Pith. sign in

REVIEW 3 major objections 6 minor 42 references

Towards Text-free Graph Foundation Models: Rethinking Multi-Domain Graph Contrastive Learning

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A domain-aware contrastive objective for multi-domain graph pre-training, claiming transfer gains of up to 19.33% in accuracy.

desk verdict MDGCL is a practical, well-tested recipe for text-free multi-domain graph pre-training, but its central claim that the new contrastive objective captures domain differences is undermined by a token shortcut that the ablations never rule out. read the letter →

arxiv 2506.22510 v1 pith:BV5H3255 submitted 2025-06-26 cs.CL cs.AI

Zihao Zhao , Xinlong Zhai , Jinyu Yang , Chuan Shi This is my paper
classification cs.CLcs.AI
keywords multi-domaingraphpre-trainingcontrastivelearningtext-freegraphscross-domaintransferdomaintokensattentionfoundationmodelsfew-shot
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

Most graph pre-training methods treat all negative pairs alike, even when the graphs come from different domains; this paper argues that such domain-blind contrastive learning suppresses the very differences that make cross-domain transfer hard. The proposed framework, MDGCL, instead pre-trains a GNN to decide whether two sampled subgraphs come from the same domain, so the encoder must represent what distinguishes domains. It adds a per-domain token, computed as the sum of that domain's node features, and a downstream attention over those tokens to transfer relevant domain knowledge to a new target graph. The paper reports that this recipe outperforms prior single-domain and multi-domain graph pre-training methods on five benchmark graphs, with peak gains of 19.33% in accuracy and 19.13% in Macro-F1, and that gains grow as more source domains are added.

What carries the argument

The load-bearing object is the same-domain discrimination objective: a cross-entropy loss on the prediction, made by a projection head over the pooled output of a GNN, of whether two random-walk subgraphs merged into one graph came from the same domain (Equation 4). To make that task solvable, the framework attaches a domain token to each subgraph pair, where each source domain's token is the sum of all its node features (Equation 3), and connects the two tokens inside the merged graph so the GNN can read domain-level context while processing local structure. On the downstream side, a domain-level attention mechanism computes the similarity between each target node and all source domain tokens and adds the attention-weighted token mixture to the node's features (Equation 6). These pieces work together: the discrimination loss shapes the encoder to keep domains distinct, the tokens carry global domain information, and the attention passes that information to an unseen target domain.

What would settle it

Train MDGCL with domain tokens whose entries are randomly permuted or replaced by a constant vector while keeping everything else fixed. If downstream accuracy stays close to the reported numbers, the discrimination objective can be solved from the token alone and the claimed structural transfer is not the source of the gains; if accuracy collapses, the encoder is genuinely using subgraph content.

Watch

Extended reading notes

Core claim

MDGCL's central claim is that the dominant contrastive pre-training objective, built for a single domain, is the wrong tool for multi-domain graph pre-training because it asks the model to keep same-domain and cross-domain subgraphs equally far apart. Treating every cross-domain pair as a plain negative pair hides domain identity. MDGCL replaces that with a binary same-domain discrimination task: positive pairs are two subgraphs from the same source graph, negative pairs are subgraphs from different source graphs, and the GNN must predict which is which on the merged pair. The paper further claims that fixed domain tokens, each the sum-pool of a source domain's node features, give the encoder a compact global signal per domain, and that a downstream attention mechanism that mixes source-domain tokens into target node features enables fine-grained transfer. On five benchmark graphs, one-shot and few-shot node and graph classification experiments show MDGCL consistently outperforming the comparison methods, and a data-ablation study shows accuracy increasing with the number of pre-training source domains.

Load-bearing premise

The method assumes that when the GNN solves the same-domain discrimination task, it is learning transferable properties of graph structure and features, rather than simply reading the fixed domain token attached to each subgraph; the paper does not analyze this shortcut.

