Pith. sign in

REVIEW 4 major objections 5 minor 47 references

Matching feature dimensions is not enough: graph foundation models need four properties of true feature unification, which SliGFM targets with smoothness-ordered sliding windows and reconstruction.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

Ordering node features by topological smoothness and encoding them with a shared sliding-window transformer plus reconstruction yields transferable cross-domain graph representations without fine-tuning.

T0 review reviewed 2026-07-31 challenge →

load-bearing objection Solid GFM systems paper with a clean framing and a real soft spot on whether smoothness order actually unifies semantics across domains. the 4 major comments →

arxiv 2607.27966 v1 pith:D7CQES72 submitted 2026-07-30 cs.LG

What Makes Graph Unified? Principles and Generative Sliding-Window Transformer for Graph Foundation Models

Dongxiao He , Siqi Liu , Jitao Zhao , Yawen Li , Yi Wang , Di Jin This is my paper
classification cs.LG
keywords Graph Foundation Modelscross-domain graph learningfeature unificationsliding-window tokenizationtopological smoothnessgenerative reconstructionfew-shot node classificationzero-shot graph classification
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Graph foundation models aim to pretrain once and reuse knowledge across many graphs, but node features differ wildly in size and meaning from dataset to dataset. Most prior work forces those features into a fixed width and treats that as unification. This paper argues that is not enough, and states four requirements: the features must share a common format, carry meanings that transfer across domains, keep enough of the original information to reconstruct it, and fit the inductive bias of the backbone. SliGFM meets those requirements by sorting each graph’s feature channels by how smoothly they vary along edges, chopping the sorted channels into overlapping windows with a shared encoder, modeling relations among the resulting tokens with a smoothness-biased transformer, reconstructing the original patches during pretraining, and fusing multi-hop neighborhoods with weights driven by graph-level structure statistics. The pretrained model is then used frozen on new graphs. Experiments on few-shot node classification and zero-shot graph classification show gains over prior graph foundation models, supporting the claim that principled feature unification, not dimension matching alone, is what makes cross-domain graph transfer work.

Core claim

Effective cross-domain graph feature unification is defined by four desiderata—formal uniformity, cross-domain transferability, information preservation, and backbone compatibility—not by dimensional alignment alone. Ordering features by topological smoothness, encoding them as fixed tokens with a shared sliding window, modeling token relations with a relative-smoothness bias, and training with generative reconstruction yields unified node representations that transfer across heterogeneous graphs without downstream tuning.

What carries the argument

Topology-aware sliding-window feature tokenization: feature dimensions are ranked by edge-wise smoothness, partitioned into overlapping patches, and mapped by one shared encoder into ordered fixed-dimensional tokens; a smoothness-biased intra-node transformer and a generative patch reconstructor then enforce transferable structure and information preservation.

Load-bearing premise

The load-bearing premise is that sorting feature channels by how smoothly they change along edges creates a shared semantic order across domains, so tokens at similar ranks mean comparable things on citation, shopping, Wikipedia, and other graphs.

What would settle it

Ablate or reverse the smoothness ordering on held-out domains (or replace it with random or original column order) while keeping window size, reconstruction, and propagation fixed: if cross-domain few-shot accuracy then collapses to or below strong baselines that only match dimensions, the claim that smoothness order is the transferable reference fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • GFM design and evaluation should score feature unifiers on transferability, reconstructibility, and backbone fit, not only on matching input width.
  • A single frozen backbone can serve few-shot node tasks and zero-shot graph tasks when features are tokenized this way and hop fusion is conditioned on graph statistics.
  • Generative reconstruction of original feature patches becomes a practical regularizer against irreversible loss in shared token spaces.
  • Adaptive multi-hop fusion from degree, clustering, and motif statistics can replace a fixed receptive field when graphs differ in sparsity and clustering.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If smoothness rank is only a weak shared reference, hybrid cues (e.g., distributional or metadata anchors alongside smoothness) may be needed before the four desiderata can be fully met on arbitrary numeric graphs.
  • The same tokenization-plus-reconstruction pattern could be stress-tested on graphs with no natural node attributes, where features are purely structural or randomly initialized.
  • Reporting reconstruction error alongside downstream accuracy would give a direct operational check of the information-preservation desideratum in future GFM benchmarks.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper argues that dimensional alignment alone is insufficient for cross-domain graph feature unification and proposes four desiderata: formal uniformity, cross-domain transferability, information preservation, and backbone compatibility. Guided by these, it introduces SliGFM: feature dimensions are ordered by topological smoothness (Eq. 7), scanned by a shared sliding-window encoder into fixed tokens, modeled by an intra-node Transformer with relative-smoothness bias, enriched by structure-guided adaptive multi-hop propagation, and trained with generative patch reconstruction plus local/global representation losses. The frozen model is evaluated on few-shot node classification across nine graphs and zero-shot graph classification on three datasets, reporting best or competitive results versus recent GFM baselines, with ablations and a window-size sensitivity study.

