Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

SAMGPT: Text-free Graph Foundation Model for Multi-domain Pre-training and Cross-domain Adaptation

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

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

Pith's one-line read Per-domain structure tokens let one graph model serve many domains.

desk verdict Empirically solid extension of MDGPT, but the 'structure alignment' claim is unsupported—the tokens are per-domain diagonal scalings, not operations on structure. read the letter →

arxiv 2502.05424 v2 pith:NGZBEBEM submitted 2025-02-08 cs.CL cs.AI

Xingtong Yu , Zechuan Gong , Chang Zhou , Yuan Fang , Hui Zhang This is my paper · ORCID
classification cs.CLcs.AI
keywords graphfoundationmodelsmulti-domainpre-trainingcross-domainadaptationstructurealignmentpromptlearningfew-shottext-freegraphs
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 graphs in real Web services carry no usable text, so a universal graph model cannot simply rely on language models to bridge domains. This paper argues that the real obstacle is structural: citation networks, co-purchase graphs, and social networks differ sharply in average degree, path length, and clustering, and ignoring those differences makes multi-domain pre-training interfere instead of synergize. The authors propose SAMGPT, which attaches a small set of learnable structure tokens to each source domain during pre-training and, at adaptation time, tunes holistic plus domain-specific prompts while freezing the encoder. On seven public datasets used in a leave-one-domain-out protocol, it reports the best one-shot node and graph classification accuracy among the compared methods, with ablations attributing the gains to the structure tokens and dual prompts rather than to feature alignment alone. If correct, this makes structure alignment a first-class ingredient for text-free graph foundation models.

What carries the argument

The carrying mechanism is the structure token: a learnable vector $\mathbf{t}^l_{S_i}$ attached to layer $l$ of domain $S_i$, applied as element-wise multiplication to neighbor messages $\mathbf{t}^l_{S_i} \odot \mathbf{h}^{l-1}_u$ inside the aggregation of Eq. (6). Because each token is specific to a domain and layer, it can absorb that domain's characteristic aggregation scale and neighborhood pattern during pre-training. The companion adaptation mechanism is the dual prompt: holistic prompts $\mathbf{p}^l_{\text{hol}}$ inject one learnable vector per layer for the target domain, while specific prompts $\mathbf{p}^l_{\text{spe}} = \sum_i \lambda^l_i \mathbf{t}^l_{S_i}$ are a learnable mixture of pre-trained source tokens, letting the frozen encoder route the target toward the most structurally similar sources. These structures are fused with feature-aligned embeddings through the scalar hyperparameters $\alpha$ and $\beta$ (Eqs. 7, 11, 12).

What would settle it

Run SAMGPT and the structure-token-free variant on source domains whose feature distributions are nearly identical but whose topologies differ sharply (for example, graphs built from the same raw attributes with very different edge densities). If the two variants reach the same target accuracy, or if the learned structure tokens converge to vectors close to all ones with near-zero variance across domains and layers, then the reported gains are not evidence that the tokens capture structural alignment.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that structural differences between text-free graph domains can be aligned by a very light mechanism: a per-domain, per-layer learnable vector that multiplies neighbor messages inside each aggregation step (Eq. 6). During pre-training the encoder weights are shared across all source domains while each domain carries its own token series, and the structure-aligned embeddings are added to feature-aligned embeddings with a scalar weight (Eq. 7). At adaptation, two prompt sets are injected into the frozen encoder: holistic prompts align the target domain with the unified multi-domain prior, and specific prompts are formed as a learnable linear mixture of the pre-trained structure tokens (Eq. 10), so the target can borrow more from structurally relevant source domains. The paper reports consistent accuracy gains in one-shot node and graph classification across seven target domains, and its ablations show that removing structure tokens, holistic prompts, or specific prompts each lowers performance.

Load-bearing premise

The load-bearing assumption is that multiplying neighbor messages by one small learnable vector per domain per layer is enough to reconcile real differences in average degree, shortest-path length, and clustering across domains while all encoder weights stay shared and frozen during adaptation.