Editorial extensions

If this is right

  • Existing single-domain contrastive objectives should be re-examined in multi-domain settings; if MDGCL is right, domain-aware negatives are not a detail but the main driver of cross-domain transfer.
  • Text-free graph foundation models can be pre-trained without text or LLM supervision, using only subgraph sampling, domain tokens, and a binary domain-discrimination head.
  • Performance scales with the number of source domains, so adding more domains during pre-training should continue to improve downstream transfer rather than saturate or hurt.
  • The same pre-trained encoder transfers to both homophilic and heterophilic target graphs, suggesting domain-aware pre-training is more robust to distribution shift than domain-blind alignment.
  • The number of training samples depends only on the number of subgraphs and domains, not on graph size, so the method remains efficient for large graphs.

Reading between the lines

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

  • The fixed sum-pooled token is a strong prior, so a learned or normalized domain prototype might transfer even better when source domains vary in size or feature scale.
  • Because the objective is a binary same-domain prediction, it implicitly assumes each dataset is one coherent domain; real-world graphs may contain multiple latent domains, and a clustering or multi-prototype extension could capture that.
  • The reported scaling with the number of domains suggests testing on much larger and more heterogeneous corpora; if the trend holds, MDGCL-style objectives could replace text bridging in graph foundation models.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes MDGCL, a multi-domain graph contrastive learning framework for text-free graphs. In pre-training, it samples subgraphs from multiple source domains, connects each subgraph to fixed domain tokens, and trains a GNN to predict whether two subgraphs come from the same domain. In the downstream stage, it augments target node features with a domain-level attention mechanism that aggregates source-domain token embeddings. The method is evaluated on five benchmark datasets for few-shot node and graph classification, with additional experiments on heterophilic graphs, and consistently outperforms eight baselines.

Significance. If the proposed mechanism is genuinely responsible for the reported gains, MDGCL offers a simple, efficient, and scalable approach to multi-domain graph pre-training. The paper's strengths include a clear problem formulation, consistent experimental improvements across multiple tasks and settings, a reasonable baseline set, and an efficiency analysis. However, the central causal claim--that the same-domain discrimination objective captures transferable domain-specific structure--is not yet supported because the objective can be solved through the fixed domain tokens, and the ablations do not isolate the proposed objective from standard contrastive learning. These gaps are addressable with additional experiments, so the contribution is potentially valuable but needs stronger evidence.

major comments (3)
  1. [Section 4.2.2, Eq. (3) and Section 4.2.3, Eq. (4)] The pre-training label y_ij is fully determined by the fixed domain tokens t_i and t_j: in the merged graph, every subgraph node is connected to its domain token and the two tokens are directly connected. A two-layer GCN can therefore solve the same-domain discrimination task by reading the token embeddings, without learning transferable graph structure. The paper provides no probe, token-masking, or feature-importance analysis to rule out this shortcut, so the claim that the encoder learns domain-specific structural differences remains unsubstantiated.
  2. [Table 4, Section 5.4] The ablation does not isolate the proposed pre-training objective. MDGCL-V1 still uses the same-domain cross-entropy objective; it only removes tokens and the downstream attention. No variant replaces the same-domain discrimination loss with a standard contrastive loss (e.g., GraphCL or SimGRACE style) under the same architecture and data. Without such a control, the comparison with GCOPE and SAMGPT does not support the conclusion in Section 5.2 that 'capturing domain-specific differences is more important than aligning features and structures'.
  3. [Section 4.3.1, Eq. (6) and Section 4.3.2, Eq. (8)] The downstream domain attention uses the same source-domain tokens as feature augmentation, and the paper does not test whether the pre-trained encoder is necessary for the reported gains. A control that applies the same attention mechanism to an untrained or randomly initialized GNN (or to a GNN pre-trained with a standard contrastive objective) would clarify whether the gains come from the proposed pre-training or from the token-based feature enhancement alone.
minor comments (6)
  1. [Abstract] The sentence 'namely MDGCL.In the pre-training stage' is missing a space after 'MDGCL.'; it should read 'MDGCL. In the pre-training stage'.
  2. [Table 10] The table formatting is broken: the entries 'SAMGPT0.52 11 588' and 'MDGCL0.48 1438' lack proper column separators, making the efficiency comparison hard to read.
  3. [Section 5.2] The text says MDGCL 'outperforms state-of-the-art significantly,' but no statistical significance tests are reported; confidence intervals or p-values for the comparisons would strengthen the claim.
  4. [Figure 3] The data-ablation plot shows no error bars or variance information, so it is unclear whether the upward trend with more source domains is reliable across seeds.
  5. [General] The paper does not include a code availability statement; releasing the code would improve reproducibility of the reported experimental results.
  6. [Section 6] The statement that MDGCL is 'the first to design the contrastive learning strategy for multi-domain pre-training scenarios' is stronger than the related-work discussion supports; a more cautious formulation with explicit comparison to prior domain-aware contrastive methods would be appropriate.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: MDGCL's claims are empirical evaluations against external baselines, and no equation-level reduction or fitted-parameter-as-prediction pattern was found.

