Pith. sign in

REVIEW 4 major objections 5 minor 44 references

AgentGFM: A Graph Foundation Model with Node-Agent Information-Flow Control

T0 review · 4 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read AgentGFM's central claim is that treating each node as an agent that controls its own information flow—deciding which neighbors to hear, which signal channel to use, and when to stop—under one shared policy, transfers across unseen graphs…

desk verdict Real conceptual step in graph foundation models, but the SOTA claim rests on tiny margins and test-set-tuned hyperparameters; worth serious refereeing, not acceptance as-is. read the letter →

arxiv 2607.26533 v1 pith:DUEPVH3S submitted 2026-07-29 cs.LG cs.AI

Jingbo Cui , Jitao Zhao , Di Jin , Dongxiao He This is my paper · ORCID
classification cs.LGcs.AI
keywords graphfoundationmodelscross-domaintransferinformation-flowcontrolnodeagentsadaptivepropagationone-shotlearningheterophilousgraphshaltingpolicy
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 proposes that a graph foundation model can transfer across unseen graphs if each node is treated as an agent that decides, at each step, where to receive information, which frequency channel to use, and when to stop, all under a single shared trainable policy. It claims this node-level information-flow control outperforms fixed propagation schemes in cross-domain one-shot classification. If true, it means adaptive propagation decisions can generalize rather than being tied to a specific graph distribution, which matters for real graphs whose local structures vary widely.

What carries the argument

The central object is the shared information-flow policy that each node executes locally. It produces edge-wise source-reception strengths γ, edge-wise channel gates ρ interpolating low- and high-frequency messages, and a node-wise halting score κ whose discrete decision is trained with a straight-through estimator. A forwarding budget b and prediction–observation feedback cues ε condition these decisions, while a feedback-gated correction T merges the predicted and actual observations; the whole policy is trained with masked attribute reconstruction, prediction–observation alignment, and three regularizers that discourage indiscriminate reception, excessive transmission, and overlong rollouts.

What would settle it

Replace the truncated-SVD feature alignment in Section 4.1 with a fixed random projection of the same dimension, keep everything else identical, and rerun the leave-one-dataset-out one-shot protocol; if accuracy does not drop on heterophilous targets, the alignment is not carrying the transfer, and if it collapses, the premise is confirmed.

Watch

Extended reading notes

Core claim

AgentGFM establishes that information-flow control can be realized as node-agent decision making under a shared end-to-end trainable policy. Each node runs a predict–act–observe–correct loop: it predicts the contextual information its current state should receive, acts through three decisions—source reception, signal-channel selection, and gain-aware node-wise halting—receives an observation from the graph, and corrects its state by the discrepancy between prediction and observation. The paper reports that this design achieves the best average rank of 1.10 across ten node-classification datasets, the best average rank of 1.00 across five graph-classification datasets, and strong results on large-scale graphs, with particularly clear gains on heterophilous graphs where fixed propagation rules absorb misleading neighbor signals.

Load-bearing premise

The load-bearing premise is that a single truncated-SVD linear projection maps very different node feature spaces into one shared 50-dimensional space that still preserves enough structural signal for the shared policy to act on; if that alignment loses transfer-relevant information, the whole cross-domain pipeline degrades.

Editorial extensions

If this is right

  • Cross-domain one-shot transfer improves on heterophilous graphs, where fixed message-passing schemes are most likely to absorb misleading neighbor signals.
  • Node-specific halting saves computation because most nodes stop early, and the method never constructs dense higher-order neighborhoods, keeping complexity at O(RT|E|d) for bounded rollout rounds and horizons.
  • The same shared policy transfers across node-level and graph-level tasks, so the mechanism is not tied to a particular task head.
  • The prediction–observation discrepancy provides a self-supervised training signal that lets the model correct node states without downstream labels.
  • The learned halting and reception distributions differ across datasets, confirming that the policy produces node-specific trajectories rather than a uniform propagation rule.

