Pith. sign in

REVIEW 4 major objections 5 minor 31 references

Enhancing Graph Representations with Neighborhood-Contextualized Message-Passing

T0 review · 4 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read Making each message depend on the entire neighborhood yields strictly more expressive GNNs.

desk verdict SINC-GCN is a simple, efficient neighborhood-context message-passing variant with decent empirical results, but the 'strictly more expressive' claim doesn't survive contact with the paper's own 1-WL admission. read the letter →

arxiv 2511.11046 v3 pith:Y5FD6BDE submitted 2025-11-14 cs.LG cs.AI

Brian Godwin Lim , Galvin Brice Lim , Renzo Roel Tan , Irwin King , Kazushi Ikeda This is my paper
classification cs.LGcs.AI
keywords graphneuralnetworksmessage-passingneighborhood-contextualizationexpressivityconvolutionsoft-isomorphicnodeclassificationpropertyprediction
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 argues that classical message-passing graph neural networks are limited because each message is computed from the center node and one neighbor in isolation. It formalizes neighborhood-contextualization—making the message depend on the entire set of neighbors—and proposes the NCMP framework, instantiated as SINC-GCN, which injects a permutation-invariant summary of all neighbors into every message. The paper claims this makes the framework strictly more expressive than classical message-passing GNNs while keeping the same asymptotic runtime. In experiments, SINC-GCN solves a synthetic catalyst-detection task perfectly and improves on eight benchmarks over one-hop baseline GNNs. If correct, this offers a simple drop-in upgrade for any one-hop GNN without extra complexity.

What carries the argument

The central object is the neighborhood-contextualized message, ψ(h_u, h_v, {h_w: w ∈ N(u)}), which replaces the pairwise message ψ(h_u, h_v). SINC-GCN operationalizes this by adding a single permutation-invariant aggregate of all neighbor features, ⊙_{w∈N(u)} W_N h_w, inside the message MLP, alongside the center and neighbor terms. This aggregate acts as a context vector that the network can use to modulate the content of each message, not just its weight. The construction preserves permutation invariance, works with any aggregator (sum, mean, max), and reduces to ordinary message-passing when W_N is zero. Its key role is to give every message access to the whole one-hop neighborhood at no e

What would settle it

Find any message-passing GNN with a pairwise message function ψ(h_u, h_v) that attains perfect accuracy on the UniqueSignature task across all W and p_edge configurations, or exhibit a pair of graphs that a standard message-passing GNN can distinguish but SINC-GCN cannot. Either would refute the strict-expressivity claim.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the neighborhood of a node should not be treated as a bag of independent pairwise interactions: the message from a neighbor v to the center u should be a function of h_u, h_v, and the aggregate of all h_w for w ∈ N(u). The paper crystallizes this as the NCMP message ψ(h_u, h_v, {h_w: w ∈ N(u)}) and gives a concrete parametrized instance, SINC-GCN: h*_u = ⊕_{v∈N(u)} W_R σ(W_Q h_u + W_K h_v + ⊙_{w∈N(u)} W_N h_w). With a particular choice of weights, SINC-GCN provably classifies the UniqueSignature task, where standard message-passing baselines score near random. Because SIR-GCN is recovered when W_N = 0, SINC-GCN inherits 1-WL-level representatio

Load-bearing premise

The claim that message-passing GNNs are strictly weaker than NCMP assumes that no standard message-passing architecture (beyond the four tested) can solve the UniqueSignature task; the paper offers only empirical evidence for this, not a proof.

Editorial extensions

If this is right

  • SINC-GCN is claimed to be strictly more expressive than classical message-passing GNNs, meaning any task solvable by a standard message-passing GNN is also solvable by NCMP, with at least some tasks (e.g., UniqueSignature) solvable by NCMP but not by the tested baselines.
  • The construction preserves the asymptotic complexity of standard one-hop GNNs, so a model like GCN or GraphSAGE can be upgraded to SINC-GCN without a change in scaling behavior.
  • Because SIR-GCN is a special case (W_N = 0), SINC-GCN retains the 1-WL-level representational power of its predecessors, and any further expressivity gains come from the neighborhood context.
  • Empirically, the approach yields statistically significant gains on most of the eight benchmark datasets, with the largest improvements on graph-level property prediction (e.g., ZINC MAE 0.256 vs. 0.278 for SIR-GCN).
  • The framework extends the taxonomy of GNN variants: instead of three (convolutional, attentional, message-passing), it proposes a fourth dimension—neighborhood-contextualized message-passing—which can be combined with the existing flavors.