Editorial extensions

If this is right

  • A text-free graph encoder can be pre-trained once on several source domains and then adapted to an unseen domain by tuning only prompts and token-mixture coefficients, not the encoder weights.
  • Adding more source domains helps SAMGPT consistently, whereas the compared feature-only multi-domain methods sometimes degrade as domains are added.
  • The advantage over feature-only alignment persists when source or target graphs are heterophilic (low homophily), not just in homophilic citation and e-commerce networks.
  • The gains are concentrated in low-shot settings ($m \le 5$) and shrink as labeled examples approach 10 per class, consistent with prompt-based adaptation being most valuable when supervision is scarce.
  • The per-layer token design makes the method applicable to any message-passing encoder and to both node-level and graph-level downstream tasks through the same subgraph-similarity task template.

Reading between the lines

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

  • One could read the learned mixture coefficients $\lambda^l_i$ as an interpretable measure of which source domain's topology the target most resembles at each layer; the paper does not analyze this, but it is directly extractable from the trained model.
  • A natural stress test would be to scale to dozens of domains or to domains with radically different connectivity (for example, temporal graphs or hypergraphs), where a single per-layer vector may be too coarse to capture within-domain structural heterogeneity.
  • If structure tokens are doing the claimed work, the same token-plus-dual-prompt design could be transplanted onto stronger feature-alignment backbones and onto graph transformers, extending the benefit beyond the GCN encoder used in the experiments.
  • The method's assumption that one token per layer suffices could be probed by comparing against tokens parameterized as functions of explicit statistics like average degree or clustering coefficient; that comparison is not in the paper.
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

4 major / 5 minor

Summary. The paper proposes SAMGPT, a text-free graph foundation model for multi-domain pre-training and cross-domain adaptation. The method inherits a feature alignment and feature adaptation backbone from MDGPT (Eqs. 3-5) and adds two novel ingredients: (i) per-domain, per-layer "structure tokens" that rescale neighbor messages during aggregation (Eq. 6), and (ii) dual prompts, consisting of holistic prompts and specific prompts that linearly combine pre-trained structure tokens, applied during frozen-encoder adaptation (Eqs. 9-10). The authors evaluate on seven public datasets for one-shot node and graph classification, report few-shot curves, and run ablations that include a variant explicitly stated to be equivalent to MDGPT.

Significance. If the central claim were established, the structure-alignment mechanism would be a useful contribution toward text-free multi-domain graph foundation models. The empirical study is broad: seven datasets, node and graph classification, 500 outcomes per setting, and an ablation with multiple component-removal variants. However, the genuine novelty is concentrated in a small set of equations, and the experiments as presented do not demonstrate that the proposed tokens actually align structural distributions. The paper also overstates its empirical superiority, since SAMGPT is not the best method on every target domain in Tables 2 and 3. The work is clearly positioned relative to MDGPT and GCOPE, but the absence of MDGPT from the main baseline tables makes the incremental contribution hard to quantify.