full rationale

The paper's central claim is that MDGCL's multi-domain contrastive pre-training objective improves cross-domain transfer. The pre-training loss in Eq. 4 is a self-supervised binary classification of whether two subgraphs come from the same domain, trained on source graphs; downstream performance is then measured on held-out target-domain tasks using standard few-shot fine-tuning. The domain tokens in Eq. 3 are computed deterministically from source features and are used both in pre-training and in the downstream attention mechanism, but the downstream accuracy numbers are not derived from those tokens by construction. No parameter is fitted to the downstream test labels and then reported as a prediction; the comparison against GCOPE, SAMGPT, GraphCL, and end-to-end baselines is an external empirical benchmark. The potential concern that the pre-training objective could be solved by recognizing the fixed domain tokens rather than learning transferable graph structure is a plausible shortcut or validity issue, but it is not a circularity in the sense of the target result being equivalent to the inputs by construction; the paper does not claim a formal derivation that reduces to its own assumptions. There is also no load-bearing self-citation: the cited works by the same authors are not used to justify the core pre-training objective or to forbid alternative designs. Therefore, the appropriate circularity score is 0.

Assumptions & free parameters 7 free parameters · 4 assumptions · 1 invented entities

The central claim rests on several domain-level assumptions that are plausible but not explicitly validated. The most consequential is the shortcut risk: the fixed domain tokens give the GNN an easy path to solve the pre-training objective without learning transferable structure. The SVD comparability assumption is also unexamined. The hyperparameters are standard choices, and none are fitted to the target result.