Significance. If the transfer story holds, the work offers a useful conceptual checklist for GFM feature unification and a concrete alternative to PCA/SVD or language-mediated pipelines that is backbone-compatible with Transformers and does not require textualization. Strengths include a clear modular design, multi-baseline comparisons with standard deviations (Tables 1–2, 5–6), component ablations (Fig. 3), and hyperparameter sweeps (Fig. 4). The desiderata framing and the combination of smoothness-ordered tokenization with explicit reconstruction are incremental but publishable contributions to graph foundation models, provided the cross-domain claims are more carefully isolated from multi-graph pretraining effects.

major comments (4)
  1. [§5.1, Table 1] §5.1 states joint pretraining on Cora, PubMed, Computers, and Chameleon, then frozen evaluation on all nine node-classification benchmarks, including those four. Table 1’s largest one-shot margins over TFSGFM fall precisely on pretraining sets (Chameleon +3.91, PubMed +3.82, Cora +2.62), while held-out Squirrel is not best. For a foundation-model / tuning-free cross-domain claim, this protocol confounds “unification enables transfer” with “strong multi-graph pretraining on overlapping domains.” A load-bearing revision is a true leave-domain-out protocol (pretrain excluding each target family) and primary reporting on fully held-out graphs.
  2. [§3.1 Eqs. 3–4; §4.2 Eq. 7; Fig. 3] Cross-domain transferability is formalized in §3.1 (Eqs. 3–4) as a shared semantic map ρ with small d_S discrepancy across domains. The mechanism in §4.2 assumes that ranking dimensions by edge-wise squared variation and windowing nearby ranks yields tokens with comparable meaning across citation, co-purchase, Wikipedia, and actor graphs. The paper never tests this assumption: there is no cross-domain token-rank correspondence, no frozen-encoder reconstruction or retrieval across datasets, and no probe that token k aligns semantically across domains. Fig. 3 shows sorting helps, but that is also consistent with a purely within-graph inductive bias for the intra-node Transformer. Without a direct measurement, desideratum (2) remains an untested load-bearing assumption behind the unification narrative.
  3. [§3.1 Eqs. 5–6; §4.5–4.6 Eq. 15] Information preservation is defined via domain-specific decoders and reconstruction error (Eqs. 5–6), yet training uses a single shared patch decoder and reports only the joint pretraining MSE objective (Eq. 15), not per-domain reconstruction quality or whether reconstructed features retain task-relevant signal on held-out graphs. Linking ε_rec (or empirical recon error) to downstream few-shot accuracy—or showing that removing reconstruction hurts transfer more on unseen domains than on pretrain domains—would make this desideratum operational rather than aspirational.
  4. [§5.3, Table 2, Appendix C] Zero-shot graph classification (§5.3, Table 2) obtains graph embeddings by mean-pooling node embeddings from a model whose pretraining objectives (§4.6) are node-centric (patch recon, edge local alignment, global dispersion). Gains over TFSGFM are small (e.g., IMDB-BINARY 64.56 vs 63.67; DD 73.83 vs 73.54). Clarify whether any graph-level pretraining signal is used (Appendix C says “pretrained at the graph level” without specifying the objective) and whether significance holds under the same random seeds/splits as the cited baselines; otherwise the graph-level transfer claim is weakly supported relative to the node-level tables.
minor comments (5)
  1. [§4.2, Figure 2] Typos and wording: “givn by” (§4.2); “Feture” / “Meomry” / “Ohter” in Figure 2; “Reconstructed Feture Patches” repeated; “equating dimensional uniformity with effective feature unification” is clear in the abstract but the intro sometimes blurs “unification” vs “dimensional alignment.”
  2. [§2–4, Table 3] Notation: ¯X is used both for the shared representation space and for the smoothness-reordered feature matrix; I_m vs node index sets could be tightened. Table 3 helps but should match in-text symbols (e.g., z^{(m)}_{i,cls} vs c^{(m)}_i).
  3. [§5.1, Appendix C] Report exact d_p, δ, λ weights, L, d_t, and K_hop used for main tables in the main text or a compact config table; Appendix C is thin on hyperparameters relative to the free-parameter list implied by §4.
  4. [Figure 3] Fig. 3 ablation is only on five datasets; including at least one fully held-out set (e.g., CS or Squirrel) in the main ablation would better support the component claims.
  5. [§3.2] Related work: briefly contrast with concurrent feature-alignment GFMs (e.g., GraphAlign, AnyGraph) on whether they satisfy the four desiderata, to sharpen the novelty claim beyond MDGPT/SAMGPT/FUG/TIG.

