Pith. sign in

REVIEW 3 major objections 5 minor 81 references

Node4All: Learning Node Representation Beyond Datasets

T0 review · 3 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read A single fixed graph encoder, pretrained only on synthetic graphs, rivals per-dataset-tuned baselines.

desk verdict A genuinely novel channel-wise graph encoder pretrained on synthetic graphs, but the 'arbitrary datasets' claim is currently a calibration hypothesis, not a demonstrated property. read the letter →

arxiv 2607.17272 v1 pith:QH5KRCGX submitted 2026-07-19 cs.LG cs.SI

Dooho Lee , Jaemin Yoo This is my paper
classification cs.LGcs.SI
keywords GraphRepresentationLearningChannelSyntheticGenerationSelf-supervisedFoundationModelsNodeClassificationTransferDataset-agnostic
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

Node4All aims to make node representation learning dataset-agnostic: one pretrained model, with no per-dataset training or hyperparameter tuning, should produce useful node embeddings for arbitrary graphs. The paper argues that this is achievable by processing each feature channel independently with a Channel Graph Transformer and pretraining the model on a stream of synthetic graphs via masked node-feature reconstruction. If correct, the result would spare practitioners the repeated training and tuning that current graph methods require, and would make node embeddings reusable across label spaces and feature dimensionalities. The paper supports this claim with node-classification experiments on 25 benchmarks, where the single fixed model ranks 1st among 21 baselines under fixed hyperparameters and 5th when baselines are tuned per dataset, and also shows one-shot and in-context learning gains over recent graph foundation models.

What carries the argument

The central object is the Channel Graph Transformer (CGT), an architecture that processes a graph channel by channel: each scalar feature channel is expanded into multi-hop tokens, each coordinate receiving a distinct receptive field, and a Transformer reads out a scalar embedding per node. This lets the model accept arbitrary feature dimensionalities with a single parameterization. The learning signal comes from a synthetic graph generator—a Chung-Lu-style structure sampler with degree-heterogeneity control, coupled to a feature generator that controls intrinsic dimensionality and graph-aware propagation—and a masked autoencoder that drops nodes and reconstructs their features from the rema

What would settle it

Take a graph with a structural or feature regime not represented in the synthetic generator—for instance, a graph with strongly disassortative mixing (homophily near 0) or with features that follow a heavy-tailed non-smooth distribution—and run the fixed Node4All model without tuning, comparing accuracy to a per-dataset-tuned GCN; a large accuracy gap would indicate the synthetic pretraining does not actually generalize beyond the generator's distribution.

Watch

Extended reading notes

Core claim

The central claim is that a node representation learner can be made dataset-agnostic at the encoder level. The paper introduces Channel Graph Transformer (CGT), an architecture that treats each feature channel as an independent graph signal and maps it to a scalar node embedding, so the same parameterization works for any feature dimensionality. Pretraining this encoder with masked node-feature reconstruction on synthetically generated graphs, rather than on any real dataset, teaches it generic structural and feature-coupling patterns. The paper reports that this single fixed model, applied uniformly to 25 real-world benchmarks, achieves competitive node-classification accuracy—1st among 21

Load-bearing premise

The model's broad applicability rests on the assumption that the synthetic graphs it is trained on resemble arbitrary real-world graphs closely enough; if a real graph has a structure or feature pattern the synthetic generator never produces, the fixed model may not transfer.

Editorial extensions

If this is right

  • Practitioners can obtain node embeddings for a new graph with a single forward pass, caching representations that are reusable across different label spaces.
  • Per-dataset hyperparameter search, often the dominant cost of existing methods, becomes unnecessary for encoder pretraining.
  • The same fixed encoder can be paired with different cheap downstream predictors, enabling one-shot and in-context node classification without retraining.
  • Pretraining on controllable synthetic graphs can substitute for collecting large, semantically aligned real-world graph corpora.

