pith. machine review for the scientific record. sign in

arxiv: 2605.09993 · v1 · submitted 2026-05-11 · 💻 cs.LG

Recognition: 1 theorem link

· Lean Theorem

Learning Graph Foundation Models on Riemannian Graph-of-Graphs

Haokun Liu , Zezhong Ding , Xike Xie
Authors on Pith no claims yet

Pith reviewed 2026-05-12 03:59 UTC · model grok-4.3

classification 💻 cs.LG
keywords graph foundation modelsRiemannian graph-of-graphsmulti-scale subgraphsdomain generalizationRiemannian manifoldsgraph machine learningstructural scale adaptation
0
0 comments X

The pith

Treating structural scale as a variable via multi-scale Riemannian graph-of-graphs reduces structural domain generalization error in graph foundation models.

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

Existing graph foundation models use fixed-hop subgraph sampling that imposes a fixed receptive field, leading to mismatches on tasks needing heterogeneous structural contexts at unknown scales. R-GFM addresses this by constructing a multi-scale graph-of-graphs from subgraphs sampled at different hop distances and learning geometry-adaptive representations on Riemannian manifolds. Theoretical analysis establishes a reduction in structural domain generalization error relative to fixed-scale approaches, while experiments report state-of-the-art results with relative improvements reaching 49 percent on downstream tasks. A sympathetic reader cares because this promises more flexible, general-purpose graph models that adapt to scale variations without requiring task-specific sampling adjustments.

Core claim

By constructing a multi-scale GoG over subgraphs sampled at different hop distances and learning geometry-adaptive representations from Riemannian manifolds, R-GFM treats structural scale as a first-class citizen and reduces structural domain generalization error compared to fixed-scale GFMs.

What carries the argument

The Riemannian Graph-of-Graphs (GoG) that over-samples subgraphs at multiple hop distances to enable geometry-adaptive learning on manifolds.

If this is right

  • Graph foundation models can better generalize across domains with varying structural scales.
  • Pretraining incorporates scale as a learnable geometric property rather than a fixed parameter.
  • Downstream task performance improves substantially, as shown by up to 49% relative gains.
  • Structural domain shifts become less problematic due to the adaptive receptive fields.

Where Pith is reading between the lines

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

  • Applying similar manifold-based multi-scale constructions could benefit foundation models in other structured data types.
  • Computational efficiency on very large graphs remains an open question that would need empirical validation.
  • Integration with existing GFM architectures might allow incremental adoption of the scale-adaptive component.

Load-bearing premise

That constructing and learning on a multi-scale graph-of-graphs over Riemannian manifolds will reliably capture heterogeneous and unknown structural contexts without introducing new sources of overfitting or computational intractability on large graphs.

What would settle it

A direct comparison where fixed-scale GFMs achieve lower or equal structural domain generalization error than R-GFM on scale-heterogeneous graph datasets would falsify the reduction claim.

Figures

Figures reproduced from arXiv: 2605.09993 by Haokun Liu, Xike Xie, Zezhong Ding.