Circularity Check

0 steps flagged

No load-bearing circularity: SliGFM is a designed encoder plus external few/zero-shot benchmarks, not a derivation that reduces to its inputs.

full rationale

This is a methods/engineering GFM paper, not a first-principles derivation. The four desiderata (§3.1) are design criteria the authors state and then implement (smoothness sort Eq. 7, sliding-window tokens Eq. 8, relative-smoothness bias Eq. 10, reconstruction Eq. 15, hop weights from graph statistics Eq. 11–12). None of the reported accuracies is forced by fitting a parameter that is then renamed a prediction, nor is any uniqueness theorem imported to forbid alternatives. Evaluation is on external labeled node/graph classification benchmarks (Tables 1–2, 5–6) with frozen pretrained weights. Same-group baselines (TFSGFM, FUG, TIG, LEDA) and the SGRL-style global dispersion term are ordinary related-work reuse; they do not make the central transfer claim true by construction. Weaknesses raised by the skeptic (unmeasured cross-domain token semantics; pretrain/eval set overlap) are empirical-support issues, not circular reductions. Minor self-citation of the authors’ SGRL scattering objective for L_global is the only self-reference in the training recipe and is not load-bearing for the unification claim. Score 1 solely for that non-central self-citation; steps left empty because no step exhibits Eq. X = input by construction.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 4 invented entities

The central transfer claim rests on domain assumptions about graphs (features vary smoothly with topology in a cross-domain-comparable way), standard deep-learning training machinery, and several hand-chosen architectural knobs (window size, stride, loss weights, hop generator inputs). No new physical entities; invented pieces are architectural modules. Free parameters are ordinary ML hyperparameters, not physical constants.