Reading between the lines

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

  • Beyond the paper, a direct test of the weakest link would be to replace the truncated-SVD feature alignment with a learned per-domain encoder; if transfer accuracy holds, the alignment is not the limiting factor, and if it collapses, the SVD step is doing more work than the paper credits.
  • Since all node agents share one policy but never exchange explicit messages about their decisions, a natural extension is to add lightweight agent-to-agent coordination or a global context token that conditions the shared policy on graph-level statistics.
  • The information-flow decisions are trained only through reconstruction and alignment losses; adding a downstream-aware objective during pretraining might sharpen source selection and halting behavior on target graphs.
  • Because over 82% of nodes halt by the second rollout step, an adaptive per-graph horizon could cut the rollout cost further without hurting accuracy, especially on large graphs.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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. This paper proposes AgentGFM, a graph foundation model that treats each node as an agent executing a shared policy for node-specific information-flow control. The model aligns node features via truncated SVD, then runs a recurrent predict–act–observe–correct process in which each node decides source reception, low-/high-frequency channel selection, and gain-aware halting. The encoder is pretrained on source graphs with masked reconstruction, prediction-observation alignment, and three regularization objectives, then evaluated under leave-one-dataset-out cross-domain 1-shot protocols on node, graph, and large-scale benchmarks. The paper claims state-of-the-art average ranking and consistent gains over existing GFMs, supported by ablations and a mechanism-level case study.

Significance. If the empirical claims hold, AgentGFM would be a meaningful step toward topology-adaptive, transferable graph representations: node-specific propagation decisions under a single shared policy are a plausible alternative to fixed message-passing schemes, and the formulation is clean and mechanism-level. The paper provides a detailed description of the policy components, comprehensive coverage of ten node datasets and five graph datasets, and ablations isolating each design choice, plus explicit hyperparameter settings and an honest discussion of computational overhead. However, the central SOTA claim rests on comparisons with small margins, without significance testing, and on a rollout horizon selected on the target benchmarks; these issues must be addressed before the claim is convincing.

major comments (4)
  1. [§5.3.4 (Table 4)] The default maximum rollout horizon Tmax = 5 is selected by comparing average rank on the same ten target datasets used for evaluation. This is test-set hyperparameter tuning: the configuration that produced the reported results was chosen to maximize performance on the benchmarks it is then evaluated on. This can inflate the reported gains and weakens the transfer claim. Please either fix Tmax a priori based on validation sources, or present Table 4 as a separate sensitivity analysis and run the main comparison with a pre-specified value; if such a value is chosen from source-domain validation alone, details of that protocol should be reported.
  2. [§4.1 and Appendix B.2] The feature-alignment step described as 'aligns their node features into a shared space using truncated SVD' is never specified: the paper does not state what matrix is decomposed, how the 50-dimensional shared space is computed, or whether the SVD is fit on source graphs only or on a combination of source and target graphs. If target-graph features are included in the SVD fit, the leave-one-dataset-out protocol is violated; if not, the procedure must be described in full to be reproducible. Given that this alignment is the first step of the cross-domain pipeline and is load-bearing for transfer, please provide the exact algorithm and a sensitivity analysis of the shared dimension.
  3. [§5.1.2, Tables 1–2] The evaluation reports only means and standard deviations over 100 support-set samples. Since all methods are evaluated under identical support-set draws, paired significance tests are the appropriate tool, and several headline wins are small relative to the reported variability (e.g., PubMed 52.96 vs 52.16 for GraphAny; Chameleon 30.54 vs 30.32 for GCOPE; Squirrel 24.20 vs 23.77 for GCOPE; MUTAG 59.52 vs 58.30 for R-GFM). Without such tests, the statements 'obtains the best performance on nine datasets' and 'ranks first on all five datasets' are not supported. At minimum, report paired bootstrap or permutation test results and effect sizes; also clarify how 'competitive' on CiteSeer (44.32 vs 51.25 for R-GFM) is assessed given that the margin is 6.93 points in favor of the baseline.
  4. [§4.3, §4.6, Eq. (28)] The 'prediction–observation' feedback is an internal self-consistency loop: the predicted observation b_o_i and the actual observation o_i are both generated by the same network (the predicted one by f_pred and the actual one by the model's own rollout), and L_pred in Eq. (28) trains the former to match the latter. The discrepancy in Eq. (23) therefore does not measure error with respect to any external ground truth, so the claim that it 'reflects the mismatch between the node's current expectation and the information returned by the graph' is only true in a model-internal sense. This is not circular for the transfer claim, which is tested on held-out datasets, but it should be acknowledged and its effect on the mechanism should be analyzed, e.g., by checking whether the correction gate in Eq. (24) is sensitive to the discrepancy or simply learns a constant.
minor comments (5)
  1. [§5.3.1] In the variant name, 'Gain-A ware Halt' should be 'Gain-Aware Halting'.
  2. [§5.2.3 (Table 3)] The large-scale evaluation uses a different pretraining corpus (eight fixed source graphs) than the leave-one-out protocol of §5.2.1; while each target is still unseen, the source set is not the same across the two tables, so the results are not directly comparable and this discrepancy should be noted.
  3. [Appendix B.2 / Table 10] The values of λ_high and θ_0 are reported in Table 10 but the paper does not describe how they were chosen; a brief explanation for at least these two hyperparameters would improve reproducibility.
  4. [§4.4, Eq. (10)] The source-reception score γ^t_{u→i} is a sigmoid and is thus not normalized over the incoming neighbors; the paper should clarify whether this is intentional and how it interacts with the normalized weights α in Eq. (12).
  5. [§5.3.2] The claim that 82.42% of nodes halt by the second rollout step would be more informative if accompanied by the standard deviation across datasets, since the distribution is described as varying by dataset.