major comments (4)
  1. [Sec. 4.2, Eq. (6)] The claimed structure alignment is not supported by the form of the structure tokens. The token t^l_{S_i} is a per-domain, per-layer vector that multiplies every neighbor message identically; it does not depend on node degree, edge identity, shortest path length, clustering coefficient, or any of the structural statistics listed in Table 1. For a GCN, the operation is algebraically equivalent to a per-domain diagonal reparameterization of the layer weight W^l on the neighbor-aggregation term (replacing W^l with W^l diag(t^l_{S_i})), which is a global rescaling of the neighbor contribution relative to the self contribution. Consequently, Table 5 shows that adding these learnable vectors improves accuracy, but it does not show that the improvement comes from aligning structure rather than from additional per-domain capacity. The paper should either provide evidence that the learned tokens correlate with or compensate for the structural statistics in Table 1, modify the tokens to be explicit functions of structural features, or reframe the contribution as per-domain adaptive aggregation without the structure-alignment interpretation.
  2. [Sec. 5.2, Tables 2 and 3] The text claims SAMGPT achieves "outstanding performance ... across various target domains" and "consistently outperforms" baselines, but the reported accuracy is not best in every setting. On Citeseer node classification, SAMGPT (36.38) is below GCOPE (38.33); on Pubmed graph classification, SAMGPT (48.69) is below GraphPrompt (49.55); and on Citeseer graph classification, SAMGPT (38.75) is below GCOPE (39.93). Since the paper repeatedly uses unqualified "state-of-the-art" and "consistently" language, the exceptions need to be acknowledged and discussed, and the claims should be qualified to "most target domains" or supported by significance testing.
  3. [Sec. 5.3, Table 5] Variant 1 is stated to be equivalent to MDGPT, and MDGPT is the source of the feature alignment and feature adaptation components (Eqs. 3-5). Yet MDGPT is not included in the main comparison Tables 2 and 3, and Table 5 reports ablations on only three of the seven target domains. To assess the true contribution of SAMGPT over its backbone, MDGPT should be added to the main tables, and the model ablation should be reported on all seven target domains rather than a subset.
  4. [Sec. 5.2, Fig. 3] The "consistently outperforms" claim for low-shot settings is not backed by statistical significance testing. With 500 outcomes per setting, the reported standard deviations are comparable to or larger than the mean differences in several cases (e.g., node classification on Facebook: SAMGPT 42.70 ± 8.73 vs GCOPE 40.63 ± 8.50). The paper should report paired significance tests or confidence intervals before claiming consistent improvement.
minor comments (5)
  1. [Appendix D] DGI is cited as reference [43], but in the bibliography [43] is the GAT paper; the DGI paper is [44]. This citation error should be corrected.
  2. [Algorithm 1] The line "Phol, Λ, Γ ← initialization" appears inside the loop over graphs, which suggests the parameters are reinitialized for each graph. Since these are global parameters tuned on the entire labeled training set, the initialization should be moved outside the per-graph loop.
  3. [Table 1] The abbreviations "Avg. nd", "Avg. spl", and "Avg. cc" are used in the table header without being defined in the caption; they are only defined in the table footnote. The caption should define them.
  4. [Sec. 4.2, Eq. (6)] The equation applies the structure token only to neighbor messages and not to the self message h^{l-1}_v. This asymmetric design choice is not discussed; the paper should explain why the self contribution is left unscaled and how this asymmetry relates to structural alignment.
  5. [Sec. 5.4, Table 6] For the heterophilic target domains, it is not specified how the dual prompts are initialized or whether they receive any structural supervision beyond the downstream labels. Please clarify the initialization and whether the prompts are purely random vectors at the start of tuning.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation is exhibited; the structure-token and dual-prompt components are tested against external and internal baselines, and the topology-independence concern is a correctness/identifiability issue rather than a circularity.

full rationale

Walking the claimed derivation chain, SAMGPT's central assertion is that per-domain structure tokens (Eq. 6) and dual prompts (Eqs. 9-10) improve cross-domain few-shot adaptation. None of these components is defined in terms of the quantities they are claimed to affect: Eq. 6 injects learnable per-layer vectors multiplicatively into neighbor aggregation, and Eqs. 9-10 inject or linearly mix such vectors during adaptation. The paper does not fit the tokens to Table 1's structural statistics (average degree, shortest path length, clustering coefficient) and then report those statistics as a prediction, so the fitted-input-called-prediction pattern does not apply. The feature-alignment backbone (Eqs. 3-5), pre-training loss (Eq. 8), and downstream task template (Eq. 13) are inherited from the authors' own prior work (GraphPrompt, GraphPrompt+, MDGPT), which is self-citation, but it is not load-bearing in a circular sense: Table 5 explicitly reports an MDGPT-equivalent Variant 1 and compares it against variants that add the proposed structure components, so the incremental gain is measured rather than assumed. The dual-prompt fusion in Eqs. 11-12 is a weighted sum of separately encoded branches, not an algebraic identity that forces the reported accuracy. The skeptic's concern that Eq. 6 tokens are topology-independent per-domain diagonal scalings is a legitimate correctness/identifiability criticism of the 'structure alignment' interpretation, but it does not exhibit a self-referential equation, a renamed known result, or a fitted parameter renamed as a prediction; it is an alternative explanation for the observed gain, not circularity. Therefore no circular step is established, and the score reflects only the presence of minor, non-load-bearing self-citations in the scaffolding.