Figure 1
Figure 1. Figure 1: Motivation of R-GFM. Existing GFMs rely on fixed-hop sampling and static manifolds, causing representation mismatch. R-GFM constructs an adaptive-hop GoG and uses a router to dynamically choose the best-fit geometry for better matching. Despite this progress, existing GFMs adopt a common de￾sign choice: they rely on fixed-hop subgraph sampling dur￾ing pretraining, where each node or subgraph is encoded usi… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of R-GFM. ⃝1 Calculate the coefficient of variation (CV) to quantify the node degree distribution, which is used to determine the candidate Riemannian expert set. ⃝2 Sample adaptive-hop subgraphs for each training node v (or an edge (u, v) in link prediction task (Zhang & Chen, 2018)) with a adaptive-hop strategy (Section 3.2). ⃝3 Encode sampled subgraphs {G (i) v } K i=1 to obtain subgraph embedd… view at source ↗
Figure 3
Figure 3. Figure 3: Robustness to structural and attribute perturbations. Photo Cora Chameleon Texas 10 15 20 25 30 35 40 45 50 55 60 65 70 Accuracy (%) Full Hyperbolic only Euclidean GoG GoG edges: chain Hyperspherical only w/o GoG pooling Stage-1 only [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Ablation results on four representative datasets (Photo, Cora, Chameleon, Texas) under 1-shot node classification. sistently achieves the best accuracy across all perturbation levels and exhibits the slowest degradation rate as p in￾creases. For example, at p = 0.5, R-GFM retains 46.91% (a drop of 5.76% points from p = 0), while the strongest baseline drops to around 41.2%. Node Masking. We randomly mask n… view at source ↗
Figure 5
Figure 5. Figure 5: Performance scaling with hop number K: a naive K￾hop extension of GCOPE vs. R-GFM under matched settings. 4.5. Scalability (RQ4) To answer RQ4, we test scalability from two perspectives: (1) scaling with structural scales (i.e., increasing K-hop), and (2) scaling with graph size (i.e., larger graphs), reporting both accuracy and resource costs. Scaling with Hop Number (K) We study how perfor￾mance scales w… view at source ↗
Figure 7
Figure 7. Figure 7: Effect of dataset scale and expert number. Cora Texas Photo Chameleon 10 15 20 25 30 35 40 45 50 55 60 65 70 Accuracy (%) Ours Top-1 Top-2 Top-3 Top-4 Top-5 [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Effect of gate sparsity (Top-m) across datasets. edges. We therefore control the edge budget by fixing the retained-edge ratio β ∈ [0, 1] (β = 2Bedge K(K−1) ). In Figure 6a, performance peaks around β ≈ 0.6. Therefore, we Bedge = 0.6 × K(K−1) 2 . More analysis is in Appendix H. Candidate Expert NumberM. Figure 7a shows that our dynamic selection strategy increases the candidate expert number as more datase… view at source ↗
Figure 9
Figure 9. Figure 9: Scalability comparison in terms of training time with respect to the number of nodes [PITH_FULL_IMAGE:figures/full_fig_p019_9.png] view at source ↗
read the original abstract

Graph foundation models (GFMs), pretrained on massive graph data, have transformed graph machine learning by supporting general-purpose reasoning across diverse graph tasks and domains. Existing GFMs pretrained with fixed-hop subgraph sampling impose a fixed receptive field, causing scale mismatch on diverse tasks, which often require heterogeneous and unknown structural contexts beyond a fixed sampling scale. We propose R-GFM, a Riemannian Graph-of-Graphs (GoG) based foundation model, that treats structural scale as a first-class citizen in modeling. R-GFM constructs a multi-scale GoG over-sampled subgraphs at different hop distances and learns geometry-adaptive representations from Riemannian manifolds. Theoretical analysis shows that R-GFM reduces structural domain generalization error compared to fixed-scale GFMs. Experiments on various datasets demonstrate that R-GFM achieves state-of-the-art performance, with up to a 49% relative improvement on downstream tasks. Our code is available at https://github.com/USTC-DataDarknessLab/R-GFM.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes R-GFM, a Riemannian Graph-of-Graphs (GoG) foundation model for graphs. It constructs multi-scale GoGs from over-sampled subgraphs at varying hop distances and learns geometry-adaptive representations on Riemannian manifolds. The central claims are a theoretical reduction in structural domain generalization error relative to fixed-scale GFMs, plus empirical SOTA results with up to 49% relative improvement on downstream tasks.

Significance. If the theoretical reduction holds under the stated manifold and sampling assumptions and the empirical gains prove robust to controls, the work would meaningfully advance graph foundation models by elevating structural scale to a learnable, geometry-aware component rather than a fixed hyperparameter. This directly targets a known limitation of current GFMs and could improve cross-domain generalization.

major comments (2)
  1. [Abstract / Theoretical analysis] Abstract and theoretical analysis section: the claim that R-GFM reduces structural domain generalization error is asserted without an explicit derivation, definition of the error term, or statement of the manifold curvature and subgraph-sampling assumptions; this is load-bearing for the central contribution and prevents verification of the reduction.
  2. [Experiments] Experiments section: the reported 49% relative improvement lacks error bars, statistical significance tests, or explicit description of controls for baseline hyperparameter tuning and subgraph sampling; without these, the SOTA claim cannot be assessed as load-bearing evidence.
minor comments (2)
  1. [Methods] Notation for the multi-scale GoG construction and Riemannian operations should be introduced with a single consistent diagram or table early in the methods section to aid readability.
  2. [Experiments] The code repository link is provided but the manuscript does not specify which exact experimental configurations (hyperparameters, dataset splits) are reproduced by the released code.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We sincerely thank the referee for their constructive and insightful comments, which help us strengthen the clarity and rigor of the manuscript. We address each major comment point by point below and commit to revisions that directly respond to the concerns raised.

read point-by-point responses

  1. Referee: [Abstract / Theoretical analysis] Abstract and theoretical analysis section: the claim that R-GFM reduces structural domain generalization error is asserted without an explicit derivation, definition of the error term, or statement of the manifold curvature and subgraph-sampling assumptions; this is load-bearing for the central contribution and prevents verification of the reduction.