Circularity Check

2 steps flagged · score 4.0 of 10

Transfer claim is tested on held-out datasets, but the internal prediction–observation discrepancy is trained to vanish (Eq. 28) and the default Tmax=5 is selected on the test targets (Table 4).

  1. self definitional [Section 4.6–4.7 (Eqs. 23–28)]
    "The predicted observation is computed as bo_i = f_pred(h_i) ... δ_i = |o_i − bo_i| ∈ R^d ... L_pred = (1/|V|) Σ_{v_i∈V} (1 − bo_i^T sg(o_i) / (||bo_i||_2 ||o_i||_2 + ε))."

    Eq. (28) optimizes bo_i to match o_i via cosine similarity with stop-gradient; at the learned optimum bo_i ≈ o_i, so the discrepancy δ_i in Eq. (23) is driven to zero by construction. The correction gate q_i = σ(f_rel(h_i, bo_i, o_i, δ_i)) (Eq. 24) and the gated fusion of Eqs. (25)–(26) are therefore conditioned on a quantity that the training objective itself eliminates. The predict–observe–correct loop is self-referential: the 'prediction' is fitted to the observation it is meant to be compared with, so the claimed discrepancy-based feedback provides no independent correction signal.

  2. fitted input called prediction [Section 5.3.4, Table 4]
    "Table 4 reports the sensitivity of AgentGFM to the maximum rollout horizon Tmax. Setting Tmax = 5 achieves the best average rank and the best performance on seven of ten datasets ... we use Tmax = 5 as the default setting."

    The default Tmax = 5 used in Tables 1–3 is chosen from accuracy computed on the same ten target datasets shown in Table 4. The reported SOTA transfer numbers therefore use a rollout horizon fitted to those very test targets, so part of the headline result is selected rather than independently predicted. This is a selection-on-test loop: the evaluation configuration is tuned on the evaluation targets before the same targets are used to certify the method.

full rationale

AgentGFM's core construction (Sections 3–4) is a model proposal whose policy, rollouts, and objectives are defined from the model's own states and graph inputs, not from the benchmark numbers. The transfer results use a leave-one-dataset-out protocol with disjoint source and target graphs, so the central cross-domain SOTA claim is not circular by construction. There is no load-bearing self-citation or imported uniqueness theorem; the only author self-citation (TIG [37]) appears as a related-work baseline and does not justify AgentGFM's mechanism. However, two issues warrant a moderate score. First, Eq. (28) trains the predictive observation to match the actual observation, so the prediction–observation discrepancy used for state correction in Eqs. (23)–(25) is minimized by the objective itself, making the feedback loop internally self-consistent rather than an independent correction signal. Second, Section 5.3.4 selects Tmax = 5 using accuracy on the same ten target datasets that later appear in Table 1, and the reported numbers use that selected horizon. These are internal self-consistency and selection-on-test problems, not a derivation that reduces entirely to its inputs, so the paper is only partially circular rather than fully so.