Reading between the lines

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

  • If the transfer claim generalizes beyond the 25 benchmarks, it would suggest that node classification performance is driven less by dataset-specific feature interactions and more by generic structural and feature-continuity patterns—an insight that could inform theory of graph out-of-distribution generalization.
  • The channel-wise design explicitly delegates cross-feature interaction modeling to the downstream predictor; this suggests a recipe for building 'feature-agnostic' encoders, potentially applicable to non-graph tabular data with relational structure.
  • A stress test worth running: apply the fixed model to graphs whose homophily or degree distributions fall far outside the synthetic generator's range (e.g., highly heterophilous, dense small-world graphs); success there would strengthen the 'arbitrary' claim, while failure would mark the boundary of the generator's coverage.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper introduces Node4All, a pretrained node representation learner that, without dataset-specific training or hyperparameter tuning, is claimed to transfer across arbitrary graph datasets. The two main components are the Channel Graph Transformer (CGT), which processes each feature channel independently and therefore handles arbitrary feature dimensionality under a single parameterization, and a self-supervised masked-autoencoding objective trained on a stream of synthetic graphs generated with a Chung–Lu-style structure model and a low-rank, propagation-based feature model. The authors report that one fixed Node4All model ranks 1st among 21 baselines under fixed hyperparameters and 5th under per-dataset-tuned hyperparameters on 25 node classification benchmarks, and that it outperforms several graph foundation models in one-shot and in-context settings. The paper also includes ablations, complexity analysis, and visualizations.

Significance. If the central claim holds, this is a meaningful step toward dataset-agnostic graph representation learning: a single encoder pretrained on synthetic data could be reused across many real graphs, with per-dataset adaptation delegated to a cheap downstream predictor. The strengths are the clean architectural idea behind CGT, the breadth of the benchmark comparison (25 datasets, 21 baselines), the additional one-shot and in-context evaluations, the release of code and checkpoints, and the mathematical motivation in Appendix A for propagation-based feature generation. The main uncertainty is whether the synthetic prior, whose parameter ranges are undisclosed, has been implicitly calibrated to the benchmark suite; if so, the 'beyond datasets' claim is weaker than stated.

major comments (3)
  1. [§5.1.2–5.1.3 and Figure 3] The synthetic generator is the load-bearing component of the transfer claim, yet the manuscript never specifies the sampling ranges or distributions of the latent variables (N, d̄, σ, α, S, β). Figure 3 validates synthetic graphs against exactly the 25 evaluation benchmarks, and §9.2 concedes that the generator lacks SBM communities and motifs. As presented, the result is a calibration hypothesis rather than a demonstrated property of generalization to arbitrary graphs. Please specify the priors and, ideally, evaluate on held-out benchmarks that were not used to design or validate the generator.
  2. [§7.1 and Figure 5] The central empirical ranking — 1st under Fixed-HP and 5th under Tuned-HP among 21 baselines — is not verifiable from the manuscript because only aggregate box plots are shown; the full 25-dataset accuracy table is placed on an external website. This is important for the main claim. The paper should include the complete table (per-dataset averages and standard deviations) in the main text or an appendix.
  3. [§7.2 and §7.3] The one-shot and in-context comparisons reuse baseline numbers from other papers: §7.2 says 'use their reported numbers on five datasets,' and §7.3 says baseline results are 'taken directly from the original papers.' Without a unified rerun or a very detailed protocol showing that splits, preprocessing, and evaluation are identical, the claim of outperforming graph foundation models may be affected by protocol differences. Please either rerun the baselines under the same setup or provide the information necessary to establish comparability.
minor comments (5)
  1. [§5.1.3] The symbol α is overloaded: it denotes the feature rank fraction and the propagation mixing strength. This makes equations hard to follow. Rename one of them.
  2. [§8.3 and Appendix B] The text says 'Replacing the backbone architecture, CGT with a GCN reduces average accuracy to 37.80%', but Table 3 lists GCN = 37.61 and GAT = 37.80. The sentence and the table disagree.
  3. [Figure 3] The caption mentions 300 synthetic graphs but does not state the parameter draws used to generate them. Please link Figure 3 to the concrete priors from §5.1.
  4. [§7.3] The selection criteria for the 11 in-context datasets ('current limitations of TabPFN on feature dimensionality and number of context samples') should be documented explicitly, including the thresholds used.
  5. [Conclusion] There is a typo: 'representations that generalizable' should read 'representations that generalize'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the main result is an empirical transfer claim; CGT and the synthetic-pretraining pipeline are not derived from the benchmark accuracies they predict.