Reading between the lines

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

  • The strict-expressivity claim is not yet backed by a separation theorem; it rests on one synthetic task where four representative message-passing baselines fail. A general adversarial argument would be needed to rule out all message-passing architectures.
  • The largest benchmark gains are on graph-level prediction (ZINC, CIFAR10), which suggests neighborhood-contextualization may matter most when the global structure of a graph influences the target, hinting that graph-level tasks in chemistry or social networks could benefit.
  • Applying the same contextualization inside attention (e.g., making the key/value or the normalization depend on the entire neighborhood) is a natural next step that the paper suggests as future work.
  • Using a single shared W_N keeps the model efficient but may lose some per-neighbor identity; an auxiliary aggregator or learnable mixing could trade a small amount of efficiency for more expressive context.
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 proposes a generalization of classical message-passing GNNs called neighborhood-contextualized message passing (NCMP), in which each message from a neighbor is a function of the center node, the neighbor, and the entire set of neighbors' features. A concrete instantiation, SINC-GCN (Eq. 15), is introduced by adding a permutation-invariant neighborhood-context term N_u to the message function of SIR-GCN. The authors claim that NCMP is strictly more expressive than classical message-passing GNNs, and they support this with a synthetic diagnostic (UniqueSignature) and eight benchmark datasets, reporting competitive performance and comparable asymptotic complexity. The paper includes an explicit parameter configuration that provably solves UniqueSignature and a runtime analysis showing SINC-GCN is as efficient as GCN/SIR-GCN while outperforming them on several tasks.

Significance. If the central strict-expressivity claim were established, this would be a simple, efficient, and theoretically grounded drop-in upgrade for one-hop GNNs. The paper has several strengths: it provides an explicit complexity analysis (Eq. 16), releases code, and demonstrates on the synthetic task that SINC-GCN can be provably configured to classify perfectly, with closed-form parameter values. The empirical results on ZINC, CIFAR10, and MNIST are strong and the architecture is simple. However, the theoretical selling point—'strictly more expressive'—is not proven and, as stated, is likely false at arbitrary depth. The paper's own statement that SINC-GCN inherits the limitations of the 1-WL test (Section III-A) contradicts the possibility of strict expressivity over arbitrary-depth sum-aggregation MPNNs, which are also 1-WL-maximal. The benchmark comparison also relies on baseline numbers from external papers with per-dataset hyperparameter choices for SINC-GCN, weakening the empirical claims. The contribution is plausibly significant if reframed as a single-layer, task-level expressivity improvement rather than a strict separation over the full message-passing class.