Assumptions & free parameters 7 free parameters · 6 assumptions · 6 invented entities

The method's central claim rests on a set of modeling choices: a single linear feature alignment, a shared policy that still makes node-specific decisions, and self-supervised objectives that are expected to transfer. These are not derived from first principles and are only empirically validated on the reported benchmarks.

free parameters (7)
  • maximum rollout horizon T_max = 5
    Chosen via Table 4 sensitivity analysis on the full benchmark; affects all results.
  • number of interaction rounds R = 2
    Set by default; not extensively ablated.
  • loss weight lambda_mae = 1.00
    Hand-chosen hyperparameter for masked reconstruction.
  • loss weight lambda_pred = 0.50
    Hand-chosen hyperparameter for observation prediction alignment.
  • loss weights lambda_gamma, lambda_mu, lambda_a = 0.001 each
    Hand-chosen regularization coefficients.
  • halt threshold theta_0 = 0.55
    Binary halting threshold in the straight-through estimator.
  • high-frequency channel coefficient lambda_high = 0.30
    Balances low- and high-frequency messages.
assumptions (6)
  • domain assumption Truncated SVD maps node features from all domains into a shared 50-dimensional space while preserving transferable information.
    Used in Section 4.1 before feature processing; if false, cross-domain transfer fails.
  • domain assumption The graph environment returns observations through differentiable message passing over original edges.
    Needed for end-to-end training of the policy.
  • domain assumption Low- and high-frequency signal decomposition captures complementary structural information across graph domains.
    Used in Eq. (13); the paper provides no evidence that this decomposition is universally appropriate.
  • ad hoc to paper Straight-through estimator provides usable gradients for the discrete halting decisions.
    Eq. (20); the paper does not justify this choice against alternatives.
  • domain assumption Self-supervised objectives (masked reconstruction and prediction-observation alignment) learned on source graphs transfer to unseen target graphs with 1-shot labels.
    Underlies the entire pretraining scheme; only tested empirically, not proven.
  • ad hoc to paper The policy can be shared across nodes and graphs while still making node-specific decisions.
    Core design assumption, empirically validated via ablations but not guaranteed.
invented entities (6)
  • Node agent (agent-based information-flow control)
    purpose: Each node makes propagation decisions via a shared policy.
    Introduced in this paper, no external falsifiable handle beyond the paper's own experiments.
  • Predict-act-observe-correct interaction
    purpose: Feedback loop for state correction.
    Model mechanism; no independent evidence.
  • Source-reception strength gamma
    purpose: Controls edge-wise information reception.
    Novel component; tested only via ablation.
  • Signal-channel selection rho
    purpose: Balances low- and high-frequency message content per edge.
    Novel component.
  • Gain-aware halting
    purpose: Node-wise stopping based on predicted information gain.
    Novel component.
  • Forwarding budget b
    purpose: Limits a node's transmitted information.
    Novel component.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AgentGFM: A Graph Foundation Model with Node-Agent Information-Flow Control." pith.science (2026). https://pith.science/paper/DUEPVH3S

@misc{pith2026260726533,
  author       = {Pith},
  title        = {Pith review of: AgentGFM: A Graph Foundation Model with Node-Agent Information-Flow Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DUEPVH3S}},
  note         = {Machine review of arXiv:2607.26533}
}
read the original abstract