free parameters (5)
  • sliding window size d_p and stride δ = d_p=256 (preferred); δ=d_p/4
    Chosen by sensitivity sweep; best reported around d_p=256 with δ=d_p/4. Directly controls token granularity and thus the unified representation.
  • loss weights λ_recon, λ_local, λ_global
    Balance reconstruction vs local alignment vs global dispersion; not derived, must be set for pretraining success.
  • relative smoothness bias scales λ_h per attention head
    Head-specific coefficients in Eq. 10 shape which token pairs interact; architectural free parameters.
  • maximum propagation hops K_hop and hop-weight MLP
    Controls adaptive multi-hop fusion; graph-statistic features and MLP mapping are design choices affecting structural transfer.
  • token dimension d_t, transformer depth L, patch encoder/decoder MLPs
    Standard capacity hyperparameters that determine whether the shared token space is expressive enough.
axioms (5)
  • domain assumption Feature dimensions can be meaningfully totally ordered by topological smoothness s_j = -avg_{(u,v)∈E}(x_u,j - x_v,j)^2 so that similar ranks share transferable semantics across domains.
    Load-bearing premise of §4.2 tokenization and of the cross-domain transferability desideratum.
  • domain assumption A parameter-shared encoder on fixed-size patches plus transformer mixing yields formally uniform, backbone-compatible tokens for heterogeneous d_m.
    Standard representation-learning assumption underlying Eqs. 8–10; not proved, supported empirically.
  • ad hoc to paper Approximate patch reconstruction (MSE ≤ ε_rec) is an adequate operationalization of information preservation for downstream transfer.
    Section 3.1 relaxes injectivity to expected reconstruction loss; paper optimizes MSE but does not validate ε_rec against semantic fidelity.
  • domain assumption Graph-level histograms (degree, clustering, motif counts) suffice for the hop-weight generator to adapt receptive fields across structural regimes.
    Section 4.4; assumes these statistics capture the heterogeneity that would otherwise break fixed-hop GNNs.
  • standard math Standard transformer attention, SGC-style polynomial propagation, and contrastive-style local/global losses are valid optimization scaffolding.
    Uses Vaswani et al. attention, ALiBi-style linear bias, and SGRL-style scattering without new theory.
invented entities (4)
  • Four desiderata for cross-domain graph feature unification no independent evidence
    purpose: Evaluation checklist: formal uniformity, transferability, information preservation, backbone compatibility.
    Conceptual contribution of §3.1; useful framing but not independently measured as a validated theory outside SliGFM’s design narrative.
  • Topology-aware sliding-window feature tokenizer (SliGFM tokens) no independent evidence
    purpose: Map variable-d features into ordered fixed-d tokens via smoothness sort + shared window encoder.
    Core architectural invention; evidence is downstream accuracy/ablations, not an external measurable quantity.
  • Relative-smoothness attention bias no independent evidence
    purpose: Bias intra-node attention toward tokens with similar smoothness ranks.
    Adaptation of ALiBi distance to smoothness difference (Eq. 10); no external falsifiable prediction beyond model ablations.
  • Structure-guided adaptive hop-weight generator no independent evidence
    purpose: Produce graph-specific multi-hop fusion weights from structural statistics.
    Module introduced in §4.4 to handle structural heterogeneity; validated only inside the paper’s benchmarks.

reviewed 2026-07-31 · how reviews work

0 comments
Cite this review

Pith. "Pith review of What Makes Graph Unified? Principles and Generative Sliding-Window Transformer for Graph Foundation Models." pith.science (2026). https://pith.science/paper/D7CQES72

@misc{pith2026260727966,
  author       = {Pith},
  title        = {Pith review of: What Makes Graph Unified? Principles and Generative Sliding-Window Transformer for Graph Foundation Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D7CQES72}},
  note         = {Machine review of arXiv:2607.27966}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Graph Foundation Models (GFMs) have recently emerged as a promising paradigm for general-purpose graph learning, aiming to learn reusable knowledge that generalizes across diverse graph domains and downstream tasks, reducing the need for specific model development. Achieving this goal requires reconciling the substantial heterogeneity in node features, graph structures, and semantic information across domains. Among them, heterogeneous node features constitute a fundamental input-level barrier, as their dimensionality and semantics vary substantially across datasets. Existing studies typically project or map heterogeneous node features into a fixed-dimensional space, often implicitly equating dimensional uniformity with effective feature unification. Yet dimensional consistency alone does not ensure that the unified features preserve informative semantics and capture transferable patterns that can support cross-domain knowledge transfer. To bridge this conceptual gap, we distill four desiderata for cross-domain graph feature unification: formal uniformity, cross-domain transferability, information preservation, and backbone compatibility. Guided by these principles, we propose SliGFM, a graph foundation model built upon topology-aware sliding-window feature encoding and generative reconstruction. SliGFM orders feature dimensions by topological smoothness and scans the reordered features with a shared sliding-window feature encoder, transforming heterogeneous features into a common space of ordered fixed-dimensional feature tokens. This formulation enables a smoothness-aware transformer to capture transferable relational patterns among feature tokens within each node, while the generative reconstruction objective encourages preservation of the original feature information.