major comments (4)
  1. [Section III, Eq. (12) and text following it] The paper claims that because NCMP generalizes message-passing, it is 'strictly more expressive.' Generalization of an architecture class does not imply strict expressivity: no separation example is provided. UniqueSignature (Table II) only shows that specific architectures (GCN, GraphSAGE, SIR-GCN, EGC-S) fail in a one-layer configuration; it does not rule out that a classical message-passing GNN can solve the task. In fact, a two-layer sum-aggregation MPNN with message ψ(h_u,h_v)=[h_v, onehot(h_v)] and an MLP update can compute both S_u and the count vector c_u, then check whether c_u[S_u]>0, solving UniqueSignature exactly. Thus the empirical evidence does not support the claimed strict expressivity advantage over classical message-passing GNNs as defined by Eq. (10).
  2. [Section III-A, paragraph on SIR-GCN and 1-WL] The paper states that SINC-GCN 'inherits the representational power and limitations of the 1-WL test.' Since sum-aggregation MPNNs with MLP updates are also 1-WL-maximal (Xu et al., 2019), this admission directly undermines the strict-expressivity claim over the general class of message-passing GNNs at graph-level discrimination. If the intended expressivity claim is limited to single-layer architectures or to task-specific reasoning with a fixed input encoding, that must be stated and proven. As written, the claim 'strictly more expressive' is internally inconsistent with the 1-WL bound.
  3. [Section IV-B and Appendix A] The benchmark results for baselines in Table III are taken from previous papers, while SINC-GCN is trained with the configuration described in Appendix A, including per-dataset choices for aggregators L, N, and graph readout (e.g., L=mean for ogbn-arxiv, L=mean/symmetric-mean/max elsewhere; N=mean or symmetric mean; readout=sum for ZINC, mean for others). The claim in Section IV-B that 'differences in model performance are solely attributed to the GNN architecture' is therefore not supported. Different hyperparameters, training budgets, and tuning protocols can account for parts of the observed gains. At minimum, the authors should reproduce baseline runs under the same training protocol or explicitly qualify the comparison as 'results reported in prior work.'
  4. [Section III-A, Eq. (15) and Section V] The paper does not define what 'expressivity' means precisely (graph-level isomorphism vs. node-level function approximation) and does not state a formal theorem about the expressivity of NCMP. The only theoretical statement is the unsupported 'strictly more expressive' claim. A rigorous treatment would define the architecture class and the feature/input space, then either provide a separation example (e.g., a pair of graphs/features distinguishable by SINC-GCN but by no classical MPNN of the same depth) or restrict the claim to a well-defined setting such as one-layer models with a given input encoding. As it stands, the central theoretical contribution is not established.
minor comments (5)
  1. [Section III-A, Eq. (15) and surrounding text] The symbol N is used both for the neighborhood N(u) and for the aggregator in Eq. (15), which is confusing. Use a different symbol, e.g., AGG_N or a script letter, for the aggregator.
  2. [Section IV-A, Table II and text] The sentence 'SIR-GCN and EGC-S exhibit near-random performance' is inaccurate for EGC-S on some configurations (e.g., 0.63±0.19 on W=3, p=0.5). It is true that performance is substantially below SINC-GCN, but 'near-random' overstates the case.
  3. [Section IV-B, Table III note] The note 'bold: statistically significant by Welch's t-test at α=0.05 vs. best baseline model' suggests multiple comparisons are made across eight datasets without correction. The reader should be told how many comparisons were performed or the test should be adjusted (e.g., Bonferroni).
  4. [Section III, Eq. (13)-(15)] The transition from Eq. (13) with node-specific W_N^{(w)} to a shared W_N in Eq. (15) is motivated by parameter efficiency and inductive learning, but the authors do not discuss any loss of expressivity caused by weight sharing. A brief comment or experiment would help.
  5. [Appendix B, Table IV] The runtime results are for a single dataset and a single GPU; they are not a substitute for a complexity analysis. Consider moving the complexity discussion to the main text and presenting runtime as supplementary evidence only.

Circularity Check

1 steps flagged · score 4.0 of 10

No fitted-input or definitional circularity in the core empirical claim; the main circularity-adjacent issue is that SINC-GCN's expressivity-inheritance/theoretical framing is load-bearing on the authors' own prior SIR-GCN paper [1] via the reduction W_N=0.

  1. self citation load bearing [Section III-A, discussion after Eq. (15)]
    "Likewise, it is also easy to see how SIR-GCN becomes an instance of SINC-GCN when W_N = 0. Hence, as SIR-GCN was shown to be comparable to a modified 1-WL test [1], it follows that SINC-GCN, as a generalization, also inherits the representational power and limitations of the 1-WL test."

    The paper's 'theoretically grounded' expressivity-inheritance claim is carried by [1], the authors' own prior work, rather than by a derivation in this paper. The reduction SINC-GCN|W_N=0 = SIR-GCN makes the inheritance formally valid, but the substantive 1-WL comparison is imported from the self-citation. This is not a fitted parameter being renamed as a prediction, and it does not drive the benchmark results, so the circularity is limited rather than central.

full rationale

The central SINC-GCN update (Eq. 15) is directly defined and then evaluated against external benchmarks (WikiCS, PATTERN, CLUSTER, MNIST, CIFAR10, ZINC, ogbn-arxiv, ogbg-molhiv) and a synthetic diagnostic. No parameter is fitted to a subset and then reported as a prediction of a closely related quantity; the hand-specified UniqueSignature weights are not learned, so the fitted-input pattern does not apply. The claim that NCMP is 'strictly more expressive' because Eq. (12) generalizes Eq. (10) is a logical/empirical gap -- no separation theorem is provided, and the paper itself concedes SINC-GCN inherits 1-WL limitations -- but that is a correctness risk, not an equivalence-by-construction circularity. The self-citations to SIR-GCN [1] are used to justify the contextualized-message design and to inherit a 1-WL expressivity bound; since SINC-GCN reduces to SIR-GCN at W_N=0, the inheritance is formally sound but load-bearing on the authors' prior result. The external benchmark results are independent of that self-citation, so the empirical headline does not collapse into the input. Overall: no significant definitional or fitted-input circularity; some self-citation reliance in the theoretical framing.