Graph Foundation Models (GFMs) aim to learn transferable knowledge from multi-domain graphs and adapt to unseen scenarios. As a fundamental source of relational semantics in graphs, the transferability of topological patterns has long been central to GFM research. However, local structural patterns may vary across graphs and even among nodes within the same graph. Despite such structural variation, most existing GFMs rely on manually designed propagation schemes and apply them to new graphs largely unchanged. Such fixed schemes may not suit the diverse structural patterns of different nodes. This raises a key question: can each node autonomously determine how information should be propagated through the graph? We refer to this capability as information-flow control. Inspired by recent advances in agent technology, we formulate this problem as agent-based decision making and treat each node as an agent. Accordingly, we propose AgentGFM, in which all node agents follow a shared end-to-end trainable policy rather than using independent models. For adaptive information-flow control, each node interacts with the graph through a predict-act-observe-correct process. During the act stage, the node makes three decisions: source reception, signal-channel selection and gain-aware node-wise halting. The resulting observation is compared with the prediction and their discrepancy is used to correct the node state and guide subsequent interactions. Extensive experiments across node-level, graph-level and large-scale transfer scenarios demonstrate the effectiveness of AgentGFM across diverse graph topologies.

Figures

Figures reproduced from arXiv: 2607.26533 by the authors.

Figure 1
Figure 1. Motivation of AgentGFM. Preset propagation may [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overall architecture of AgentGFM. Multi-domain node attributes are first aligned into a shared feature space. Each [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Mechanism-aligned ablation study on cross-domain [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Distribution of node-specific halting steps in Agent [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Node-agent information-flow trace for target node [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 19 canonical work pages

  1. [1]

    Deyu Bo, Xiao Wang, Chuan Shi, and Huawei Shen. 2021. Beyond Low-frequency Information in Graph Convolutional Networks. InThirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Appli- cations of Artificial Intelligence, IAAI 2021, The Eleventh Symposium on Educational Advances in Artificial Intelligence, ...

    doi:10.1609/aaai.v35i5.16514 2021

  2. [2]

    Karsten M Borgwardt, Cheng Soon Ong, Stefan Schönauer, SVN Vishwanathan, Alex J Smola, and Hans-Peter Kriegel. 2005. Protein function prediction via graph kernels.Bioinformatics21, suppl_1 (2005), i47–i56

    2005

  3. [3]

    Eli Chien, Jianhao Peng, Pan Li, and Olgica Milenkovic. 2021. Adaptive Universal Generalized PageRank Graph Neural Network. In9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net. https://openreview.net/forum?id=n6jl7fLxrP

    work page 2021

  4. [4]

    Rajarshi Das, Shehzaad Dhuliawala, Manzil Zaheer, Luke Vilnis, Ishan Durugkar, Akshay Krishnamurthy, Alex Smola, and Andrew McCallum. 2018. Go for a Walk and Arrive at the Answer: Reasoning Over Paths in Knowledge Bases using Reinforcement Learning. In6th International Conference on Learning Representa- tions, ICLR 2018, Vancouver, BC, Canada, April 30 - ...

    work page 2018

  5. [5]

    Paul D Dobson and Andrew J Doig. 2003. Distinguishing enzyme structures from non-enzymes without alignments.Journal of molecular biology330, 4 (2003), 771–783

    work page 2003

  6. [6]

    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...

    work page 2023

  7. [7]

    Zhenyu Hou, Xiao Liu, Yukuo Cen, Yuxiao Dong, Hongxia Yang, Chunjie Wang, and Jie Tang. 2022. GraphMAE: Self-Supervised Masked Graph Autoencoders. In KDD ’22: The 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Washington, DC, USA, August 14 - 18, 2022. ACM, 594–604

    work page 2022

  8. [8]

    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

Show all 44 references
  1. [9]

    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

    2017

  2. [10]

    Kwei-Herng Lai, Daochen Zha, Kaixiong Zhou, and Xia Hu. 2020. Policy-GNN: Aggregation Optimization for Graph Neural Networks. InKDD ’20: The 26th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Virtual Event, CA, USA, August 23-27, 2020, Rajesh Gupta, Yan Liu, Ji...