Figures

Figures reproduced from arXiv: 2607.27966 by Di Jin, Dongxiao He, Jitao Zhao, Siqi Liu, Yawen Li, Yi Wang.

Figure 1
Figure 1. Figure 1: Schematic illustration of the four desiderata for [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The overview of the proposed SliGFM model. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Ablation study on five cross-domain one-shot node [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Parameter sensitivity analysis on the sliding win [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

47 extracted references · 8 linked inside Pith

  1. [1]

    Borgwardt, Cheng Soon Ong, Stefan Schönauer, S

    Karsten M. Borgwardt, Cheng Soon Ong, Stefan Schönauer, S. V. N. Vishwanathan, Alexander J. Smola, and Hans-Peter Kriegel. 2005. Protein function prediction via graph kernels. InProceedings Thirteenth International Conference on Intelligent Systems for Molecular Biology 2005, Detroit, MI, USA, 25-29 June 2005. 47–56

    2005

  2. [2]

    Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin...

    2020

  3. [3]

    Runjin Chen, Tong Zhao, Ajay Jaiswal, Neil Shah, and Zhangyang Wang. 2024. LLaGA: large language and graph assistant. InProceedings of the 41st International Conference on Machine Learning, ICML 2024. 7809–7823

    2024

  4. [4]

    Zhikai Chen, Haitao Mao, Jingzhe Liu, Yu Song, Bingheng Li, Wei Jin, Bahare Fatemi, Anton Tsitsulin, Bryan Perozzi, Hui Liu, and Jiliang Tang. 2024. Text- space Graph Foundation Models: Comprehensive Benchmarks and New Insights. InAdvances in neural information processing systems

    2024

  5. [5]

    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 Recognition at Scale. InProceedings of the 9th International Conference on Learning Re...

    2021

  6. [6]

    Moshe Eliasof, Krishna Sri Ipsit Mantri, Beatrice Bevilacqua, Bruno Ribeiro, and Carola-Bibiane Schönlieb. 2026. Bridging Input Feature Spaces Towards Graph Foundation Models.arXiv preprint arXiv:2605.04834(2026)

    Pith/arXiv arXiv 2026

  7. [7]

    Taoran Fang, Yunchao Zhang, Yang Yang, Chunping Wang, and Lei Chen. 2023. Universal Prompt Tuning for Graph Neural Networks. InAdvances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, Alice Oh, Tristan Naumann, Amir Globerson, Kate Sae...

    2023

  8. [8]

    Mikhail Galkin, Xinyu Yuan, Hesham Mostafa, Jian Tang, and Zhaocheng Zhu

  9. [9]

    Dongxiao He, Lianze Shan, Jitao Zhao, Hengrui Zhang, Zhen Wang, and Weixiong Zhang. 2024. Exploitation of a Latent Mechanism in Graph Contrastive Learning: Representation Scattering. InAdvances in Neural Information Processing Systems 37: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15...

    2024

  10. [10]

    Xiaodong He, Haolan He, Ruiyi Fang, Ming Sun, and Zhao Kang. 2026. Structure-Centric Graph Foundation Model via Geometric Bases.arXiv preprint arXiv:2605.08689(2026)

    Pith/arXiv arXiv 2026

  11. [11]

    Yufei He, Yuan Sui, Xiaoxin He, and Bryan Hooi. 2025. UniGraph: Learning a Uni- fied Cross-Domain Foundation Model for Text-Attributed Graphs. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining, V.1, KDD 2025, Toronto, ON, Canada, August 3-7, 2025, Yizhou Sun, Flavio Chierichetti, Hady W. Lauw, Claudia Perlich, Wee Hyon...