full rationale

The paper's derivation chain is self-contained. CGT is defined as a channel-wise architecture (Def 4.1-4.2) with multi-hop tokenization and Transformer readout; the self-supervised objective is masked feature reconstruction on synthetic graphs (Section 5.2). No equation in the paper reduces a predicted quantity to a fitted input: the model is pretrained once on generated graphs and then applied to the 25 benchmarks, so the 1st/5th rankings and the one-shot/in-context numbers are genuine out-of-sample evaluations relative to the pretraining objective. The self-citations [40] and [41] are used only as related-work comparisons and as examples of GNN baselines; they are not invoked to justify the transfer claim or to forbid alternatives, so no self-citation chain is load-bearing. The main weakness is not circularity but external validity: Section 5.1.2-5.1.3 never specifies the sampling ranges for (N, dbar, sigma) or (S, alpha, beta), and the only evidence of coverage is Figure 3 comparing synthetic graphs to the same 25 evaluation benchmarks; Section 9.2 also concedes the generator lacks SBM communities and motifs. This means the 'arbitrary graph' claim is partly a calibration hypothesis rather than a demonstrated property, and out-of-distribution structures may not be covered. But that is a missing-support / generalization-risk concern, not a step in which the prediction equals the input by construction. Accordingly, no circular step is identified and the score is 0.

Assumptions & free parameters 8 free parameters · 4 assumptions · 0 invented entities

Node4All introduces no new physical or mathematical entities. Its free parameters are mostly unspecified generator and architecture choices that are not fitted to labels but are manually chosen, while its axioms are domain assumptions about predictability of node attributes, channel-wise independence, and representativeness of the synthetic generator. The weakest part of the ledger is the ad hoc synthetic-generator assumption: because hyperparameter ranges are unreported and visually matched to the 25 benchmark datasets, the cross-dataset generalization claim rests on a distribution-matching choice rather than an independent principle.

free parameters (8)
  • Synthetic structure generator ranges (\bar d, \sigma) = not reported
    Expected average degree and degree-heterogeneity ranges are sampled in the generator but their numerical ranges are absent from the paper; they determine which graph structures the model sees during pretraining.
  • Feature rank fraction \alpha in synthetic generator = not reported
    Controls intrinsic dimensionality of generated features; sampled but no ranges or distribution are given.
  • Propagation operator parameters (S, \alpha, \beta) = not reported
    Depth, mixing strength, and mode selector of stochastic propagation are latent variables but their ranges are not specified in the manuscript.
  • CGT hyperparameters (token dimension d, layers L, transformer layers R) = not reported
    The architecture depends on these dimensions and depths, but no values or sensitivity analysis are given in the main text.
  • DropNode ratio = not reported
    The fraction of dropped nodes in masked autoencoding is not specified, though it controls the difficulty of the reconstruction task.
  • Pretraining schedule: iterations and number of synthetic graphs = 500 iterations over 500 synthetic graphs
    Reported in Section 8.1 but no ablation for pretraining length or number of graphs is shown.
  • One-shot predictor settings (SVD dimension and ridge alpha) = 32 and 10.0
    Fixed uniformly across all one-shot datasets but no justification for these values is given; they directly affect the Table 1 results.
  • Downstream MLP hidden dimension in main experiments = not reported
    All representation methods are said to use a one-hidden-layer MLP predictor, but the hidden size is not stated.
assumptions (4)
  • domain assumption Masked reconstruction assumes node attributes are at least partially predictable from graph context.
    Section 5.2 states this assumption; if a target graph has purely independent node attributes, the DropNode reconstruction objective provides no transferable learning signal.
  • domain assumption Channel-wise independence is sufficient: cross-feature interactions can be left to a downstream predictor.
    Section 9.1 explicitly chooses not to model cross-feature interactions, and notes that adding linear self-attention lowers average accuracy from 71.55% to 70.78%; this is an inductive bias rather than a proven property.
  • ad hoc to paper The synthetic graph generator covers the structural and feature diversity of arbitrary real graphs.
    Section 5.1 and Figure 3 compare only degree distribution, community size, and feature smoothness; other graph properties such as heterophily, motifs, and label-dependent structure are not modeled, yet 'arbitrary datasets' is claimed.
  • domain assumption Randomly assigned node features carry negligible graph-aligned energy, so stochastic propagation is necessary.
    Appendix A proves only that random features have at most k/N expected energy in any k-dimensional high-frequency eigenspace; it does not establish that this concentration is harmful or that the proposed propagation improves downstream representations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Node4All: Learning Node Representation Beyond Datasets." pith.science (2026). https://pith.science/paper/QH5KRCGX