    2020

  3. [11]

    Haokun Liu, Zezhong Ding, and Xike Xie. 2026. Learning Graph Founda- tion Models on Riemannian Graph-of-Graphs.CoRRabs/2605.09993 (2026). arXiv:2605.09993 doi:10.48550/ARXIV.2605.09993

    2026 doi arXiv

  4. [12]

    Jiawei Liu, Cheng Yang, Zhiyuan Lu, Junze Chen, Yibo Li, Mengmei Zhang, Ting Bai, Yuan Fang, Lichao Sun, Philip S Yu, et al. 2023. Towards graph foundation models: A survey and beyond.arXiv preprint arXiv:2310.11829(2023)

    2023 arXiv

  5. [13]

    Lei Liu, Xingyu Xia, Qianqian Xie, Ben Liu, Wenjie Xu, and Min Peng. 2025. Enhanced Expert Merging for Mixture-of-Experts in Graph Foundation Models. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems

    2025

  6. [14]

    Zemin Liu, Xingtong Yu, Yuan Fang, and Xinming Zhang. 2023. GraphPrompt: Unifying Pre-Training and Downstream Tasks for Graph Neural Networks. In Proceedings of the ACM Web Conference 2023, WWW 2023, Austin, TX, USA, 30 April 2023 - 4 May 2023, Ying Ding, Jie Tang, Juan F. Seq...

    2023 doi

  7. [15]

    Iván Palomares, Carlos Porcel, Luiz Pizzato, Ido Guy, and Enrique Herrera-Viedma

  8. [16]

    Hongbin Pei, Bingzhe Wei, Kevin Chen-Chuan Chang, Yu Lei, and Bo Yang. 2020. Geom-GCN: Geometric Graph Convolutional Networks. In8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net

    2020

  9. [17]

    Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and Max Welling

    Michael Sejr Schlichtkrull, Thomas N. Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and Max Welling. 2018. Modeling Relational Data with Graph Convo- lutional Networks. InThe Semantic Web - 15th International Conference, ESWC 2018, Heraklion, Crete, Greece, June 3-7, 201...

    2018

  10. [18]

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

    2018 arXiv

  11. [19]

    Borgwardt

    Nino Shervashidze, Pascal Schweitzer, Erik Jan van Leeuwen, Kurt Mehlhorn, and Karsten M. Borgwardt. 2011. Weisfeiler-Lehman Graph Kernels.J. Mach. Learn. Res.12 (2011), 2539–2561. doi:10.5555/1953048.2078187

    2011

  12. [20]

    Li Sun, Zhenhao Huang, Suyang Zhou, Qiqi Wan, Hao Peng, and Philip S. Yu

  13. [21]

    Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. 2018. Graph Attention Networks. In6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedin...

    2018

  14. [22]

    Hamilton, Pietro Liò, Yoshua Bengio, and R

    Petar Velickovic, William Fedus, William L. Hamilton, Pietro Liò, Yoshua Bengio, and R. Devon Hjelm. 2019. Deep Graph Infomax. In7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net

    2019

  15. [23]

    Sebastian Vlaic, Theresia Conrad, Christian Tokarski-Schnelle, Mika Gustafsson, Uta Dahmen, Reinhard Guthke, and Stefan Schuster. 2018. ModuleDiscoverer: Identification of regulatory modules in protein-protein interaction networks. Scientific reports8, 1 (2018), 433

    2018

  16. [24]

    Chawla, Chuxu Zhang, and Yanfang Ye

    Zehong Wang, Zheyuan Zhang, Nitesh V. Chawla, Chuxu Zhang, and Yanfang Ye. 2024. GFT: Graph Foundation Model with Transferable Tree Vocabulary. InAdvances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024...

    2024

  17. [25]

    Shu Wu, Yuyuan Tang, Yanqiao Zhu, Liang Wang, Xing Xie, and Tieniu Tan

  18. [26]

    Lianghao Xia, Ben Kao, and Chao Huang. 2024. OpenGraph: Towards Open Graph Foundation Models. InFindings of the Association for Computational Lin- guistics: EMNLP 2024, Miami, Florida, USA, November 12-16, 2024 (Findings of ACL, Vol. EMNLP 2024), Yaser Al-Onaizan, Mohit Bansal...