free parameters (7)
  • Unified feature dimension (d~) = 50
    Hyperparameter set in Section 5.1 and Appendix C; used for SVD dimension reduction of all source and target features.
  • Hidden dimension = 256
    GNN hidden size used for all pre-training methods (Section 5.1).
  • Number of subgraphs per source domain (K) = 50
    Set in Section 5.1 and Appendix C; controls the number of training pairs.
  • Random walk length = 50
    Set in Appendix C; analyzed in Appendix F.1.
  • Learning rate = 1e-4
    Optimizer setting for pre-training (Section 5.1).
  • Attention heads = 2
    Multi-head attention in downstream domain-attention (Section 5.1).
  • Number of negative pairs per domain pair (N) = Not stated in main text
    Introduced in Appendix E and swept in Appendix F.2 (40-65), but the default used for Tables 1-9 is not given.
assumptions (4)
  • domain assumption Each benchmark dataset can be treated as a single coherent domain with a learnable domain-level representation.
    Section 3 defines each source graph G_i as a domain; the method assumes domain identity is a meaningful and stable property of the data.
  • domain assumption After per-domain SVD projection to 50 dimensions, features from different domains live in a space where dot products (Eq. 6) are meaningful for attention.
    The domain-attention mechanism (Eq. 6) computes attention scores via dot products between target node features and source domain tokens; if the per-domain SVD spaces are not aligned, these scores could be arbitrary. The paper does not justify cross-domain comparability.
  • domain assumption Random-walk subgraphs preserve domain-specific discriminative information.
    The pre-training pairs are built from random-walk sampled subgraphs (Section 4.2.2); the method assumes these subgraphs contain enough signal to determine the source domain, which may fail for very small or poorly sampled subgraphs.
  • ad hoc to paper The same-domain discrimination task is not trivially solved by the fixed domain tokens, so the GNN must learn transferable structure.
    The domain tokens (Eq. 3) are connected to every node in the merged graph, giving the model a direct signal of domain identity. The paper provides no analysis (e.g., token-free pre-training or probing) to confirm the encoder learns more than token matching.
invented entities (1)
  • Domain token t_i = sum(X~i)
    purpose: A fixed vector per source domain, connected into pre-training graphs and used as keys/values in downstream attention to transfer domain-level knowledge.
    This is a new model component introduced in Eq. 3. It has no falsifiable handle outside the reported experiments; its utility is only demonstrated on the paper's benchmarks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Text-free Graph Foundation Models: Rethinking Multi-Domain Graph Contrastive Learning." pith.science (2026). https://pith.science/paper/BV5H3255

@misc{pith2026250622510,
  author       = {Pith},
  title        = {Pith review of: Towards Text-free Graph Foundation Models: Rethinking Multi-Domain Graph Contrastive Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BV5H3255}},
  note         = {Machine review of arXiv:2506.22510}
}
read the original abstract

Foundation models have achieved great success in natural language processing (NLP) and computer vision (CV). Their success largely stems from the ability to integrate multi-domain knowledge in pre-training and transfer it to target domains. Considering graph data, especially graphs without textual features, is ubiquitous in real-world applications such as social networks and recommendation systems, some researchers have attempted to extend this paradigm to the graph field, aiming to construct graph foundation models. However, unlike CV and NLP, there are huge gaps among the semantics and properties of graphs in different domains, while current works still adopt traditional contrastive pre-training strategies designed in the single-domain scenario, which regard contrastive samples from different domains as equivalent. From experimental investigations, we discovered that inherent domain-specific differences prevent these strategies from effectively absorbing knowledge from different domains to generate informative representations. In this paper, we propose a novel multi-domain pre-training and cross-domain transfer framework, namely MDGCL.In the pre-training stage, we design a contrastive learning strategy to substantially recognize and capture domain differences, and introduce domain tokens to encode domain-level global information. In the downstream stage, we introduce a domain attention mechanism to enable fine-grained domain knowledge transfer. Extensive experiments on five benchmark datasets have demonstrated that our method outperforms state-of-the-art significantly, with the maximum improvement of 19.33\% on accuracy and 19.13\% on Macro-F1 score.

Figures

Figures reproduced from arXiv: 2506.22510 by the authors.

Figure 1
Figure 1. Visualization for initial features and learned representations of SAMGPT, GCOPE and [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Overall framework of MDGCL. The upper part represents the pre-training stage, and the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Data ablation study with dif￾ferent numbers of source domains, fix￾ing Cora as the target domain. A prominent feature of foundation models is the ability to absorb knowledge from more different data domains and obtain better downstream performance. Thus, in this sec￾tion, we investigate the impact of the source domain num￾bers on the downstream target domain performance of our method and other text-free graph founda… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Analysis of the walk length. We evaluate the performance of MDGCL with different walk lengths, as shown in [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: Analysis of the hyperparameter N. As mentioned in Section E, the number of negative samples during the pre-training stage is controlled by the hyperparameter N. We evaluate the performance of MDGCL with different numbers of negative samples, as shown in [PITH_FULL_IMA…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 25 canonical work pages

  1. [1]

