Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

The paper claims that pre-training on a mixture of homogeneous and heterogeneous graphs, with one expert per graph and class-wise fusion, delivers stable few-shot transfer and outperforms type-specific pre-training methods.

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

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

T0 review · deepseek-v4-flash

2026-08-02 23:34 UTC pith:DZAEASUU

load-bearing objection GPH2 is an honest empirical study of per-domain graph encoders with attention fusion, but the 'unified pre-training' claim is oversold: the gains likely come from ensembling, and no ensemble baseline is run. the 4 major comments →

arxiv 2602.13075 v2 pith:DZAEASUU submitted 2026-02-13 cs.LG

Unified Multi-Domain Graph Pre-training for Homogeneous and Heterogeneous Graphs via Domain-Specific Expert Encoding

Chundong Liang , Yongqi Huang , Dongxiao He , Peiyuan Li , Yawen Li , Di Jin , Weixiong Zhang This is my paper
classification cs.LG
keywords graph pre-traininghomogeneous graphsheterogeneous graphsmulti-domain transfermeta-pathexpert encodingfew-shot node classificationself-supervised learning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper argues that the long-standing split between homogeneous and heterogeneous graph pre-training is unnecessary: a balanced mixture of both graph types during pre-training consistently improves few-shot node classification downstream. It proposes GPH2, which turns any graph into a set of edge-only views — edge-dropped topologies for homogeneous graphs, meta-path-induced topologies for heterogeneous graphs — so a single ordinary GNN can encode either type. To avoid cross-domain conflict, one expert encoder is pre-trained per source graph, then frozen and fused at adaptation time by a class-wise attention mechanism after per-expert linear alignment. On 3- and 5-shot node classification, GPH2 beats the strongest homogeneous baseline by an average of 4.59 F1 points and the strongest heterogeneous baseline by 5.98. If correct, this is a step toward graph foundation models that treat mixed real-world graph collections as one pre-training corpus.

Core claim

GPH2 claims that homogeneous and heterogeneous graphs can be pre-trained jointly in one framework without any graph-type-specific encoder. The trick is Unified Multi-View Graph Construction: each source graph is rewritten as several views that differ only in edge sets — random edge dropping for homogeneous graphs, meta-path-induced rewiring for heterogeneous graphs — so the encoder never needs to know the original type. Because a single shared encoder across all domains would fight distribution shift, each graph gets its own expert, pre-trained independently with a DGI-style Infomax objective. Downstream, every frozen expert encodes the target graph; an expert-specific linear map aligns the

What carries the argument

Unified Multi-View Graph Construction: maps a homogeneous graph to P edge-dropped subgraphs and a heterogeneous graph to P meta-path-induced subgraphs (a meta-path is a chain of node and edge types such as paper-author-paper that defines a homogeneous adjacency). All views share the same nodes and features, so a vanilla GNN — here a 2-layer SGC — can encode any graph. Domain-Specific Expert Encoding: one GNN plus view-attention vector per source graph, pre-trained independently with the Infomax objective to avoid cross-domain negative transfer. Task-oriented Expert Fusion: per-expert linear alignment, then C learnable class vectors produce softmax attention over experts for each class, fusin

Load-bearing premise

The whole method rests on the assumption that converting each heterogeneous graph into a few hand-picked meta-path views — and each homogeneous graph into random edge-dropped views — preserves the semantics any downstream task will need, even though node and edge type information is thrown away in the process.

What would settle it

Construct a heterogeneous benchmark where the label is determined by a relation that is not among the chosen meta-paths (for example, in a DBLP-like network, predict an author's field using only co-authorship through a specific venue relation absent from the APA/APCPA/APTPA meta-path set). Pre-train GPH2 exactly as specified, fine-tune with 3 labeled nodes per class, and compare against a heterogeneous encoder that consumes the full typed graph. If GPH2 matches or exceeds it, the unified view construction captures the needed semantics; if the typed encoder wins by a large margin, the meta-path