Assumptions & free parameters 5 free parameters · 4 assumptions · 3 invented entities

SAMGPT introduces no new physical or mathematical entities. The structure tokens and dual prompts are learnable parameter vectors whose only support is downstream accuracy on benchmark datasets. The most important inherited assumptions are the SVD dimension alignment and the MDGPT feature alignment procedure, both taken from prior work. The strongest paper-specific assumption is that per-domain multiplicative rescaling of neighbor messages is sufficient to align structural distributions while sharing encoder weights.

free parameters (5)
  • Domain structure tokens t^l_Si = learned during pre-training, dimension equals hidden size (256)
    Eq. 6: per-domain, per-layer element-wise multipliers applied to neighbor embeddings. These are the core novelty and are fit to the pre-training data.
  • Holistic prompts p^l_hol = learned during adaptation
    Eq. 9: per-layer learnable vectors injected into aggregation to align the target domain with unified source-domain knowledge.
  • Specific prompt mixture coefficients lambda^l_i = learned during adaptation
    Eq. 10: per-layer weights combining the K pre-trained structure tokens, allowing the target domain to weight source domains differently.
  • Fusion weight alpha = 1
    Eqs. 7 and 12: controls the balance between feature-aligned and structure-aligned embeddings. Set to 1 based on sensitivity analysis in Appendix G on target domains Cora, Photo, and Facebook.
  • Fusion weight beta = 1
    Eq. 11: controls the balance between holistic and specific prompts. Also set to 1 based on the same sensitivity analysis.
assumptions (4)
  • domain assumption SVD-based dimension alignment preserves enough information for cross-domain transfer
    Eq. 3 inherits SVD alignment from prior work [64, 68]. If the projection discards domain-specific signal, the feature alignment stage could fail independently of structure alignment.
  • domain assumption The feature alignment method of Yu et al. [64] effectively unifies semantic spaces across text-free domains
    Eq. 4 is taken from MDGPT, and the paper states it follows Yu et al. for FAL and FAD. SAMGPT's measured gains sit on top of this inherited component.
  • domain assumption GraphCL-style contrastive loss provides a universal pre-training objective transferable across domains
    Eq. 8 adopts GraphCL, relying on graph augmentations and a readout to create graph-level embeddings that are useful across node and graph classification.
  • ad hoc to paper Shared encoder weights can be reused across domains when neighbor messages are rescaled by per-domain tokens
    Eq. 6 assumes per-domain multiplicative scaling is sufficient to reconcile structural differences such as average degree, path length, and clustering coefficient. No theoretical or empirical proof is offered beyond downstream accuracy.
invented entities (3)
  • Structure tokens
    purpose: Per-domain, per-layer learnable multiplicative vectors that modify neighbor aggregation to align structural distributions across source domains.
    Their only observable effect is downstream task accuracy; the paper provides no independent falsifiable prediction tied to these vectors outside the benchmark results.
  • Holistic prompts
    purpose: Learnable per-layer vectors that align the target domain with unified multi-domain structural knowledge during adaptation.
    They are internal model parameters evaluated only through final classification accuracy.
  • Specific prompts
    purpose: Learnable per-layer mixtures of pre-trained source structure tokens, adapting fine-grained domain-specific structural knowledge to the target domain.
    They are internal model parameters with no external falsifiable handle outside the reported downstream performance.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SAMGPT: Text-free Graph Foundation Model for Multi-domain Pre-training and Cross-domain Adaptation." pith.science (2026). https://pith.science/paper/NGZBEBEM

@misc{pith2026250205424,
  author       = {Pith},
  title        = {Pith review of: SAMGPT: Text-free Graph Foundation Model for Multi-domain Pre-training and Cross-domain Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NGZBEBEM}},
  note         = {Machine review of arXiv:2502.05424}
}
read the original abstract

Graphs are able to model interconnected entities in many online services, supporting a wide range of applications on the Web. This raises an important question: How can we train a graph foundational model on multiple source domains and adapt to an unseen target domain? A major obstacle is that graphs from different domains often exhibit divergent characteristics. Some studies leverage large language models to align multiple domains based on textual descriptions associated with the graphs, limiting their applicability to text-attributed graphs. For text-free graphs, a few recent works attempt to align different feature distributions across domains, while generally neglecting structural differences. In this work, we propose a novel Structure Alignment framework for text-free Multi-domain Graph Pre-Training and cross-domain adaptation (SAMGPT). It is designed to learn multi-domain knowledge from graphs originating in multiple source domains, which can then be adapted to address applications in an unseen target domain. Specifically, we introduce a set of structure tokens to harmonize structure-based aggregation across source domains during the pre-training phase. Next, for cross-domain adaptation, we design dual prompts, namely, holistic prompts and specific prompts, which adapt unified multi-domain structural knowledge and fine-grained, domain-specific information, respectively, to a target domain. Finally, we conduct comprehensive experiments on seven public datasets to evaluate and analyze the effectiveness of SAMGPT.