@misc{pith2026260717272,
  author       = {Pith},
  title        = {Pith review of: Node4All: Learning Node Representation Beyond Datasets},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QH5KRCGX}},
  note         = {Machine review of arXiv:2607.17272}
}
read the original abstract

Node representation learning has advanced rapidly, yet most existing methods rely on per-dataset training and hyperparameter tuning. This dataset-specific optimization comes from the difficulty of designing reusable graph models that generalize across diverse graph datasets. In this work, we introduce Node4All, a node representation learner applicable to arbitrary graph datasets without any dataset-specific optimization. Node4All is built on two complementary ideas. At the architectural level, we introduce the Channel Graph Transformer (CGT), which enables a single fixed parameterization to process arbitrary graph datasets. At the learning level, we propose a self-supervised learning based on a series of synthetic graphs. Together, these components enable generalization beyond individual datasets, which is infeasible with existing architectures and learning frameworks. We extensively evaluate Node4All on node classification across 25 benchmarks against 21 baselines, covering both supervised and self-supervised methods. Despite all baselines being trained and optimized for each dataset, a single Node4All, applied uniformly across the datasets, achieves a competitive ranking of 5th among 21 baselines. Moreover, Node4All supports one-shot and in-context learning with an appropriate predictor and outperforms recent graph foundation models (GFMs) in these settings. These results demonstrate that Node4All not only achieves reusability across arbitrary graph datasets, but also remains an effective solution in practice. Code and model checkpoints are available in https://github.com/dooho00/node4all.

Figures

Figures reproduced from arXiv: 2607.17272 by the authors.

Figure 1
Figure 1. Overview. Node4All is a pretrained node representa￾tion learner that generalizes across arbitrary graph datasets. 1 Introduction A graph is a data structure consisting of a set of nodes𝑉 and a set of edges 𝐸 that connect pairs of nodes. Graphs provide a natural way to describe many real-world systems by representing entities and their relationships, such as social networks, molecules, knowledge graphs, and biologica… view at source ↗
Figure 2
Figure 2. The Channel Graph Transformer (CGT) maps each [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Comparison between 25 real-world graph datasets [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Node4All trains CGT via masked autoencoding on [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Box plots show the distribution of average test ac [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Scatter plots of raw features (left) and Node4All [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: t-SNE visualizations of raw features and Node4All [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Scatter plots of raw features (top) and Node4All [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

81 extracted references · 8 linked inside Pith

  1. [1]

    William N Anderson Jr and Thomas D Morley. 1985. Eigenvalues of the Laplacian of a graph.Linear and multilinear algebra18, 2 (1985), 141–145

    1985

  2. [2]

    Austin R Benson, David F Gleich, and Jure Leskovec. 2016. Higher-order organi- zation of complex networks.Science353, 6295 (2016), 163–166

    2016

  3. [3]

    Shaked Brody, Uri Alon, and Eran Yahav. 2021. How attentive are graph attention networks?arXiv preprint arXiv:2105.14491(2021)

    arXiv 2021

  4. [4]

    Fenxiao Chen, Yun-Cheng Wang, Bin Wang, and C-C Jay Kuo. 2020. Graph representation learning: a survey.APSIPA Transactions on Signal and Information Processing9 (2020), e15

    2020

  5. [5]

    Haibo Chen, Xin Wang, Zeyang Zhang, Haoyang Li, Ling Feng, and Wenwu Zhu

  6. [6]

    Jinsong Chen, Kaiyuan Gao, Gaichao Li, and Kun He. 2023. NAGphormer: A Tokenized Graph Transformer for Node Classification in Large Graphs. InThe Eleventh International Conference on Learning Representations

    2023

  7. [7]

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

    2020

  8. [8]

    Runjin Chen, Tong Zhao, Ajay Kumar Jaiswal, Neil Shah, and Zhangyang Wang

Show all 81 references
  1. [9]

    Zhiqiang Chen, Hongxin Mao, Jun Liu, Yizhou Song, Bing Li, Wei Jin, Yizhou Ren, Yifan Hou, Pan Li, Chao Wang, Carl Yang, and Jiliang Tang. 2024. Text- space graph foundation models: Comprehensive benchmarks and new insights. InAdvances in Neural Information Processing Systems,...