    Llama: Open and efficient foundation language models

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timo- thée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023

    arXiv 2023

  2. [2]

    All nlp tasks are generation tasks: A general pretraining framework

    Zhengxiao Du, Yingyi Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. All nlp tasks are generation tasks: A general pretraining framework. arXiv: Computation and Language,arXiv: Computation and Language, Mar 2021. 13

    work page 2021

  3. [3]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023

    arXiv 2023

  4. [4]

    Understanding and improving visual prompting: A label-mapping perspective

    Aochuan Chen, Yuguang Yao, Pin-Yu Chen, Yihua Zhang, and Sijia Liu. Understanding and improving visual prompting: A label-mapping perspective. Nov 2022

    work page 2022

  5. [5]

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and GeoffreyE. Hinton. A simple framework for contrastive learning of visual representations. Cornell University - arXiv,Cornell University - arXiv, Feb 2020

    work page 2020

  6. [6]

    Imagenet-21k pretraining for the masses

    Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik-Manor. Imagenet-21k pretraining for the masses. arXiv: Computer Vision and Pattern Recognition,arXiv: Computer Vision and Pattern Recognition, Apr 2021

    work page 2021

  7. [7]

    Selectivity drives productivity: Efficient dataset pruning for enhanced transfer learning

    Yihua Zhang, Yimeng Zhang, Aochuan Chen, Jinghan Jia, Jiancheng Liu, Gaowen Liu, Mingyi Hong, Shiyu Chang, and Sijia Liu. Selectivity drives productivity: Efficient dataset pruning for enhanced transfer learning. Oct 2023

    work page 2023

  8. [8]

    Hhan: Comprehensive infectious disease source tracing via heterogeneous hypergraph neural network

    Qiang He, Yunting Bao, Hui Fang, Yuting Lin, and Hao Sun. Hhan: Comprehensive infectious disease source tracing via heterogeneous hypergraph neural network. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 291–299, 2025

    work page 2025

Show all 42 references
  1. [9]

    Social recommendation via graph-level counterfactual augmentation

    Yinxuan Huang, Ke Liang, Yanyi Huang, Xiang Zeng, Kai Chen, and Bin Zhou. Social recommendation via graph-level counterfactual augmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 334–342, 2025

    2025

  2. [10]

    Hypergraph convolutional network for user-oriented fairness in recommender systems

    Zhongxuan Han, Chaochao Chen, Xiaolin Zheng, Li Zhang, and Yuyuan Li. Hypergraph convolutional network for user-oriented fairness in recommender systems. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 9...

    2024

  3. [11]

    Graph contrastive learning with augmentations

    Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. Graph contrastive learning with augmentations. Advances in neural information processing systems, 33:5812–5823, 2020

    2020

  4. [12]

    Graphprompt: Unifying pre-training and downstream tasks for graph neural networks

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

    2023

  5. [13]

    All in one: Multi-task prompt- ing for graph neural networks

    Xiangguo Sun, Hong Cheng, Jia Li, Bo Liu, and Jihong Guan. All in one: Multi-task prompt- ing for graph neural networks. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2120–2131, 2023

    2023

  6. [14]

    Universal prompt tun- ing for graph neural networks

    Taoran Fang, Yunchao Zhang, Yang Yang, Chunping Wang, and Lei Chen. Universal prompt tun- ing for graph neural networks. Advances in Neural Information Processing Systems, 36:52464– 52489, 2023

    2023

  7. [15]

    Gppt: Graph pre-training and prompt tuning to generalize graph neural networks

    Mingchen Sun, Kaixiong Zhou, Xin He, Ying Wang, and Xin Wang. Gppt: Graph pre-training and prompt tuning to generalize graph neural networks. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 1717–1727, 2022

    2022

  8. [16]

    Simgrace: A simple framework for graph contrastive learning without data augmentation