Figures

Figures reproduced from arXiv: 2502.05424 by the authors.

Figure 1
Figure 1. Motivation of SAMGPT. Research problem. Thus, it is crucial to pre-train a graph model on a wide range of multi-domain (i.e., multi-dataset) graphs and achieve cross-domain adaptation. However, graph structures from different datasets often exhibit markedly distinct characteristics. For instance, the structural patterns in a social network might not be directly applicable to a citation or e-commerce graph. Such dive… view at source ↗
Figure 2
Figure 2. Overall framework of SAMGPT. 4.1 Overall Framework SAMGPT consists of two phases: multi-domain pre-training, and cross-domain adaptation, as shown in [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Impact of number of shots on node and graph clas [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Sensitivity study of 𝛼 and 𝛽. dimensions set to 256. Specifically, we select edge dropping as the augmentations, with a default augmentation ratio of 0.2. For GPPT [39], we utilize a 2-layer GraphSAGE as its backbone, setting the hidden dimensions to 256. We employ a m…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

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

    cs.CL 2025-06 conditional novelty 6.0 of 10

    MDGCL pre-trains graph encoders on multiple domains using same-domain discrimination and a downstream domain-attention mechanism, outperforming existing text-free graph foundation models.

Reference graph

Works this paper leans on

77 extracted references · 56 canonical work pages · cited by 1 Pith paper

  1. [1]

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

    arXiv 2023

  2. [2]

    Vibhor Agarwal, Sagar Joglekar, Anthony P Young, and Nishanth Sastry. 2022. GraphNLI: A Graph-based Natural Language Inference Model for Polarity Pre- diction in Online Debates. In WWW. 2729–2737

    work page 2022

  3. [3]

    Karsten M Borgwardt and Hans-Peter Kriegel. 2005. Shortest-path kernels on graphs. In ICDM. IEEE, 8–pp

    work page 2005

  4. [4]

    Kaize Ding, Kai Shu, Xuan Shan, Jundong Li, and Huan Liu. 2021. Cross-domain graph anomaly detection. IEEE TNNLS 33, 6 (2021), 2406–2415

    work page 2021

  5. [5]

    Junfeng Fang, Xinglin Li, Yongduo Sui, Yuan Gao, Guibin Zhang, Kun Wang, Xi- ang Wang, and Xiangnan He. 2024. EXGC: Bridging Efficiency and Explainability in Graph Condensation. In WWW. 721–732

    work page 2024

  6. [6]

    Junfeng Fang, Wei Liu, Yuan Gao, Zemin Liu, An Zhang, Xiang Wang, and Xiangnan He. 2023. Evaluating Post-hoc Explanations for Graph Neural Networks via Robustness Analysis. In Thirty-seventh Conference on Neural Information Processing Systems

    work page 2023

  7. [7]

    Taoran Fang, Yunchao Zhang, Yang Yang, Chunping Wang, and Lei Chen. 2023. Universal Prompt Tuning for Graph Neural Networks. In NeurIPS, Vol. 36

    work page 2023

  8. [8]

    Christos Giatsidis, Fragkiskos Malliaros, Dimitrios Thilikos, and Michalis Vazir- giannis. 2014. CoreCluster: A degeneracy based graph clustering framework. In AAAI, Vol. 28

    work page 2014

Show all 77 references
  1. [9]

    Kaveh Hassani. 2022. Cross-domain few-shot graph classification. In AAAI, Vol. 36. 6856–6864

    2022

  2. [10]

    Zhenyu Hou, Xiao Liu, Yukuo Cen, Yuxiao Dong, Hongxia Yang, Chunjie Wang, and Jie Tang. 2022. Graphmae: Self-supervised masked graph autoencoders. In KDD. 594–604

    2022

  3. [11]

    Ziniu Hu, Yuxiao Dong, Kuansan Wang, Kai-Wei Chang, and Yizhou Sun. 2020. Gpt-gnn: Generative pre-training of graph neural networks. SIGKDD (2020), 1857–1867

    2020

  4. [12]