    2024

  2. [10]

    Eli Chien, Jianhao Peng, Pan Li, and Olgica Milenkovic. 2021. Adaptive Universal Generalized PageRank Graph Neural Network. InInternational Conference on Learning Representations

    2021

  3. [11]

    Jeongwhan Choi, Jongwoo Kim, Woosung Kang, and Noseong Park. 2026. Learn- ing Posterior Predictive Distributions for Node Classification from Synthetic Graph Priors. InThe Fourteenth International Conference on Learning Representa- tions

    2026

  4. [12]

    Fan Chung and Linyuan Lu. 2002. Connected components in random graphs with given expected degree sequences.Annals of combinatorics6, 2 (2002), 125–145

    2002

  5. [13]

    Michaël Defferrard, Xavier Bresson, and Pierre Vandergheynst. 2016. Convolu- tional neural networks on graphs with fast localized spectral filtering.Advances in neural information processing systems29 (2016)

    2016

  6. [14]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xi- aohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. 2021. An Image is Worth 16x16 Words: Transformers for Image Recogn...

    2021

  7. [15]

    Vijay Prakash Dwivedi and Xavier Bresson. 2020. A generalization of transformer networks to graphs.arXiv preprint arXiv:2012.09699(2020)

    2020 arXiv

  8. [16]

    Taoran Fang, Yunchao Zhang, Yang Yang, Chunping Wang, and Lei Chen. 2023. Universal prompt tuning for graph neural networks.Advances in Neural Infor- mation Processing Systems36 (2023), 52464–52489

    2023

  9. [17]

    Yi Fang, Dongzhe Fan, Sirui Ding, Ninghao Liu, and Qiaoyu Tan. 2025. UniGLM: Training One Unified Language Model for Text-Attributed Graphs Embedding. InProceedings of the Eighteenth ACM International Conference on Web Search and Data Mining. 973–981

    2025

  10. [18]

    Wenzheng Feng, Jie Zhang, Yuxiao Dong, Yu Han, Huanbo Luan, Qian Xu, Qiang Yang, Evgeny Kharlamov, and Jie Tang. 2020. Graph random neural networks for semi-supervised learning on graphs.Advances in neural information processing systems33 (2020), 22092–22103

    2020

  11. [19]

    Matthias Fey and Jan Eric Lenssen. 2019. Fast graph representation learning with PyTorch Geometric.arXiv preprint arXiv:1903.02428(2019)

    2019 arXiv

  12. [20]

    Ben Finkelshtein, İsmail İlkan Ceylan, Michael Bronstein, and Ron Levie. 2025. Equivariance Everywhere All At Once: A Recipe for Graph Foundation Models

    2025

  13. [21]

    Dongqi Fu, Zhigang Hua, Yan Xie, Jin Fang, Si Zhang, Kaan Sancak, Hao Wu, Andrey Malevich, Jingrui He, and Bo Long. 2024. VCR-Graphormer: A Mini- batch Graph Transformer via Virtual Connections. InThe Twelfth International Conference on Learning Representations

    2024

  14. [22]

    Johannes Gasteiger, Aleksandar Bojchevski, and Stephan Günnemann. 2019. Combining Neural Networks with Personalized PageRank for Classification on Graphs. InInternational Conference on Learning Representations

    2019

  15. [23]

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

    2017

  16. [24]

    Aditya Grover and Jure Leskovec. 2016. node2vec: Scalable feature learning for networks. InProceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining. 855–864

    2016

  17. [25]

    Will Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs.Advances in neural information processing systems30 (2017)

    2017

  18. [26]

    Kaveh Hassani. 2022. Cross-domain few-shot graph classification. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 36. 6856–6864

    2022

  19. [27]

    Kaveh Hassani and Amir Hosein Khasahmadi. 2020. Contrastive multi-view representation learning on graphs. InInternational conference on machine learning. PMLR, 4116–4126

