Eleventh International Conference on Learning Representations, 2023

  13. [21]

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

  14. [22]

    Lee, M., Yu, G., Dai, H., and Li, G. Y. Graph neural networks meet wireless communications: M otivation, applications, and future directions. IEEE Wireless Communications, 29 0 (5): 0 12--19, 2022

  15. [23]

    R., Wang, Y., and Wang, Y

    Liao, W., Bak-Jensen, B., Pillai, J. R., Wang, Y., and Wang, Y. A review of graph neural networks and their applications in power systems. Journal of Modern Power Systems and Clean Energy, 10 0 (2): 0 345--360, 2021

  16. [24]

    Provably powerful graph networks

    Maron, H., Ben-Hamu, H., Serviansky, H., and Lipman, Y. Provably powerful graph networks. Advances in neural information processing systems, 32, 2019

  17. [25]

    L., Lenssen, J

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

  18. [26]

    Weisfeiler and Leman go sparse: T owards scalable higher-order graph embeddings

    Morris, C., Rattan, G., and Mutzel, P. Weisfeiler and Leman go sparse: T owards scalable higher-order graph embeddings. Advances in Neural Information Processing Systems, 33: 0 21824--21840, 2020

  19. [27]

    Graph neural networks for materials science and chemistry

    Reiser, P., Neubert, M., Eberhard, A., Torresi, L., Zhou, C., Shao, C., Metni, H., van Hoesel, C., Schopmans, H., Sommer, T., and Friederich, P. Graph neural networks for materials science and chemistry. Communications Materials, 3, 2022. URL https://api.semanticscholar.org/Co...

  20. [28]

    C., Hagenbuchner, M., and Monfardini, G

    Scarselli, F., Gori, M., Tsoi, A. C., Hagenbuchner, M., and Monfardini, G. The graph neural network model. IEEE transactions on neural networks, 20 0 (1): 0 61--80, 2008

  21. [29]

    Graph neural networks in particle physics

    Shlomi, J., Battaglia, P., and Vlimant, J.-R. Graph neural networks in particle physics. Machine Learning: Science and Technology, 2 0 (2): 0 021001, 2020

  22. [30]

    Graph attention networks

    Veli c kovi \'c , P., Cucurull, G., Casanova, A., Romero, A., Lio, P., and Bengio, Y. Graph attention networks. arXiv preprint arXiv:1710.10903, 2017

  23. [31]

    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

  24. [32]

    A review on graph neural network methods in financial applications

    Wang, J., Zhang, S., Xiao, Y., and Song, R. A review on graph neural network methods in financial applications. arXiv preprint arXiv:2111.15367, 2021

  25. [33]

    and Leman, A

    Weisfeiler, B. and Leman, A. The reduction of a graph to canonical form and the algebra which appears therein. NTI, Series, 2 0 (9): 0 12--16, 1968

  26. [34]

    Wildman, S. A. and Crippen, G. M. Prediction of physicochemical parameters by atomic contributions. Journal of chemical information and computer sciences, 39 0 (5): 0 868--873, 1999

  27. [35]

    Wu, Z., Pan, S., Chen, F., Long, G., Zhang, C., and Philip, S. Y. A comprehensive survey on graph neural networks. IEEE transactions on neural networks and learning systems, 32 0 (1): 0 4--24, 2020

  28. [36]

    How powerful are graph neural networks? arXiv preprint arXiv:1810.00826, 2018

    Xu, K., Hu, W., Leskovec, J., and Jegelka, S. How powerful are graph neural networks? arXiv preprint arXiv:1810.00826, 2018

  29. [37]

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

    Xu, K., Hu, W., Leskovec, J., and Jegelka, S. How powerful are graph neural networks? In International Conference on Learning Representations, 2019

  30. [38]

    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

  31. [39]

    Rethinking the expressive power of GNN s via graph biconnectivity

    Zhang, B., Luo, S., Wang, L., and He, D. Rethinking the expressive power of GNN s via graph biconnectivity. arXiv preprint arXiv:2301.09505, 2023

  32. [40]

    and Li, P

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

  33. [41]

    Graph neural networks and their current applications in bioinformatics

    Zhang, X.-M., Liang, L., Liu, L., and Tang, M.-J. Graph neural networks and their current applications in bioinformatics. Frontiers in genetics, 12: 0 690049, 2021

  34. [42]

    From stars to subgraphs: U plifting any GNN with local structure awareness

    Zhao, L., Jin, W., Akoglu, L., and Shah, N. From stars to subgraphs: U plifting any GNN with local structure awareness. arXiv preprint arXiv:2110.03753, 2021

  35. [43]

    A practical, progressively-expressive GNN

    Zhao, L., Shah, N., and Akoglu, L. A practical, progressively-expressive GNN . Advances in Neural Information Processing Systems, 35: 0 34106--34120, 2022

  36. [44]

    Graph neural networks: A review of methods and applications

    Zhou, J., Cui, G., Hu, S., Zhang, Z., Yang, C., Liu, Z., Wang, L., Li, C., and Sun, M. Graph neural networks: A review of methods and applications. AI Open, 1: 0 57--81, 2020

Pith tools

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