    2025

  12. [12]

    Zhenyu Hou, Haozhan Li, Yukuo Cen, Jie Tang, and Yuxiao Dong. 2024. GraphAlign: Pretraining One Graph Neural Network on Multiple Graphs via Feature Alignment.arXiv preprint arXiv:2406.02953(2024)

    Pith/arXiv arXiv 2024

  13. [13]

    Mohsen Jamali and Martin Ester. 2010. A matrix factorization technique with trust propagation for recommendation in social networks. InProceedings of the 2010 ACM Conference on Recommender Systems, RecSys 2010, Barcelona, Spain, September 26-30, 2010, Xavier Amatriain, Marc Torrens, Paul Resnick, and Markus Zanker (Eds.). ACM, 135–142

    2010

  14. [14]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. 2017. Semi-Supervised Classification with Graph Convolutional Networks. In5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net. https://openreview.net/forum?id=SJU4ayYgl

    2017

  15. [15]

    Yuhan Li, Peisong Wang, Zhixun Li, Jeffrey Xu Yu, and Jia Li. 2024. ZeroG: Investigating Cross-dataset Zero-shot Transferability in Graphs. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD 2024, Barcelona, Spain, August 25-29, 2024, Ricardo Baeza-Yates and Francesco Bonchi (Eds.). ACM, 1725–1735

    2024

  16. [16]

    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, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net

    2024

  17. [17]

    Yu, and Chuan Shi

    Jiawei Liu, Cheng Yang, Zhiyuan Lu, Junze Chen, Yibo Li, Mengmei Zhang, Ting Bai, Yuan Fang, Lichao Sun, Philip S. Yu, and Chuan Shi. 2023. Towards Graph Foundation Models: A Survey and Beyond.arXiv preprint arXiv:2310.11829 (2023)

    Pith/arXiv arXiv 2023

  18. [18]

    McAuley, Christopher Targett, Qinfeng Shi, and Anton van den Hengel

    Julian J. McAuley, Christopher Targett, Qinfeng Shi, and Anton van den Hengel

  19. [19]

    Andrew Kachites McCallum, Kamal Nigam, Jason Rennie, and Kristie Seymore

  20. [20]

    Smith, and Mike Lewis

    Ofir Press, Noah A. Smith, and Mike Lewis. 2022. Train Short, Test Long: Attention with Linear Biases Enables Input Length Extrapolation. InThe Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net

    2022

  21. [21]

    Benedek Rozemberczki, Carl Allen, and Rik Sarkar. 2021. Multi-Scale attributed node embedding.J. Complex Networks9, 2 (2021)

    2021

  22. [22]

    Prithviraj Sen, Galileo Namata, Mustafa Bilgic, Lise Getoor, Brian Gallagher, and Tina Eliassi-Rad. 2008. Collective Classification in Network Data.AI Mag.29, 3 (2008), 93–106

    2008

  23. [23]

    Lianze Shan, Jitao Zhao, Dongxiao He, Siqi Liu, Jiaxu Cui, and Weixiong Zhang

  24. [24]

    Oleksandr Shchur, Maximilian Mumme, Aleksandar Bojchevski, and Stephan Günnemann. 2018. Pitfalls of Graph Neural Network Evaluation.arXiv preprint arXiv:1811.05868(2018)

    Pith/arXiv arXiv 2018

  25. [25]

    Kanatsoulis, Jure Leskovec, and Bruno Ribeiro

    Yangyi Shen, Jincheng Zhou, Beatrice Bevilacqua, Joshua Robinson, Charilaos I. Kanatsoulis, Jure Leskovec, and Bruno Ribeiro. 2025. Zero-Shot Generalization of GNNs over Distinct Attribute Domains. InForty-second International Conference on Machine Learning, ICML 2025, Vancouver, BC, Canada, July 13-19, 2025 (Proceed- ings of Machine Learning Research, Vo...

    2025

  26. [26]

    Arnab Sinha, Zhihong Shen, Yang Song, Hao Ma, Darrin Eide, Bo-June Paul Hsu, and Kuansan Wang. 2015. An Overview of Microsoft Academic Service (MAS) and Applications. InProceedings of the 24th International Conference on World Wide Web Companion, WWW 2015, Florence, Italy, May 18-22, 2015 - Companion Volume, Aldo Gangemi, Stefano Leonardi, and Alessandro ...