    2024 doi

  19. [27]

    Wenhan Xiong, Thien Hoang, and William Yang Wang. 2017. DeepPath: A Reinforcement Learning Method for Knowledge Graph Reasoning. InProceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, EMNLP 2017, Copenhagen, Denmark, September 9-11, 2017, Mar...

    2017

  20. [28]

    Fuzhao Xue, Aixin Sun, Hao Zhang, and Eng Siong Chng. 2021. GDPNet: Refining Latent Multi-View Graph for Relation Extraction. InThirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Appli- cations of Artificial Intelligence, ...

    2021 doi

  21. [29]

    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. W...

    2015

  22. [30]

    Zhilin Yang, William Cohen, and Ruslan Salakhudinov. 2016. Revisiting semi- supervised learning with graph embeddings. InInternational conference on ma- chine learning. PMLR, 40–48

    2016

  23. [31]

    Zhe-Rui Yang, Jindong Han, Chang-Dong Wang, and Hao Liu. 2025. GraphLoRA: Structure-Aware Contrastive Low-Rank Adaptation for Cross-Graph Transfer Learning. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining, V.1, KDD 2025, Toronto, ON, Cana...

    2025

  24. [32]

    Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. 2020. Graph Contrastive Learning with Augmentations. InAdvances in Neural Information Processing Systems 33: Annual Conference on Neural In- formation Processing Systems 2020, NeurIPS 2020, Decem...

    2020

  25. [33]

    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 20...

    2025

  26. [34]

    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)

    2024 arXiv

  27. [35]

    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

    2025

  28. [36]

    Haihong Zhao, Aochuan Chen, Xiangguo Sun, Hong Cheng, and Jia 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, KDD 2024, Barcelona...

    2024

  29. [37]

    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 (United Arab Emirates)(WWW ’26). Association for...

    2026

  30. [38]

    Bronstein, and Jian Tang

    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, ICLR 2025, Singapore, April 24-28, 2025. OpenR...

    2025

  31. [39]

    Qifang Zhao, Weidong Ren, Tianyu Li, Hong Liu, Xingsheng He, and Xiaoxiao Xu. 2025. GraphGPT: Generative Pre-trained Graph Eulerian Transformer. In Forty-second International Conference on Machine Learning, ICML 2025, Vancouver, BC, Canada, July 13-19, 2025 (Proceedings of Mac...

    2025

  32. [40]

    Yun Zhu, Yaoke Wang, Haizhou Shi, Zhenshuo Zhang, Dian Jiao, and Siliang Tang. 2024. GraphControl: Adding Conditional Control to Universal Graph Pre- trained Models for Graph Domain Transfer Learning. InProceedings of the ACM Web Conference 2024 (WWW ’24). ACM, 539–550. doi:10...

    2024

  33. [573]

    doi:10.18653/V1/D17-1060

    doi

  34. [2019]

    Session-Based Recommendation with Graph Neural Networks. InThe Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, The Thirty- First Innovative Applications of Artificial Intelligence Conference, IAAI 2019, The Ninth AAAI Symposium on Educational Advances in Ar...

    2019

  35. [2021]

    Fusion69 (2021), 103–127

    Reciprocal Recommender Systems: Analysis of state-of-art literature, chal- lenges and opportunities towards social recommendation.Inf. Fusion69 (2021), 103–127. doi:10.1016/j.inffus.2020.12.001

    2021 doi

  36. [2025]

    RiemannGFM: Learning a Graph Foundation Model from Riemannian Geometry. InProceedings of the ACM on Web Conference 2025, WWW 2025, Sydney, NSW, Australia, 28 April 2025- 2 May 2025, Guodong Long, Michale Blumestein, Yi Chang, Liane Lewin-Eytan, Zi Helen Huang, and Elad Yom-Tov...

    2025

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.