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

The paper introduces a conceptual framework (NCMP) and a new architecture (SINC-GCN) but no new physical or scientific entities. The main epistemic load is carried by self-cited expressivity results from [1] and by an unproven separation assumption for the synthetic task.

free parameters (3)
  • Aggregator L per dataset = mean / symmetric mean / max (chosen per dataset)
    Appendix A selects L among mean, symmetric mean, or max for benchmark datasets without reporting all configurations; the reported result may depend on the best-performing choice.
  • Aggregator N per dataset = mean (Benchmarking GNNs), symmetric mean (OGB)
    N is set differently across benchmarks; a design choice that affects the neighborhood-context vector.
  • Graph readout aggregator = sum (ZINC), mean (MNIST, CIFAR10, ogbg-molhiv)
    Readout choice affects graph-level results and is not controlled across baselines.
assumptions (4)
  • domain assumption SIR-GCN is comparable to a modified 1-WL test and has maximal expressivity within its class (from [1])
    Used in Section III.A to argue SINC-GCN inherits 1-WL representational power; [1] is an author-written prior paper.
  • domain assumption Two-layer MLP messages are universal function approximators of (h_u, h_v) (from [1])
    Motivates parametrizing ψ as a two-layer MLP; relies on [1].
  • ad hoc to paper Permutation-invariant aggregators over neighbor sets are sufficient to encode neighborhood context
    Eq. (14) assumes a single aggregated vector N_u captures the information needed; no expressivity analysis for different aggregators is provided.
  • ad hoc to paper The UniqueSignature task is unsolvable by any message-passing GNN, not just the baselines tested
    Needed for the strict-expressivity claim; not proven in the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Graph Representations with Neighborhood-Contextualized Message-Passing." pith.science (2026). https://pith.science/paper/Y5FD6BDE

@misc{pith2026251111046,
  author       = {Pith},
  title        = {Pith review of: Enhancing Graph Representations with Neighborhood-Contextualized Message-Passing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Y5FD6BDE}},
  note         = {Machine review of arXiv:2511.11046}
}
read the original abstract

Graph neural networks (GNNs) have become an indispensable tool for analyzing relational data. Classical GNNs are broadly classified into three variants: convolutional, attentional, and message-passing. While the standard message-passing variant is expressive, its typical pair-wise messages only consider the features of the center node and each neighboring node individually. This design fails to incorporate contextual information contained within the broader local neighborhood, potentially hindering its ability to learn meaningful relationships within the entire set of neighboring nodes. To address this, the paper first refines the concept of neighborhood-contextualization within GNNs, leveraging ideas from set-based aggregation methods and a key property of the attentional variant. This then serves as the basis for generalizing the message-passing variant to the proposed neighborhood-contextualized message-passing (NCMP) framework. To demonstrate its utility, a simple, mathematically grounded method to parametrize and operationalize NCMP is presented, leading to the development of the proposed Soft-Isomorphic Neighborhood-Contextualized Graph Convolution Network (SINC-GCN). Across a diverse set of synthetic and benchmark datasets, SINC-GCN strikes a highly favorable balance between expressivity and efficiency. Notably, while more complex models incur significant computational overhead, SINC-GCN delivers substantial performance gains with considerable effect sizes over baseline GNN models while maintaining a highly efficient asymptotic runtime complexity, further underscoring the distinctive utility of neighborhood-contextualization. Overall, by integrating multiset neighborhood context, the proposed NCMP framework serves as a practical and scalable path toward enhancing the graph representational power of classical GNNs.

Figures

Figures reproduced from arXiv: 2511.11046 by the authors.

Figure 1
Figure 1. Graph Neural Network Architecture Variants. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Neighborhood-Contextualized Message-Passing. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 2 linked inside Pith

  1. [1]

    Contextualized messages boost graph representations,

    B. G. Lim, G. B. Lim, R. R. Tan, and K. Ikeda, “Contextualized messages boost graph representations,”Transactions on Machine Learning Research, 2025. [Online]. Available: https://openreview.net/ forum?id=sXr1fRjs1N

    2025

  2. [2]

    Semi-supervised classification with graph convolutional networks,

    T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” inInternational Conference on Learning Rep- resentations, 2017

    2017

  3. [3]

    The reduction of a graph to a canon- ical form and an algebra arising during this reduction,

    B. Weisfeiler and A. Leman, “The reduction of a graph to a canon- ical form and an algebra arising during this reduction,”Nauchno- Technicheskaya Informatsia, vol. 2, no. 9, pp. 12–16, 1968

    1968

  4. [4]

    How powerful are graph neural networks?