    Xinke Jiang, Zidi Qin, Jiarong Xu, and Xiang Ao. 2023. Incomplete graph learning via attribute-structure decoupled variational auto-encoder. In WSDM. 304–312

    2023

  5. [13]

    Xinke Jiang, Rihong Qiu, Yongxin Xu, Wentao Zhang, Yichen Zhu, Ruizhe Zhang, Yuchen Fang, Xu Chu, Junfeng Zhao, and Yasha Wang. 2024. RAGraph: A General Retrieval-Augmented Graph Learning Framework. NeurIPS (2024)

    2024

  6. [14]

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

    2016 arXiv

  7. [15]

    Thomas N Kipf and Max Welling. 2017. Semi-supervised classification with graph convolutional networks. In ICLR

    2017

  8. [16]

    Jintang Li, Ruofan Wu, Wangbin Sun, Liang Chen, Sheng Tian, Liang Zhu, Changhua Meng, Zibin Zheng, and Weiqiang Wang. 2023. What’s Behind the Mask: Understanding Masked Graph Modeling for Graph Autoencoders. In KDD. 1268–1279

    2023

  9. [17]

    Rongfan Li, Xinke Jiang, Ting Zhong, Goce Trajcevski, Jin Wu, and Fan Zhou

  10. [18]

    Hao Liu, Jiarui Feng, Lecheng Kong, Ningyue Liang, Dacheng Tao, Yixin Chen, and Muhan Zhang. 2024. One for All: Towards Training One Graph Model for All Classification Tasks. In ICLR

    2024

  11. [19]

    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

  12. [20]

    Zemin Liu, Yuan Fang, Chenghao Liu, and Steven CH Hoi. 2021. Relative and absolute location embedding for few-shot node classification on graph. In AAAI, Vol. 35. 4267–4275

    2021

  13. [21]

    Zemin Liu, Xingtong Yu, Yuan Fang, and Xinming Zhang. 2023. GraphPrompt: Unifying pre-training and downstream tasks for graph neural networks. InWWW. 417–428

    2023

  14. [22]

    Yuanfu Lu, Xunqiang Jiang, Yuan Fang, and Chuan Shi. 2021. Learning to pre-train graph neural networks. In AAAI, Vol. 35. 4276–4284

    2021

  15. [23]

    Chenglong Ma, Yongli Ren, Pablo Castells, and Mark Sanderson. 2024. Temporal Conformity-aware Hawkes Graph Network for Recommendations. In WWW. 3185–3194

    2024

  16. [24]

    Yao Ma, Xiaorui Liu, Neil Shah, and Jiliang Tang. 2022. Is homophily a necessity for graph neural networks?. In ICLR

    2022

  17. [25]

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

  18. [26]

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

  19. [27]

    Kentaro Miyake, Hiroyoshi Ito, Christos Faloutsos, Hirotomo Matsumoto, and Atsuyuki Morishima. 2024. NETEVOLVE: Social Network Forecasting using Multi-Agent Reinforcement Learning with Interpretable Features. In WWW. 2542–2551

    2024

  20. [28]

    Xuelian Ni, Fei Xiong, Yu Zheng, and Liang Wang. 2024. Graph Contrastive Learning with Kernel Dependence Maximization for Social Recommendation. In WWW. 481–492

    2024

  21. [29]

    Hongbin Pei, Bingzhe Wei, Kevin Chen-Chuan Chang, Yu Lei, and Bo Yang. 2020. Geom-GCN: Geometric graph convolutional networks. In ICLR

    2020

  22. [30]

    Jiezhong Qiu, Qibin Chen, Yuxiao Dong, Jing Zhang, Hongxia Yang, Ming Ding, Kuansan Wang, and Jie Tang. 2020. Gcc: Graph contrastive coding for graph neural network pre-training. In SIGKDD. 1150–1160

    2020

  23. [31]

    Faisal Rahutomo, Teruaki Kitasuka, Masayoshi Aritsugi, et al. 2012. Semantic cosine similarity. In ICAST, Vol. 4. 1

    2012

  24. [32]

    Ladislav Rampášek, Michael Galkin, Vijay Prakash Dwivedi, Anh Tuan Luu, Guy Wolf, and Dominique Beaini. 2022. Recipe for a general, powerful, scalable graph transformer. NeurIPS 35 (2022), 14501–14515