    Jun Xia, Lirong Wu, Jintao Chen, Bozhen Hu, and Stan Z Li. Simgrace: A simple framework for graph contrastive learning without data augmentation. In Proceedings of the ACM web conference 2022, pages 1070–1079, 2022

    2022

  9. [17]

    Higpt: Heterogeneous graph language model

    Jiabin Tang, Yuhao Yang, Wei Wei, Lei Shi, Long Xia, Dawei Yin, and Chao Huang. Higpt: Heterogeneous graph language model. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2842–2853, 2024

    2024

  10. [18]

    One for all: Towards training one graph model for all classification tasks

    Hao Liu, Jiarui Feng, Lecheng Kong, Ningyue Liang, Dacheng Tao, Yixin Chen, and Muhan Zhang. One for all: Towards training one graph model for all classification tasks. arXiv preprint arXiv:2310.00149, 2023. 14

    2023 arXiv

  11. [19]

    Opengraph: Towards open graph foundation models

    Lianghao Xia, Ben Kao, and Chao Huang. Opengraph: Towards open graph foundation models. arXiv preprint arXiv:2403.01121, 2024

    2024 arXiv

  12. [20]

    Augmenting low-resource text classification with graph-grounded pre-training and prompting

    Zhihao Wen and Yuan Fang. Augmenting low-resource text classification with graph-grounded pre-training and prompting. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 506–516, 2023

    2023

  13. [21]

    Learning on large-scale text-attributed graphs via variational inference

    Jianan Zhao, Meng Qu, Chaozhuo Li, Hao Yan, Qian Liu, Rui Li, Xing Xie, and Jian Tang. Learning on large-scale text-attributed graphs via variational inference. arXiv preprint arXiv:2210.14709, 2022

    2022 arXiv

  14. [22]

    All in one and one for all: A simple yet effective method towards cross-domain graph pretraining

    Haihong Zhao, Aochuan Chen, Xiangguo Sun, Hong Cheng, and Jia Li. All in one and one for all: A simple yet effective method towards cross-domain graph pretraining. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 4443–4454, 2024

    2024

  15. [23]

    Samgpt: Text-free graph foundation model for multi-domain pre-training and cross-domain adaptation

    Xingtong Yu, Zechuan Gong, Chang Zhou, Yuan Fang, and Hui Zhang. Samgpt: Text-free graph foundation model for multi-domain pre-training and cross-domain adaptation. arXiv preprint arXiv:2502.05424, 2025

    2025 arXiv

  16. [24]

    Collective classification in network data

    Prithviraj Sen, Galileo Namata, Mustafa Bilgic, Lise Getoor, Brian Galligher, and Tina Eliassi- Rad. Collective classification in network data. AI Magazine, page 93, Jul 2017

    2017

  17. [25]

    Multi-scale attributed node embedding

    Benedek Rozemberczki, Carl Allen, and Rik Sarkar. Multi-scale attributed node embedding. Journal of Complex Networks, 9(2):cnab014, 2021

    2021

  18. [26]

    Revisiting semi-supervised learning with graph embeddings

    Zhilin Yang, William Cohen, and Ruslan Salakhudinov. Revisiting semi-supervised learning with graph embeddings. In International conference on machine learning, pages 40–48. PMLR, 2016

    2016

  19. [27]

    Pitfalls of graph neural network evaluation

    Oleksandr Shchur, Maximilian Mumme, Aleksandar Bojchevski, and Stephan Günnemann. Pitfalls of graph neural network evaluation. arXiv: Learning,arXiv: Learning, Nov 2018

    2018

  20. [28]

    Image-based recommendations on styles and substitutes

    Julian McAuley, Christopher Targett, Qinfeng Shi, and Anton Van Den Hengel. Image-based recommendations on styles and substitutes. In Proceedings of the 38th international ACM SIGIR conference on research and development in information retrieval, pages 43–52, 2015

    2015

  21. [29]

    Semi-supervised classification with graph convolutional networks

    Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016

    2016 arXiv

  22. [30]