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

If this is right

  • Homogeneous-only and heterogeneous-only pre-training pipelines can be replaced by one method that accepts both graph types, so real-world mixed graph collections become usable pre-training data.
  • Few-shot node classification on either graph type improves when the pre-training corpus mixes types — by about 4.6 F1 points over the best homogeneous baseline and 6.0 over the best heterogeneous baseline in the paper's experiments.
  • Pre-training an expert per source graph sidesteps the conflicting optimization signals that arise when one shared encoder must absorb graphs with very different topology and feature distributions.
  • Because expert fusion is class-wise, a downstream model can weight the same expert differently for different classes, letting domain-relevant knowledge engage precisely where it helps.
  • The design is objective-agnostic: any self-supervised loss and any GNN can be plugged into the expert encoders, so future improvements in self-supervised graph learning can transfer directly into GPH2.

Where Pith is reading between the lines

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

  • The view-construction step discards heterogeneous type information; if an omitted relation carries the label signal, the unified encoder cannot recover it, so automatic meta-path selection or type-aware views is a natural next step.
  • The paper itself reports that adding heterogeneous graphs did not improve PubMed (Section 4.3), and its motivational study shows homogeneous-only pre-training can sometimes beat mixed pre-training on citation tasks — both suggest the benefit depends on domain similarity rather than graph-type diversity per se.
  • The 'one expert per graph' design means cost and memory grow linearly with the number of pre-training graphs; grouping similar graphs into shared experts or adding a learned routing layer could scale it to large graph corpora.
  • The learned class-wise attention weights could double as an interpretable diagnostic of which pre-training domains matter for a downstream task, effectively a domain-relevance probe; the paper's attention analyses hint at this but do not develop it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes GPH2, a two-stage framework for graph pre-training that aims to handle mixed homogeneous and heterogeneous graphs. In the pre-training stage, each source graph is transformed into a multi-view representation (edge-dropped views for homogeneous graphs, meta-path-induced views for heterogeneous graphs) and a separate GNN encoder is pre-trained per source graph with an Infomax objective. In the downstream stage, frozen expert encoders produce representations of the target graph, which are linearly aligned and fused via class-wise attention vectors to produce predictions. The authors report few-shot node classification results on five homogeneous and four heterogeneous datasets, claiming average F1 gains of 4.59 points over the strongest homogeneous baseline and 5.98 points over the strongest heterogeneous baseline, and argue that combining homogeneous and heterogeneous graphs in pre-training is beneficial.

Significance. If fully supported, a method that genuinely unifies pre-training across homogeneous and heterogeneous graphs would be a useful contribution to the graph foundation model agenda. The paper identifies a real gap—existing methods are typed—and the proposed multi-view construction is a simple and plausible way to give both graph types a common input format. The downstream class-wise fusion is also reasonable. However, as written, the central contribution is significantly overstated: the method is an ensemble of independently pre-trained encoders, the main experiments are performed under matched-type pre-training rather than the claimed mixed-type setting, and the baseline comparisons are hampered by unclear or potentially unfair adaptations. The paper is a solid empirical systems study of an ensemble-style approach, but it does not yet support the claim of a unified multi-domain pre-training method that outperforms existing methods in mixed-type scenarios.