    2020

  20. [28]

    Yufei He, Yifan Sui, Xiangnan He, and Bryan Hooi. 2025. Unigraph: Learning a unified cross-domain foundation model for text-attributed graphs. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Volume 1. 448–459

    2025

  21. [29]

    Noah Hollmann, Samuel Müller, Katharina Eggensperger, and Frank Hutter. 2022. Tabpfn: A transformer that solves small tabular classification problems in a second.arXiv preprint arXiv:2207.01848(2022)

    2022 arXiv

  22. [30]

    Noah Hollmann, Samuel Müller, Lennart Purucker, Arjun Krishnakumar, Max Körfer, Shi Bin Hoo, Robin Tibor Schirrmeister, and Frank Hutter. 2025. Accurate predictions on small data with a tabular foundation model.Nature637, 8045 (2025), 319–326

    2025

  23. [31]

    Zhenyu Hou, Yufei He, Yukuo Cen, Xiao Liu, Yuxiao Dong, Evgeny Kharlamov, and Jie Tang. 2023. Graphmae2: A decoding-enhanced masked self-supervised graph learner. InProceedings of the ACM web conference 2023. 737–746

    2023

  24. [32]

    Zhenyu Hou, Xiao Liu, Yukuo Cen, Yuxiao Dong, Hongxia Yang, Chunjie Wang, and Jie Tang. 2022. Graphmae: Self-supervised masked graph autoencoders. In Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining. 594–604

    2022

  25. [33]

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

    2020

  26. [34]

    Weihua Hu, Bowen Liu, Joseph Gomes, Marinka Zitnik, Percy Liang, Vijay Pande, and Jure Leskovec. 2020. Strategies for Pre-training Graph Neural Networks. In International Conference on Learning Representations

    2020

  27. [35]

    Wei Ju, Zheng Fang, Yiyang Gu, Zequn Liu, Qingqing Long, Ziyue Qiao, Yifang Qin, Jianhao Shen, Fang Sun, Zhiping Xiao, et al. 2024. A comprehensive survey on deep graph representation learning.Neural Networks173 (2024), 106207

    2024

  28. [36]

    Brian Karrer and Mark EJ Newman. 2011. Stochastic blockmodels and community structure in networks.Physical Review E—Statistical, Nonlinear, and Soft Matter Physics83, 1 (2011), 016107

    2011

  29. [37]

    Shima Khoshraftar and Aijun An. 2024. A survey on graph representation learning methods.ACM Transactions on Intelligent Systems and Technology15, 1 (2024), 1–55

    2024

  30. [38]

    Thomas N Kipf and Max Welling. 2016. Variational graph auto-encoders.arXiv preprint arXiv:1611.07308(2016)

    2016 arXiv

  31. [39]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. 2017. Semi-Supervised Classification with Graph Convolutional Networks. InInternational Conference on Learning Repre- sentations

    2017

  32. [40]

    Dooho Lee, Myeong Kong, Sagad Hamid, Cheonwoo Lee, and Jaemin Yoo. 2025. Aggregation Buffer: Revisiting DropEdge with a New Parameter Block. InInter- national Conference on Machine Learning. PMLR, 33181–33204

    2025

  33. [41]

    Dooho Lee, Myeong Kong, Minho Jeong, and Jaemin Yoo. 2026. View Space: Learning Representation across Arbitrary Graphs. arXiv:2512.11561

    2026 arXiv

  34. [42]

    Jure Leskovec and Andrej Krevl. 2014. SNAP Datasets: Stanford Large Network Dataset Collection. http://snap.stanford.edu/data

    2014

  35. [43]

    Haoyang Li, Xin Wang, Ziwei Zhang, and Wenwu Zhu. 2025. Out-of-distribution generalization on graphs: A survey.IEEE Transactions on Pattern Analysis and Machine Intelligence(2025)

    2025

  36. [44]

    Jintang Li, Ruofan Wu, Wangbin Sun, Liang Chen, Sheng Tian, Liang Zhu, Changhua Meng, Zibin Zheng, and Weiqiang Wang. 2023. What’s behind the mask: Understanding masked graph modeling for graph autoencoders. InPro- ceedings of the 29th ACM SIGKDD Conference on Knowledge Discov...