    K. Xu, W. Hu, J. Leskovec, and S. Jegelka, “How powerful are graph neural networks?” inInternational Conference on Learning Represen- tations, 2019

    2019

  5. [5]

    Random features strengthen graph neural networks,

    R. Sato, M. Yamada, and H. Kashima, “Random features strengthen graph neural networks,” inProceedings of the 2021 SIAM International Conference on Data Mining (SDM), 2021, pp. 333–341

    2021

  6. [6]

    DeeperGCN: All you need to train deeper GCNs,

    G. Li, C. Xiong, A. Thabet, and B. Ghanem, “DeeperGCN: All you need to train deeper GCNs,” 2020, arXiv:2006.07739

    arXiv 2020

  7. [7]

    GraphSAGE-based traffic speed forecasting for segment network with sparse data,

    J. Liu, G. P. Ong, and X. Chen, “GraphSAGE-based traffic speed forecasting for segment network with sparse data,”IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 3, pp. 1755–1766, 2022

    2022

  8. [8]

    Understanding graph isomorphism network for rs-fMRI functional connectivity analysis,

    B.-H. Kim and J. C. Ye, “Understanding graph isomorphism network for rs-fMRI functional connectivity analysis,”Frontiers in Neuroscience, vol. 14, p. 630, 2020

    2020

Show all 31 references
  1. [9]

    Benchmarking graph neural networks,

    V . P. Dwivedi, C. K. Joshi, A. T. Luu, T. Laurent, Y . Bengio, and X. Bresson, “Benchmarking graph neural networks,”Journal of Machine Learning Research, vol. 24, no. 43, pp. 1–48, 2023

    2023

  2. [10]

    Open graph benchmark: Datasets for machine learning on graphs,

    W. Hu, M. Fey, M. Zitnik, Y . Dong, H. Ren, B. Liu, M. Catasta, and J. Leskovec, “Open graph benchmark: Datasets for machine learning on graphs,” inAdvances in Neural Information Processing Systems, vol. 33, 2020, pp. 22 118–22 133

    2020

  3. [11]

    Graph attention networks,

    P. Veli ˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P. Li `o, and Y . Bengio, “Graph attention networks,” inInternational Conference on Learning Representations, 2018

    2018

  4. [12]

    How attentive are graph attention networks?

    S. Brody, U. Alon, and E. Yahav, “How attentive are graph attention networks?” inInternational Conference on Learning Representations, 2022

    2022

  5. [13]

    EGAT: Edge-featured graph attention network,

    Z. Wang, J. Chen, and H. Chen, “EGAT: Edge-featured graph attention network,” inArtificial Neural Networks and Machine Learning – ICANN

  6. [14]

    FinGAT: Financial graph attention networks for recommending top-k profitable stocks,

    Y .-L. Hsu, Y .-C. Tsai, and C.-T. Li, “FinGAT: Financial graph attention networks for recommending top-k profitable stocks,”IEEE Transactions on Knowledge and Data Engineering, vol. 35, no. 1, pp. 469–481, 2023

    2023

  7. [15]

    GATrust: A multi-aspect graph attention network model for trust assessment in OSNs,

    N. Jiang, J. Wen, J. Li, X. Liu, and D. Jin, “GATrust: A multi-aspect graph attention network model for trust assessment in OSNs,”IEEE Transactions on Knowledge and Data Engineering, vol. 35, no. 6, pp. 5865–5878, 2023

    2023

  8. [16]

    FinSIR: Financial SIR-GCN for market-aware stock recommendation,

    B. G. Lim, J. Liu, H. J. Ong, J. Adrian Chan, R. R. Tan, I. King, and K. Ikeda, “FinSIR: Financial SIR-GCN for market-aware stock recommendation,” in2025 International Joint Conference on Neural Networks (IJCNN), 2025, pp. 1–8

    2025

  9. [17]

    AGTCNet: A graph-temporal approach for principled motor imagery EEG classification,