major comments (4)
  1. [§4.2, Tables 1 and 2] The main comparisons are explicitly run in 'matched graph-type settings' (homogeneous-to-homogeneous and heterogeneous-to-heterogeneous). This does not test the paper's central claim, stated in the Abstract and §1, of unified pre-training across mixed homogeneous and heterogeneous graphs. The only mixed-type evidence is the motivational study (Fig. 2) and Table 3, which lacks baseline comparisons. The reported gains of 4.59 and 5.98 points therefore do not support the headline claim. Please add experiments where the pre-training set mixes both graph types, and compare GPH2 against all baselines under that same mixed setting.
  2. [§3.3–§3.4, Eqs. (2), (8), (11)–(16)] GPH2 pre-trains a separate encoder per source graph with non-shared parameters and then fuses their frozen outputs at downstream. This is by construction an ensemble of independently pre-trained models. The superiority over single-encoder baselines could be driven by the ensemble effect and by the extra downstream capacity (per-expert projections W_i, class-wise attention q_c) rather than by the proposed unified multi-view construction or task-oriented fusion. The ablation 'w/o Expert' (Fig. 4) is not a control for this effect because it removes all experts at once. Please add controls that isolate the contribution: (a) a single shared encoder pre-trained on all source graphs with the same multi-view construction; (b) an ensemble of single-view encoders with simple logit averaging; and (c) an oracle that selects the best single expert per downstream task.
  3. [§A.3, Baselines] The paper states that DMGI and HeCo were modified to share parameters across views to adapt them to multi-domain pre-training. This is a non-trivial modification: DMGI uses relation-specific encoders and HeCo uses view-specific encoders, and sharing parameters removes a core design element. If this weakens these baselines, the reported gains over them are inflated. Please validate the modified baselines by showing they retain their original performance on the original single-domain setting, or use unmodified versions with a clearly documented multi-domain adaptation procedure.
  4. [§4.2, §A.3] For baselines such as DGI, GRACE, and GraphMAE—which are designed for single-graph pre-training—the protocol for pre-training on 'all remaining graphs' is unspecified. Is a single encoder trained on the union of graphs (and if so, how are differing graph structures handled), or are separate encoders pre-trained and then combined or selected? The ambiguity makes the comparison difficult to reproduce and interpret. Please specify the exact pre-training protocol used for every baseline.
minor comments (5)
  1. [Eq. (3)] Notation inconsistency: Eq. (3) uses index k, while the surrounding text and Eqs. (4)–(7) use p. Please unify.
  2. [§4.1 and Table 1] The text mentions GraphPrompt as a baseline, but it does not appear in Table 1. Also, the text refers to 'SCOPE' while Table 1 and the reference list use 'GCOPE'; please use consistent naming.
  3. [Table 6] The learning rate entries are written as '0.015×10^-4', '0.0055×10^-4', etc., which is ambiguous. Please state the actual numeric values (e.g., 1.5e-6) or use a consistent scientific notation.
  4. [§4.3, Table 3] It is not clear whether Table 3 reports results using GPH2 or a simpler pre-training model. If it uses GPH2, say so explicitly; if not, clarify which model was used. This is important because the table is used to support the mixed-type claim.
  5. [Abstract and §2.2] The abstract mentions a 'balanced mixture' of homogeneous and heterogeneous graph pre-training, but the motivational study does not define or test any notion of balance. Please either define it or rephrase.

Circularity Check

0 steps flagged

No significant circularity: the paper is an empirical systems contribution; the minor self-citation is not load-bearing.

full rationale

GPH2 does not derive its headline numbers from its own assumptions. The components (edge-dropping views for homogeneous graphs, meta-path views for heterogeneous graphs, per-domain Infomax pre-training, and downstream class-wise attention fusion) are defined independently of the reported F1 values, and the main comparisons are against external baselines with their own published settings. The only self-citation, reference [15] (overlapping authors), appears in Section 3.3 to support the generic claim that cross-domain discrepancies exist, alongside independent reference [53]; it is not a premise from which the framework's equations or results are forced. The motivational study in Section 2.2 is an empirical observation, not an input to the method's construction. The ablations remove components and measure the resulting drop, which is a standard control rather than a circular prediction. The paper itself notes that comparisons were conducted under matched graph-type settings 'due to the limitations of baselines,' which limits the strength of the cross-type transfer claim but is an external-validity concern, not circularity. A potential experimental gap is the absence of a simple per-domain ensemble baseline, but this is a correctness/evaluation issue, not a definitional equivalence: no equation reduces the claimed gain to a fitted parameter renamed as a prediction. Therefore no circular step can be exhibited with quote-level specificity.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

Most of the paper's machinery is standard: edge-dropping augmentation, meta-path views, DGI objective, attention fusion. The novel step is the workflow that combines these and the empirical claim that per-domain experts plus class-wise attention transfer across graph types. No new entities or constants are introduced; the free parameters are hyperparameters and view-construction settings.