    Authors: We acknowledge that while Section 4 of the manuscript contains the theoretical analysis under the Riemannian manifold setting, the presentation in the abstract and the main theoretical section could be more explicit to facilitate verification. The analysis relies on the structural domain generalization error defined as the expected discrepancy in node representations across domains due to scale mismatch, with the reduction shown via a bound that depends on the learnable curvature and multi-scale sampling. In the revised manuscript, we will (i) add a concise statement of the key assumptions (constant negative curvature on the manifold and uniform over-sampling of subgraphs at hop distances {1,2,...,K}) directly in the abstract, (ii) include the formal definition of the error term at the beginning of Section 4, and (iii) expand the proof to provide a complete, self-contained derivation of the error reduction relative to fixed-scale GFMs. These changes will make the central theoretical claim fully verifiable without altering the underlying result. revision: yes

  2. Referee: [Experiments] Experiments section: the reported 49% relative improvement lacks error bars, statistical significance tests, or explicit description of controls for baseline hyperparameter tuning and subgraph sampling; without these, the SOTA claim cannot be assessed as load-bearing evidence.

    Authors: We agree that the empirical evaluation would benefit from greater statistical rigor and transparency. The 49% relative improvement is computed from mean performance across datasets, but variability and controls are not reported. In the revised version, we will: (i) report mean performance with standard deviation error bars computed over five independent runs using different random seeds for initialization and sampling; (ii) add statistical significance tests (paired t-tests or Wilcoxon signed-rank tests with p-values) comparing R-GFM against each baseline; and (iii) include a new appendix subsection that explicitly documents the hyperparameter search grids and tuning protocols applied to all baselines, as well as the precise subgraph sampling procedure (including the set of hop distances, over-sampling ratios, and how the multi-scale GoG is constructed). These additions will allow readers to assess the robustness of the SOTA claims. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper presents R-GFM as a novel architecture that constructs multi-scale Riemannian GoG representations to address fixed-scale limitations in existing GFMs. The central theoretical claim—that this reduces structural domain generalization error—is stated as following from the multi-scale manifold construction itself rather than from any fitted parameter, self-referential definition, or load-bearing self-citation. No equations or steps are shown that rename a fitted quantity as a prediction or import uniqueness via prior author work. The modeling decisions are introduced explicitly as design choices for heterogeneous structural contexts, keeping the derivation self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The model rests on the standard assumption that Riemannian manifolds provide a suitable geometry for learning adaptive graph representations; the multi-scale GoG construction is introduced without independent evidence beyond the claimed empirical gains.

axioms (1)
  • domain assumption Riemannian manifolds can be used to model geometry-adaptive representations on graphs
    Invoked to justify placing multi-scale subgraphs on curved spaces for scale adaptation.
invented entities (1)
  • Riemannian Graph-of-Graphs (GoG) no independent evidence
    purpose: To represent multi-scale structural contexts as first-class objects in the foundation model
    New construction proposed in the paper; no independent falsifiable evidence provided in abstract.

pith-pipeline@v0.9.0 · 5468 in / 1360 out tokens · 67619 ms · 2026-05-12T03:59:55.760370+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

107 extracted references · 107 canonical work pages

  1. [1]

    Hinton , title =

    Ting Chen and Simon Kornblith and Mohammad Norouzi and Geoffrey E. Hinton , title =. ICML , pages =

    work page

  2. [2]

    CoRR , volume =

    Jinze Zhao and Peihao Wang and Zhangyang Wang , title =. CoRR , volume =

    work page

  3. [3]

    NeurIPS , year=

    DuetGraph: Coarse-to-Fine Knowledge Graph Reasoning with Dual-Pathway Global-Local Fusion , author=. NeurIPS , year=

    work page

  4. [4]

    NeurIPS , year =

    Nicolas Keriven , title =. NeurIPS , year =

    work page

  5. [5]

    ICML , pages =

    Mingze Dong and Yuval Kluger , title =. ICML , pages =

    work page

  6. [6]

    ICML , pages =

    Denis Lukovnikov and Asja Fischer , title =. ICML , pages =

    work page

  7. [7]