    2015

  27. [27]

    Mingchen Sun, Kaixiong Zhou, Xin He, Ying Wang, and Xin Wang. 2022. GPPT: Graph Pre-training and Prompt Tuning to Generalize Graph Neural Networks. InKDD ’22: The 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Washington, DC, USA, August 14 - 18, 2022, Aidong Zhang and Huzefa Rangwala (Eds.). ACM, 1717–1727

    2022

  28. [28]

    Jie Tang, Jimeng Sun, Chi Wang, and Zi Yang. 2009. Social influence analysis in large-scale networks. InProceedings of the 15th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Paris, France, June 28 - July 1, 2009, John F. Elder IV, Françoise Fogelman-Soulié, Peter A. Flach, and Mohammed Javeed Zaki (Eds.). ACM, 807–816

    2009

  29. [29]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is All you Need. InAdvances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, Isabelle Guyon, Ulrike von Luxburg, ...

    2017

  30. [30]

    Kai Wang and Siqiang Luo. 2024. Towards Graph Foundation Models: The Perspective of Zero-shot Reasoning on Knowledge Graphs.arXiv preprint arXiv:2410.12609(2024)

    Pith/arXiv arXiv 2024

  31. [31]

    Song Wang, Yushun Dong, Xiao Huang, Chen Chen, and Jundong Li. 2022. FAITH: Few-Shot Graph Classification with Hierarchical Task Graphs. InProceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI 2022, Vienna, Austria, 23-29 July 2022, Luc De Raedt (Ed.). ijcai.org, 2284–2290

    2022

  32. [32]

    Shuo Wang, Bokui Wang, Zhixiang Shen, Boyan Deng, and Zhao Kang. 2025. Multi-Domain Graph Foundation Models: Robust Knowledge Transfer via Topol- ogy Alignment. InForty-second International Conference on Machine Learning, ICML 2025, Vancouver, BC, Canada, July 13-19, 2025 (Proceedings of Machine Learning Research, Vol. 267), Aarti Singh, Maryam Fazel, Dan...

    2025

  33. [33]

    Yi Wang, Jitao Zhao, Dongxiao He, Jia Li, Yuxiao Huang, and Zhiyong Feng. 2026. Topology-Aware Feature Sorting Enables Universal Modeling on Homophilic and Heterophilic Graphs. InProceedings of the ACM Web Conference 2026, WWW 2026, Dubai, United Arab Emirates, originally scheduled for April 13-17, 2026, rescheduled for June 29 - July 3, 2026. ACM, 475–486

    2026

  34. [34]

    Chawla, Chuxu Zhang, and Yanfang Ye

    Zehong Wang, Zheyuan Liu, Tianyi Ma, Jiazheng Li, Zheyuan Zhang, Xingbo Fu, Yiyang Li, Zhengqing Yuan, Wei Song, Yijun Ma, Qingkai Zeng, Xiusi Chen, Jianan Zhao, Jundong Li, Meng Jiang, Pietro Lio, Nitesh V. Chawla, Chuxu Zhang, and Yanfang Ye. 2025. Graph Foundation Models: A Comprehensive Survey. arXiv preprint arXiv:2505.15116(2025)

    Pith/arXiv arXiv 2025

  35. [35]

    Chawla, Chuxu Zhang, and Yanfang Ye

    Zehong Wang, Zheyuan Zhang, Tianyi Ma, Nitesh V. Chawla, Chuxu Zhang, and Yanfang Ye. 2025. Towards Graph Foundation Models: Learning Generalities Across Graphs via Task-Trees. InProceedings of the 42nd International Conference on Machine Learning, ICML 2025, Vol. 267. PMLR

    2025

  36. [36]

    Lianghao Xia and Chao Huang. 2026. AnyGraph: Graph Foundation Model in the Wild. InFindings of the Association for Computational Linguistics, ACL 2026, San Diego, California, United States, July 2-7, 2026. Association for Computational Linguistics, 882–896