    2023

  37. [45]

    Hao Liu, Jiarui Feng, Lecheng Kong, Ningyue Liang, Dacheng Tao, Yixin Chen, and Muhan Zhang. 2024. One For All: Towards Training One Graph Model For All Classification Tasks. InThe Twelfth International Conference on Learning Representations

    2024

  38. [46]

    Zemin Liu, Xingtong Yu, Yuan Fang, and Xinming Zhang. 2023. Graphprompt: Unifying pre-training and downstream tasks for graph neural networks. InPro- ceedings of the ACM web conference 2023. 417–428

    2023

  39. [47]

    Yuankai Luo, Lei Shi, and Xiao-Ming Wu. 2024. Classic gnns are strong base- lines: Reassessing gnns for node classification.Advances in Neural Information Processing Systems37 (2024), 97650–97669

    2024

  40. [48]

    Yuankai Luo, Lei Shi, and Xiao-Ming Wu. 2025. Can Classic GNNs Be Strong Baselines for Graph-level Tasks? Simple Architectures Meet Excellence. InForty- second International Conference on Machine Learning

    2025

  41. [49]

    Jiahong Ma, Mingguo He, and Zhewei Wei. 2024. Polyformer: Scalable node-wise filters via polynomial graph transformer. InProceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining. 2118–2129

    2024

  42. [50]

    Laurens van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-SNE. Journal of machine learning research9, Nov (2008), 2579–2605. Node4All: Learning Node Representation Beyond Datasets KDD 2026, August 9–13, 2026, Jeju Island, Republic of Korea

    2008

  43. [51]

    Haitao Mao, Zhikai Chen, Wei Jin, Haoyu Han, Yao Ma, Tong Zhao, Neil Shah, and Jiliang Tang. 2023. Demystifying structural disparity in graph neural networks: Can one size fit all?Advances in neural information processing systems36 (2023), 37013–37067

    2023

  44. [52]

    Mingdong Ou, Peng Cui, Jian Pei, Ziwei Zhang, and Wenwu Zhu. 2016. Asym- metric transitivity preserving graph embedding. InProceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining. 1105– 1114

    2016

  45. [53]

    Bryan Perozzi, Rami Al-Rfou, and Steven Skiena. 2014. Deepwalk: Online learning of social representations. InProceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining. 701–710

    2014

  46. [54]

    Oleg Platonov, Denis Kuznedelev, Michael Diskin, Artem Babenko, and Liudmila Prokhorenkova. 2023. A critical look at the evaluation of GNNs under heterophily: Are we really making progress?. InThe Eleventh International Conference on Learning Representations

    2023

  47. [55]

    Sam T Roweis and Lawrence K Saul. 2000. Nonlinear dimensionality reduction by locally linear embedding.science290, 5500 (2000), 2323–2326

    2000

  48. [56]

    Olivier Roy and Martin Vetterli. 2007. The effective rank: A measure of effective dimensionality. In2007 15th European signal processing conference. IEEE, 606–610

    2007

  49. [57]

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

    2018 arXiv

  50. [58]

    Yunsheng Shi, Zhengjie Huang, Shikun Feng, Hui Zhong, Wenjing Wang, and Yu Sun. 2021. Masked Label Prediction: Unified Message Passing Model for Semi-Supervised Classification. InProceedings of the Thirtieth International Joint Conference on Artificial Intelligence. Internatio...

    2021

  51. [59]

    Yifan Sun, Yizhou Yang, Xiaoyu Feng, Zhe Wang, Haoran Zhong, Chen Wang, and Lei Chen. 2025. Handling feature heterogeneity with learnable graph patches. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Volume 1. 1313–1324

    2025

  52. [60]

    Jian Tang, Meng Qu, Mingzhe Wang, Ming Zhang, Jun Yan, and Qiaozhu Mei

  53. [61]

    Jiabin Tang, Yuhao Yang, Wei Wei, Lei Shi, Lixin Su, Suqi Cheng, Dawei Yin, and Chao Huang. 2024. Graphgpt: Graph instruction tuning for large language models. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 491–500

    2024

  54. [62]

    Shantanu Thakoor, Corentin Tallec, Mohammad Gheshlaghi Azar, Rémi Munos, Petar Veličković, and Michal Valko. 2021. Bootstrapped representation learning on graphs. InICLR 2021 workshop on geometrical and topological representation learning