    ICLR , year =

    Uri Alon and Eran Yahav , title =. ICLR , year =

    work page

  8. [8]

    NeurIPS , year =

    Kai Han and Yunhe Wang and Jianyuan Guo and Yehui Tang and Enhua Wu , title =. NeurIPS , year =

    work page

  9. [9]

    Langley , title =

    P. Langley , title =. 2000 , pages =

    work page 2000

  10. [10]

    T. M. Mitchell. The Need for Biases in Learning Generalizations. 1980

    work page 1980

  11. [11]

    M. J. Kearns , title =

    work page

  12. [12]

    Machine Learning: An Artificial Intelligence Approach, Vol. I. 1983

    work page 1983

  13. [13]

    R. O. Duda and P. E. Hart and D. G. Stork. Pattern Classification. 2000

    work page 2000

  14. [14]

    Suppressed for Anonymity , author=

    work page

  15. [15]

    Newell and P

    A. Newell and P. S. Rosenbloom. Mechanisms of Skill Acquisition and the Law of Practice. Cognitive Skills and Their Acquisition. 1981

    work page 1981

  16. [16]

    A. L. Samuel. Some Studies in Machine Learning Using the Game of Checkers. IBM Journal of Research and Development. 1959

    work page 1959

  17. [17]

    Collective Classification in Network Data , journal =

    Prithviraj Sen and Galileo Namata and Mustafa Bilgic and Lise Getoor and Brian Gallagher and Tina Eliassi. Collective Classification in Network Data , journal =

    work page

  18. [18]

    Cohen and Ruslan Salakhutdinov , title =

    Zhilin Yang and William W. Cohen and Ruslan Salakhutdinov , title =

    work page

  19. [19]

    8th International Conference on Learning Representations,

    Pei, Hongbin and Wei, Bingzhe and Chang, Kevin Chen-Chuan and Lei, Yu and Yang, Bo , journal=. 8th International Conference on Learning Representations,

    work page

  20. [20]

    McAuley, Julian and Targett, Christopher and Shi, Qinfeng and Van Den Hengel, Anton , booktitle =

    work page

  21. [21]

    Pitfalls of Graph Neural Network Evaluation

    Pitfalls of graph neural network evaluation , author=. arXiv preprint arXiv:1811.05868 , year=

    work page Pith review arXiv

  22. [22]

    Graph Attention Networks , journal =

    Petar Velickovic and Guillem Cucurull and Arantxa Casanova and Adriana Romero and Pietro Li. Graph Attention Networks , journal =

    work page

  23. [23]

    Deyu Bo and Xiao Wang and Chuan Shi and Huawei Shen , title =

    work page

  24. [24]

    Jiong Zhu and Yujun Yan and Lingxiao Zhao and Mark Heimann and Leman Akoglu and Danai Koutra , title =

    work page

  25. [25]

    Hamilton and Pietro Li

    Petar Velickovic and William Fedus and William L. Hamilton and Pietro Li. Deep Graph Infomax , booktitle =

    work page

  26. [26]

    NeurIPS , year =

    Yuning You and Tianlong Chen and Yongduo Sui and Ting Chen and Zhangyang Wang and Yang Shen , title =. NeurIPS , year =

    work page

  27. [27]

    NeurIPS , year =

    Teng Xiao and Zhengyu Chen and Zhimeng Guo and Zeyang Zhuang and Suhang Wang , title =. NeurIPS , year =

    work page

  28. [28]

    NeurIPS , year =

    Teng Xiao and Huaisheng Zhu and Zhengyu Chen and Suhang Wang , title =. NeurIPS , year =

    work page

  29. [29]

    Mingchen Sun and Kaixiong Zhou and Xin He and Ying Wang and Xin Wang , title =

    work page

  30. [30]

    Zemin Liu and Xingtong Yu and Yuan Fang and Xinming Zhang , title =

    work page

  31. [31]

    Xingtong Yu and Zhenghao Liu and Yuan Fang and Zemin Liu and Sihong Chen and Xinming Zhang , title =

    work page

  32. [32]

    Ruiyi Fang and Bingheng Li and Zhao Kang and Qiuhao Zeng and Nima Hosseini Dashtbayaz and Ruizhi Pu and Charles Ling and Boyu Wang , title =

    work page

  33. [33]

    Haihong Zhao and Aochuan Chen and Xiangguo Sun and Hong Cheng and Jia Li , title =

    work page

  34. [34]