    2026

  37. [37]

    Pinar Yanardag and S. V. N. Vishwanathan. 2015. Deep Graph Kernels. InProceed- ings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Sydney, NSW, Australia, August 10-13, 2015, Longbing Cao, Chengqi Zhang, Thorsten Joachims, Geoffrey I. Webb, Dragos D. Margineantu, and Graham Williams (Eds.). ACM, 1365–1374

    2015

  38. [38]

    Xingtong Yu, Zechuan Gong, Chang Zhou, Yuan Fang, and Hui 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, WWW 2025, Sydney, NSW, Australia, 28 April 2025- 2 May 2025. ACM, 1142–1153

    2025

  39. [39]

    Xingtong Yu, Chang Zhou, Yuan Fang, and Xinming Zhang. 2024. Text-Free Multi-domain Graph Pre-training: Toward Graph Foundation Models.arXiv preprint arXiv:2405.13934(2024)

    Pith/arXiv arXiv 2024

  40. [40]

    Haonan Yuan, Qingyun Sun, Junhua Shi, Xingcheng Fu, Bryan Hooi, Jianxin Li, and Philip S. Yu. 2025. How Much Can Transfer? BRIDGE: Bounded Multi- Domain Graph Foundation Model with Generalization Guarantees. InForty- second International Conference on Machine Learning, ICML 2025, Vancouver, BC, Canada, July 13-19, 2025 (Proceedings of Machine Learning Res...

    2025

  41. [41]

    Jitao Zhao, Di Jin, Meng Ge, Lianze Shan, Xin Wang, Dongxiao He, and Zhiyong Feng. 2024. FUG: Feature-Universal Graph Contrastive Pre-training for Graphs with Diverse Node Features. InAdvances in Neural Information Processing Systems 37: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15,...

    2024

  42. [42]

    Jitao Zhao, Yi Wang, Yawen Li, Dongxiao He, Di Jin, Zhiyong Feng, and Weixiong Zhang. 2026. Towards Graph Foundation Model: Node Feature Transfer Invariant Modeling on General Graphs. InProceedings of the ACM Web Conference 2026, WWW 2026, Dubai, United Arab Emirates, originally scheduled for April 13-17, 2026, rescheduled for June 29 - July 3, 2026, Haki...

    2026

  43. [43]

    Bronstein, and Jian Tang

    Jianan Zhao, Zhaocheng Zhu, Mikhail Galkin, Hesham Mostafa, Michael M. Bronstein, and Jian Tang. 2025. Fully-inductive Node Classification on Arbi- trary Graphs. InProceedings of the 13th International Conference on Learning Representations, ICLR 2025. A Notations and Definitions The notations used in this paper and their descriptions are summa- rized in ...

    2025

  44. [2000]

    Automating the Construction of Internet Portals with Machine Learning. Inf. Retr.3, 2 (2000), 127–163

    2000

  45. [2015]

    Image-Based Recommendations on Styles and Substitutes. InProceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Retrieval, Santiago, Chile, August 9-13, 2015, Ricardo Baeza-Yates, Mounia Lalmas, Alistair Moffat, and Berthier A. Ribeiro-Neto (Eds.). ACM, 43–52

    2015

  46. [2024]

    InThe Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024

    Towards Foundation Models for Knowledge Graph Reasoning. InThe Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net

    2024

  47. [2026]

    InProceedings of the ACM Web Conference 2026, WWW 2026, Dubai, United Arab Emirates, originally scheduled for April 13-17, 2026, rescheduled for June 29 - July 3, 2026

    LEDA: Latent Semantic Distribution Alignment for Multi-domain Graph Pre-training. InProceedings of the ACM Web Conference 2026, WWW 2026, Dubai, United Arab Emirates, originally scheduled for April 13-17, 2026, rescheduled for June 29 - July 3, 2026. ACM, 614–625

    2026

This paper was first reviewed by grok-4.5 on July 31, 2026.