    2021

  55. [63]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need.Advances in neural information processing systems30 (2017)

    2017

  56. [64]

    Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. 2018. Graph Attention Networks. InInternational Con- ference on Learning Representations

    2018

  57. [65]

    Minjie Wang, Da Zheng, Zihao Ye, Quan Gan, Mufei Li, Xiang Song, Jinjing Zhou, Chao Ma, Lingfan Yu, Yu Gai, et al. 2019. Deep graph library: A graph- centric, highly-performant package for graph neural networks.arXiv preprint arXiv:1909.01315(2019)

    2019 arXiv

  58. [66]

    Quan Wang, Zhendong Mao, Bin Wang, and Li Guo. 2017. Knowledge graph embedding: A survey of approaches and applications.IEEE transactions on knowledge and data engineering29, 12 (2017), 2724–2743

    2017

  59. [67]

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

    2019

  60. [68]

    Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. 2018. How Powerful are Graph Neural Networks?. InInternational Conference on Learning Representa- tions

    2018

  61. [69]

    Keyulu Xu, Chengtao Li, Yonglong Tian, Tomohiro Sonobe, Ken-ichi Kawarabayashi, and Stefanie Jegelka. 2018. Representation learning on graphs with jumping knowledge networks. InInternational Conference on Machine Learn- ing. PMLR, 5453–5462

    2018

  62. [70]

    Ruosong Ye, Caiqi Zhang, Runhui Wang, Shuyuan Xu, and Yongfeng Zhang. 2024. Language is all a graph needs. InFindings of the association for computational linguistics: EACL 2024. 1955–1973

    2024

  63. [71]

    Gilad Yehudai, Ethan Fetaya, Eli Meirom, Gal Chechik, and Haggai Maron. 2021. From local structures to size generalization in graph neural networks. InInterna- tional Conference on Machine Learning. PMLR, 11975–11986

    2021

  64. [72]

    Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, and Tie-Yan Liu. 2021. Do transformers really perform badly for graph representation?Advances in neural information processing systems34 (2021), 28877–28888

    2021

  65. [73]

    Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. 2020. Graph contrastive learning with augmentations.Advances in neural information processing systems33 (2020), 5812–5823

    2020

  66. [74]

    Xinke Yu, Zeyu Gong, Chen Zhou, Yuxiao Fang, and Hongyang Zhang. 2025. SAMGPT: Text-free graph foundation model for multi-domain pre-training and cross-domain adaptation. InProceedings of the ACM on Web Conference 2025. 1142–1153

    2025

  67. [75]

    Hengrui Zhang, Qitian Wu, Junchi Yan, David Wipf, and Philip S Yu. 2021. From canonical correlation analysis to self-supervised graph neural networks.Advances in Neural Information Processing Systems34 (2021), 76–89

    2021

  68. [76]

    Hao Zhao, An Chen, Xiaoqian Sun, Hong Cheng, and Jiajun Li. 2024. All in one and one for all: A simple yet effective method towards cross-domain graph pretraining. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 4443–4454

    2024

  69. [77]

    Jianan Zhao, Zhaocheng Zhu, Mikhail Galkin, Hesham Mostafa, Michael M Bronstein, and Jian Tang. 2025. Fully-inductive Node Classification on Arbitrary Graphs. InThe Thirteenth International Conference on Learning Representations

    2025

  70. [78]

    Hao Zhu and Piotr Koniusz. 2021. Simple Spectral Graph Convolution. InInter- national Conference on Learning Representations. A Limitations of Random Node Assignment A key challenge in synthetic graph generation is producing node features that reflect the underlying graph stru...

    2021

  71. [2015]

    InProceedings of the 24th international conference on world wide web

    Line: Large-scale information network embedding. InProceedings of the 24th international conference on world wide web. 1067–1077

  72. [2024]

    InInternational Conference on Machine Learning

    LLaGA: Large Language and Graph Assistant. InInternational Conference on Machine Learning. PMLR, 7809–7823

  73. [2025]

    InForty-second International Conference on Machine Learning

    Autogfm: Automated graph foundation model with adaptive architecture customization. InForty-second International Conference on Machine Learning

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.