    Inductive representation learning on large graphs

    Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. Advances in neural information processing systems, 30, 2017

    2017

  23. [31]

    Graph attention networks

    Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. arXiv preprint arXiv:1710.10903, 2017

    2017 arXiv

  24. [32]

    Graph transformer networks

    Seongjun Yun, Minbyul Jeong, Raehyun Kim, Jaewoo Kang, and Hyunwoo J Kim. Graph transformer networks. Advances in neural information processing systems, 32, 2019

    2019

  25. [33]

    G. W. Stewart. On the early history of the singular value decomposition. SIAM Review, page 551–566, Dec 1993

    1993

  26. [34]

    The pagerank citation ranking: Bringing order to the web

    Lawrence Page, Sergey Brin, Rajeev Motwani, and Terry Winograd. The pagerank citation ranking: Bringing order to the web. Technical report, Stanford infolab, 1999

    1999

  27. [35]

    Beyond low-frequency information in graph convolutional networks

    Deyu Bo, Xiao Wang, Chuan Shi, and Huawei Shen. Beyond low-frequency information in graph convolutional networks. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 3950–3957, 2021

    2021

  28. [36]

    Is homophily a necessity for graph neural networks? Learning,Learning, Jun 2021

    Yao Ma, Xiaorui Liu, Neil Shah, and Jiliang Tang. Is homophily a necessity for graph neural networks? Learning,Learning, Jun 2021. 15

    2021

  29. [37]

    The heterophilic graph learning hand- book: Benchmarks, models, theoretical analysis, applications and challenges

    Sitao Luan, Chenqing Hua, Qincheng Lu, Liheng Ma, Lirong Wu, Xinyu Wang, Minkai Xu, Xiao-Wen Chang, Doina Precup, Rex Ying, et al. The heterophilic graph learning hand- book: Benchmarks, models, theoretical analysis, applications and challenges. arXiv preprint arXiv:2407.09618, 2024

    2024 arXiv

  30. [38]

    A critical look at the evaluation of gnns under heterophily: Are we really making progress? arXiv preprint arXiv:2302.11640, 2023

    Oleg Platonov, Denis Kuznedelev, Michael Diskin, Artem Babenko, and Liudmila Prokhorenkova. A critical look at the evaluation of gnns under heterophily: Are we really making progress? arXiv preprint arXiv:2302.11640, 2023

    2023 arXiv

  31. [39]

    Harnessing language model for cross-heterogeneity graph knowledge transfer

    Jinyu Yang, Ruijia Wang, Cheng Yang, Bo Yan, Qimin Zhou, Yang Juan, and Chuan Shi. Harnessing language model for cross-heterogeneity graph knowledge transfer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 13026–13034, 2025

    2025

  32. [40]

    Text-space graph foundation models: Comprehensive benchmarks and new insights

    Zhikai Chen, Haitao Mao, Jingzhe Liu, Yu Song, Bingheng Li, Wei Jin, Bahare Fatemi, Anton Tsitsulin, Bryan Perozzi, Hui Liu, et al. Text-space graph foundation models: Comprehensive benchmarks and new insights. arXiv preprint arXiv:2406.10727, 2024

    2024 arXiv

  33. [41]

    Towards foundation models for knowledge graph reasoning

    Mikhail Galkin, Xinyu Yuan, Hesham Mostafa, Jian Tang, and Zhaocheng Zhu. Towards foundation models for knowledge graph reasoning. arXiv preprint arXiv:2310.04562, 2023

    2023 arXiv

  34. [42]

    Gft: Graph foundation model with transferable tree vocabulary.Advances in Neural Information Processing Systems, 37:107403–107443, 2024

    Zehong Wang, Zheyuan Zhang, Nitesh Chawla, Chuxu Zhang, and Yanfang Ye. Gft: Graph foundation model with transferable tree vocabulary.Advances in Neural Information Processing Systems, 37:107403–107443, 2024. 16

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.