    Xingtong Yu and Zechuan Gong and Chang Zhou and Yuan Fang and Hui Zhang , title =

    work page

  35. [35]

    Shuo Wang and Bokui Wang and Zhixiang Shen and Boyan Deng and Zhao Kang , title =

    work page

  36. [36]

    Yu , title =

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

    work page

  37. [37]

    Zhe Xu and Yuzhong Chen and Qinghai Zhou and Yuhang Wu and Menghai Pan and Hao Yang and Hanghang Tong , title =

    work page

  38. [38]

    NeurIPS , year =

    Taoran Fang and Yunchao Zhang and Yang Yang and Chunping Wang and Lei Chen , title =. NeurIPS , year =

    work page

  39. [39]

    Xiangguo Sun and Hong Cheng and Jia Li and Bo Liu and Jihong Guan , title =

    work page

  40. [40]

    Hanchen Wang and Defu Lian and Ying Zhang and Lu Qin and Xuemin Lin , title =

    work page

  41. [41]

    Learning Mixed-Curvature Representations in Product Spaces , booktitle =

    Albert Gu and Frederic Sala and Beliz Gunel and Christopher R. Learning Mixed-Curvature Representations in Product Spaces , booktitle =

    work page

  42. [42]

    Jiezhong Qiu and Qibin Chen and Yuxiao Dong and Jing Zhang and Hongxia Yang and Ming Ding and Kuansan Wang and Jie Tang , title =

    work page

  43. [43]

    Hamilton and Zhitao Ying and Jure Leskovec , title =

    William L. Hamilton and Zhitao Ying and Jure Leskovec , title =. NeuIPS , pages =

    work page

  44. [44]

    NeuIPS , year=

    Link prediction based on graph neural networks , author=. NeuIPS , year=

    work page

  45. [45]

    Jie Chen and Tengfei Ma and Cao Xiao , title =

    work page

  46. [46]

    Representation Learning with Contrastive Predictive Coding , journal =

    A. Representation Learning with Contrastive Predictive Coding , journal =

    work page

  47. [47]

    Prasanna , title =

    Hanqing Zeng and Hongkuan Zhou and Ajitesh Srivastava and Rajgopal Kannan and Viktor K. Prasanna , title =. 8th International Conference on Learning Representations,

    work page

  48. [48]

    Gel , title =

    Zhiwei Zhen and Yuzhou Chen and Murat Kantarcioglu and Yulia R. Gel , title =

    work page

  49. [49]

    Constant Curvature Graph Convolutional Networks , booktitle =

    Gregor Bachmann and Gary B. Constant Curvature Graph Convolutional Networks , booktitle =

    work page

  50. [50]

    Feinberg and Joseph Gomes and Caleb Geniesse and Aneesh S

    Zhenqin Wu and Bharath Ramsundar and Evan N. Feinberg and Joseph Gomes and Caleb Geniesse and Aneesh S. Pappu and Karl Leswing and Vijay S. Pande , title =. CoRR , year =

    work page

  51. [51]

    Schoenholz and Patrick F

    Justin Gilmer and Samuel S. Schoenholz and Patrick F. Riley and Oriol Vinyals and George E. Dahl , title =

    work page

  52. [52]

    Hamilton and Jure Leskovec , title =

    Rex Ying and Ruining He and Kaifeng Chen and Pong Eksombatchai and William L. Hamilton and Jure Leskovec , title =

    work page

  53. [53]

    LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation , booktitle =

    Xiangnan He and Kuan Deng and Xiang Wang and Yan Li and Yong. LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation , booktitle =

    work page

  54. [54]

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

    Michael Sejr Schlichtkrull and Thomas N. Kipf and Peter Bloem and Rianne van den Berg and Ivan Titov and Max Welling , title =

    work page

  55. [55]

    Yaguang Li and Rose Yu and Cyrus Shahabi and Yan Liu , title =

    work page

  56. [56]

    Yu , title =

    Zonghan Wu and Shirui Pan and Fengwen Chen and Guodong Long and Chengqi Zhang and Philip S. Yu , title =. CoRR , year =

    work page

  57. [57]

    Jie Zhou and Ganqu Cui and Shengding Hu and Zhengyan Zhang and Cheng Yang and Zhiyuan Liu and Lifeng Wang and Changcheng Li and Maosong Sun , title =

    work page

  58. [58]

    Yu , title =

    Yixin Liu and Ming Jin and Shirui Pan and Chuan Zhou and Yu Zheng and Feng Xia and Philip S. Yu , title =

    work page

  59. [59]