    G. B. S. Lim, B. G. S. Lim, A. A. Bandala, J. A. C. Jose, T. S. C. Chu, and E. Sybingco, “AGTCNet: A graph-temporal approach for principled motor imagery EEG classification,”IEEE Access, vol. 13, pp. 187 383– 187 409, 2025

    2025

  10. [18]

    Neural message passing for quantum chemistry,

    J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, and G. E. Dahl, “Neural message passing for quantum chemistry,” inProceedings of the 34th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 70, 2017, pp. 1263–1272

    2017

  11. [19]

    Principal neighbourhood aggregation for graph nets,

    G. Corso, L. Cavalleri, D. Beaini, P. Li `o, and P. Veli ˇckovi´c, “Principal neighbourhood aggregation for graph nets,” inAdvances in Neural Information Processing Systems, vol. 33, 2020, pp. 13 260–13 271

    2020

  12. [20]

    Do we need anisotropic graph neural networks?

    S. A. Tailor, F. Opolka, P. Lio, and N. D. Lane, “Do we need anisotropic graph neural networks?” inInternational Conference on Learning Representations, 2022

    2022

  13. [21]

    Weisfeiler and leman go neural: higher-order graph neu- ral networks,

    C. Morris, M. Ritzert, M. Fey, W. L. Hamilton, J. E. Lenssen, G. Rattan, and M. Grohe, “Weisfeiler and leman go neural: higher-order graph neu- ral networks,” inProceedings of the Thirty-Third AAAI Conference on Artificial Intelligence and Thirty-First Innovative Applications ...

    2019

  14. [22]

    Expressive power of invariant and equiv- ariant graph neural networks,

    W. Azizian and M. Lelarge, “Expressive power of invariant and equiv- ariant graph neural networks,” inInternational Conference on Learning Representations, 2021

    2021

  15. [23]

    Weisfeiler and Lehman go cellular: CW networks,

    C. Bodnar, F. Frasca, N. Otter, Y . Wang, P. Li `o, G. F. Montufar, and M. Bronstein, “Weisfeiler and Lehman go cellular: CW networks,” in Advances in Neural Information Processing Systems, vol. 34, 2021, pp. 2625–2640

    2021

  16. [24]

    Do transformers really perform badly for graph representation?

    C. Ying, T. Cai, S. Luo, S. Zheng, G. Ke, D. He, Y . Shen, and T.-Y . Liu, “Do transformers really perform badly for graph representation?” inAdvances in Neural Information Processing Systems, vol. 34, 2021, pp. 28 877–28 888

    2021

  17. [25]

    Adaptive graph diffusion networks,

    C. Sun, J. Hu, H. Gu, J. Chen, and M. Yang, “Adaptive graph diffusion networks,” 2022, arXiv:2012.15024

    2022 arXiv

  18. [26]

    Improving graph neural network expressivity via subgraph isomorphism count- ing,

    G. Bouritsas, F. Frasca, S. Zafeiriou, and M. M. Bronstein, “Improving graph neural network expressivity via subgraph isomorphism count- ing,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 1, pp. 657–668, 2023

    2023

  19. [27]

    Towards expressive graph representations for graph neural networks,

    C. Mao, L. Yao, and Y . Luo, “Towards expressive graph representations for graph neural networks,” in2024 IEEE International Conference on Data Mining (ICDM), 2024, pp. 797–802

    2024

  20. [28]

    Towards dynamic mes- sage passing on graphs,

    J. Sun, C. Yang, X. Ji, Q. Huang, and S. Wang, “Towards dynamic mes- sage passing on graphs,” inAdvances in Neural Information Processing Systems, vol. 37, 2024, pp. 80 936–80 964

    2024

  21. [29]

    A simple and expressive graph neural network based method for structural link rep- resentation,

    V . Lachi, F. Ferrini, A. Longa, B. Lepri, and A. Passerini, “A simple and expressive graph neural network based method for structural link rep- resentation,” inProceedings of the Geometry-grounded Representation Learning and Generative Modeling Workshop (GRaM), ser. Proceedin...

    2024

  22. [30]

    Neural execution of graph algorithms,

    P. Veli ˇckovi´c, R. Ying, M. Padovano, R. Hadsell, and C. Blundell, “Neural execution of graph algorithms,” inInternational Conference on Learning Representations, 2020

    2020

  23. [2021]

    Cham: Springer, 2021, pp. 253–264

    2021

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.