    2022

  25. [33]

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

    2021

  26. [34]

    Benedek Rozemberczki and Rik Sarkar. 2020. Characteristic functions on graphs: Birds of a feather, from statistical descriptors to parametric models. In CIKM. 1325–1334

    2020

  27. [35]

    Prithviraj Sen, Galileo Namata, Mustafa Bilgic, Lise Getoor, Brian Galligher, and Tina Eliassi-Rad. 2008. Collective classification in network data. AI magazine 29, 3 (2008), 93–93

    2008

  28. [36]

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

    2018 arXiv

  29. [37]

    Gilbert W Stewart. 1993. On the early history of the singular value decomposition. SIAM review 35, 4 (1993), 551–566

    1993

  30. [38]

    Fan-Yun Sun, Jordan Hoffman, Vikas Verma, and Jian Tang. 2019. InfoGraph: Unsupervised and Semi-supervised Graph-Level Representation Learning via Mutual Information Maximization. In ICLR

    2019

  31. [39]

    Mingchen Sun, Kaixiong Zhou, Xin He, Ying Wang, and Xin Wang. 2022. Gppt: Graph pre-training and prompt tuning to generalize graph neural networks. In SIGKDD. 1717–1727

    2022

  32. [40]

    Xiangguo Sun, Hong Cheng, Jia Li, Bo Liu, and Jihong Guan. 2023. All in One: Multi-Task Prompting for Graph Neural Networks. SIGKDD (2023), 2120–2131

    2023

  33. [41]

    Jiabin Tang, Yuhao Yang, Wei Wei, Lei Shi, Long Xia, Dawei Yin, and Chao Huang

  34. [42]

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

    2023 arXiv

  35. [43]

    Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. 2018. Graph attention networks. In ICLR

    2018

  36. [44]

    Petar Veličković, William Fedus, William L Hamilton, Pietro Liò, Yoshua Bengio, and R Devon Hjelm. 2018. Deep Graph Infomax. In ICLR

    2018

  37. [45]

    Chen Wang, Yueqing Liang, Zhiwei Liu, Tao Zhang, and S Yu Philip. 2021. Pre- training graph neural network for cross domain recommendation. In CogMI. 140–145

    2021

  38. [46]

    Ning Wang, Minnan Luo, Kaize Ding, Lingling Zhang, Jundong Li, and Qinghua Zheng. 2020. Graph Few-shot Learning with Attribute Matching. In CIKM. 1545– 1554

    2020

  39. [47]

    Qizhou Wang, Guansong Pang, Mahsa Salehi, Wray Buntine, and Christopher Leckie. 2023. Cross-domain graph anomaly detection via anomaly-aware con- trastive alignment. In AAAI, Vol. 37. 4676–4684

    2023

  40. [48]

    Zeyuan Wang, Qiang Zhang, HU Shuang-Wei, Haoran Yu, Xurui Jin, Zhichen Gong, and Huajun Chen. 2022. Multi-level Protein Structure Pre-training via Prompt Learning. In ICLR

    2022

  41. [49]

    Zhihao Wen and Yuan Fang. 2023. Augmenting Low-Resource Text Classification with Graph-Grounded Pre-training and Prompting. In SIGIR. 506–516

    2023

  42. [50]

    Zhihao Wen and Yuan Fang. 2024. Prompt tuning on graph-augmented low- resource text classification. IEEE TKDE (2024)

    2024

  43. [51]

    Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and S Yu Philip. 2020. A comprehensive survey on graph neural networks. IEEE TNNLS 32, 1 (2020), 4–24

    2020

  44. [52]

    Jun Xia, Lirong Wu, Jintao Chen, Bozhen Hu, and Stan Z Li. 2022. Simgrace: A simple framework for graph contrastive learning without data augmentation. In WWW. 1070–1079

    2022

  45. [53]

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

    2024 arXiv

  46. [54]

    Minghao Xu, Hang Wang, Bingbing Ni, Hongyu Guo, and Jian Tang. 2021. Self- supervised graph-level representation learning with local and global structure. In ICML. 11548–11558

    2021

  47. [55]

    Baoyao Yang and Pong C Yuen. 2019. Cross-domain visual representations via unsupervised graph alignment. In AAAI, Vol. 33. 5613–5620

    2019

  48. [56]

    Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, and Tie-Yan Liu. 2021. Do transformers really perform badly for graph representation? NeurIPS 34 (2021), 28877–28888

    2021

  49. [57]

    Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. 2020. Graph contrastive learning with augmentations. NeurIPS 33 WWW ’25, April 28–May 2, 2025, Sydney, NSW, Australia. Xingtong Yu, Zechuan Gong, Chang Zhou, Yuan Fang, and Hui Zhang (2020), 5812–5823

    2020

  50. [58]

    Xingtong Yu, Zhenghao Liu, Yuan Fang, Zemin Liu, Sihong Chen, and Xinming Zhang. 2024. Generalized Graph Prompt: Toward a Unification of Pre-Training and Downstream Tasks on Graphs. IEEE TKDE (2024)

    2024

  51. [59]

    Xingtong Yu, Zemin Liu, Yuan Fang, and Xinming Zhang. 2023. Learning to count isomorphisms with graph neural networks. AAAI 37, 4 (2023), 4845–4853

    2023

  52. [60]

    Xingtong Yu, Zemin Liu, Yuan Fang, and Xinming Zhang. 2024. HGPROMPT: Bridging Homogeneous and Heterogeneous Graphs for Few-shot Prompt Learn- ing. In AAAI. 16578–16586

    2024

  53. [61]

    Xingtong Yu, Zhenghao Liu, Yuan Fang, and Xinming Zhang. 2025. Node-Time Conditional Prompt Learning in Dynamic Graphs. In ICLR

    2025

  54. [62]

    Xingtong Yu, Jie Zhang, Yuan Fang, and Renhe Jiang. 2025. Non-Homophilic Graph Pre-Training and Prompt Learning. In SIGKDD

    2025

  55. [63]

    Xingtong Yu, Chang Zhou, Yuan Fang, and Xinming Zhang. 2024. MultiGPrompt for Multi-Task Pre-Training and Prompting on Graphs. InWWW. 515–526

    2024

  56. [64]

    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

  57. [65]

    Seongjun Yun, Minbyul Jeong, Raehyun Kim, Jaewoo Kang, and Hyunwoo J Kim

  58. [66]

    Delvin Ce Zhang, Menglin Yang, Rex Ying, and Hady W Lauw. 2024. Text- attributed graph representation learning: Methods, applications, and challenges. In WWW. 1298–1301

    2024

  59. [67]

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. 2023. Adding conditional control to text-to-image diffusion models. In ICCV. 3836–3847

    2023

  60. [68]

    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. In SIGKDD. 4443–4454

    2024

  61. [69]

    Jianan Zhao, Meng Qu, Chaozhuo Li, Hao Yan, Qian Liu, Rui Li, Xing Xie, and Jian Tang. 2023. Learning on Large-scale Text-attributed Graphs via Variational Inference. In ICLR

    2023

  62. [70]

    Jiong Zhu, Yujun Yan, Lingxiao Zhao, Mark Heimann, Leman Akoglu, and Danai Koutra. 2020. Beyond homophily in graph neural networks: Current limitations and effective designs. NeurIPS (2020), 7793–7804

    2020

  63. [71]

    Yun Zhu, Yaoke Wang, Haizhou Shi, Zhenshuo Zhang, Dian Jiao, and Siliang Tang

  64. [77]

    GraphControl: Adding Conditional Control to Universal Graph Pre-trained Models for Graph Domain Transfer Learning. In WWW. 539–550. Appendices A Algorithm Our algorithm consists of two stages, multi-domain graph pre- training and downstream adaptation. In the multi-domain pre-...

    2025

  65. [2000]

    Information Retrieval 3 (2000), 127–163

    Automating the construction of internet portals with machine learning. Information Retrieval 3 (2000), 127–163

    2000

  66. [2015]

    In SIGIR

    Image-based recommendations on styles and substitutes. In SIGIR. 43–52

  67. [2019]

    NeurIPS 32 (2019)

    Graph transformer networks. NeurIPS 32 (2019)

    2019

  68. [2022]

    In SIGKDD

    Mining spatio-temporal relations via self-paced graph contrastive learning. In SIGKDD. 936–944

  69. [2024]

    In SIGKDD

    HiGPT: Heterogeneous Graph Language Model. In SIGKDD. 2842–2853

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.