    Kaveh Hassani and Amir Hosein Khas Ahmadi , title =

    work page

  60. [60]

    Zhenyu Hou and Xiao Liu and Yukuo Cen and Yuxiao Dong and Hongxia Yang and Chunjie Wang and Jie Tang , title =

    work page

  61. [61]

    Jacob Devlin and Ming

    work page

  62. [62]

    Masked Autoencoders Are Scalable Vision Learners , booktitle =

    Kaiming He and Xinlei Chen and Saining Xie and Yanghao Li and Piotr Doll. Masked Autoencoders Are Scalable Vision Learners , booktitle =

    work page

  63. [63]

    Chawla and Chuxu Zhang and Yanfang Ye , title =

    Zehong Wang and Zheyuan Liu and Tianyi Ma and Jiazheng Li and Zheyuan Zhang and Xingbo Fu and Yiyang Li and Zhengqing Yuan and Wei Song and Yijun Ma and Qingkai Zeng and Xiusi Chen and Jianan Zhao and Jundong Li and Meng Jiang and Pietro Lio and Nitesh V. Chawla and Chuxu Zhang and Yanfang Ye , title =. CoRR , year =

    work page

  64. [64]

    CoRR , year =

    Chen Cai and Yusu Wang , title =. CoRR , year =

    work page

  65. [65]

    Recipe for a General, Powerful, Scalable Graph Transformer , booktitle =

    Ladislav Ramp. Recipe for a General, Powerful, Scalable Graph Transformer , booktitle =

    work page

  66. [66]

    NeurIPS , year =

    Shurui Gui and Xiner Li and Limei Wang and Shuiwang Ji , title =. NeurIPS , year =

    work page

  67. [67]

    Boshen Shi and Yongqing Wang and Fangda Guo and Jiangli Shao and Huawei Shen and Xueqi Cheng , title =

    work page

  68. [68]

    NeuIPS , pages =

    Maximilian Nickel and Douwe Kiela , title =. NeuIPS , pages =

    work page

  69. [69]

    Hyperbolic Graph Convolutional Neural Networks , booktitle =

    Ines Chami and Zhitao Ying and Christopher R. Hyperbolic Graph Convolutional Neural Networks , booktitle =

    work page

  70. [70]

    Pande and Jure Leskovec , title =

    Weihua Hu and Bowen Liu and Joseph Gomes and Marinka Zitnik and Percy Liang and Vijay S. Pande and Jure Leskovec , title =

    work page

  71. [71]

    Kipf and Max Welling , title =

    Thomas N. Kipf and Max Welling , title =. CoRR , year =

    work page

  72. [72]

    CoRR , year =

    Andrea Cavallo and Claas Grohnfeldt and Michele Russo and Giulio Lovisotto and Luca Vassio , title =. CoRR , year =

    work page

  73. [73]

    Yang Liu and Xiang Ao and Zidi Qin and Jianfeng Chi and Jinghua Feng and Hao Yang and Qing He , title =

    work page

  74. [74]

    Yu Wang and Yuying Zhao and Neil Shah and Tyler Derr , title =

    work page

  75. [75]

    Borgwardt , title =

    Nino Shervashidze and Pascal Schweitzer and Erik Jan van Leeuwen and Kurt Mehlhorn and Karsten M. Borgwardt , title =. J. Mach. Learn. Res. , pages =

    work page

  76. [76]

    Neural Networks , pages =

    Hui Shen and Huifang Ma and Jiyuan Sun and Yuwei Gao and Zhixin Li , title =. Neural Networks , pages =

    work page

  77. [77]

    Jackson and Pietro Tebaldi , title =

    Francis Bloch and Matthew O. Jackson and Pietro Tebaldi , title =. Soc. Choice Welf. , pages =

    work page

  78. [78]

    Deeper Insights Into Graph Convolutional Networks for Semi-Supervised Learning , booktitle =

    Qimai Li and Zhichao Han and Xiao. Deeper Insights Into Graph Convolutional Networks for Semi-Supervised Learning , booktitle =

    work page

  79. [79]

    CoRR , year =

    Lianghao Xia and Chao Huang , title =. CoRR , year =

    work page

  80. [80]

    Provably Powerful Graph Neural Networks for Directed Multigraphs , booktitle =

    B. Provably Powerful Graph Neural Networks for Directed Multigraphs , booktitle =

    work page

Showing first 80 references.