free parameters (3)
  • number of views P for homogeneous graphs = 3
    Chosen by hand in Section A.3; sensitivity analysis in Appendix B.1 shows stable performance but no principled selection.
  • edge-drop ratio for homogeneous view construction = 0.3
    Chosen by hand (Section A.3); Figure 6(b) shows robustness across 0.2-0.4.
  • per-dataset downstream hyperparameters (SGC depth, learning rate, dropout) = see Table 6
    Grid-searched per dataset, standard practice but means reported numbers are post-selection.
axioms (4)
  • ad hoc to paper Homogeneous and heterogeneous graphs can both be represented as a set of single-edge-set views acceptable to one GNN encoder
    Section 3.2 Eq. (3)-(5); the paper introduces this abstraction without proving it preserves relational semantics that type-aware encoders normally use.
  • domain assumption Meta-path-based connectivity captures the meaningful semantics of a heterogeneous graph
    Borrowed from metapath2vec/HAN line of work; assumed per dataset, not validated against alternative relational views.
  • domain assumption Infomax/DGI self-supervised pre-training transfers across domains
    Section 3.3; the objective is taken from DGI and assumed to produce transferable representations.
  • domain assumption 3-shot node classification accuracy is a proxy for pre-training quality
    All evaluations use the ProG benchmark; no other transfer tasks (link prediction, graph classification) are considered.

pith-pipeline@v1.3.0-alltime-deepseek · 19510 in / 9855 out tokens · 83545 ms · 2026-08-02T23:34:14.004601+00:00 · methodology

0 comments
read the original abstract

Graph pre-training has achieved remarkable success in recent years, delivering transferable representations for downstream adaptation. However, most existing methods are designed for either homogeneous or heterogeneous graphs, thereby hindering unified graph modeling across diverse graph types. This separation contradicts real-world applications, where mixed homogeneous and heterogeneous graphs are ubiquitous, and distribution shifts between upstream pre-training and downstream deployment are common. In this paper, we empirically demonstrate that a balanced mixture of homogeneous and heterogeneous graph pre-training benefits downstream tasks and propose a unified multi-domain \textbf{G}raph \textbf{P}re-training method across \textbf{H}omogeneous and \textbf{H}eterogeneous graphs ($\mathbf{GPH^{2}}$). To address the lack of a unified encoder for homogeneous and heterogeneous graphs, we propose a Unified Multi-View Graph Construction that simultaneously encodes both without explicit graph-type-specific designs. To cope with the increased cross-domain distribution discrepancies arising from mixed graphs, we introduce domain-specific expert encoding. Each expert is independently pre-trained on a single graph to capture domain-specific knowledge, thereby shielding the pre-training encoder from the adverse effects of cross-domain discrepancies. For downstream tasks, we further design a Task-oriented Expert Fusion Strategy that adaptively integrates multiple experts based on their discriminative strengths. Extensive experiments on mixed graphs demonstrate that $\text{GPH}^{2}$ enables stable transfer across graph types and domains, significantly outperforming existing graph pre-training methods.

Figures

Figures reproduced from arXiv: 2602.13075 by Chundong Liang, Di Jin, Dongxiao He, Peiyuan Li, Weixiong Zhang, Yawen Li, Yongqi Huang.

Figure 1
Figure 1. Figure 1: The fragmented progress of graph pre-training [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Performance of 3-shot node classification under [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The workflow of GPH2 . GPH2 is designed for learning across mix graphs (left). It consists of two stages: multi-domain pre-training (top) and downstream task adaptation (bottom). In the pre-training stage, each source graph is independently processed by a domain-specific expert, and in the downstream stage, all pre-trained experts are transferred for adaptation. encodings. Subsequently, multiple GNN encode… view at source ↗
Figure 4
Figure 4. Figure 4: Ablation study of on ACM and Cora. From the results, we observed that increasing the number of pre-training graphs generally improved downstream performance on Cora and PubMed, indicating that pre-training domain diver￾sity remains an important factor. Notably, when replacing Photo and Computer with other academic networks, i.e., ACM and DBLP, we observed an improvement in downstream performance on the Cor… view at source ↗
Figure 5
Figure 5. Figure 5: Attention scores of the Task-Oriented Expert Fu [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Hyper-parameter analysis of the number of views [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Attention scores of the Task-Oriented Expert Fu [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 2 Pith papers

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

  1. CHoE: Cross-Domain Heterogeneous Graph Prompt Learning via Structure-Conditioned Experts

    cs.LG 2026-05 unverdicted novelty 5.0

    CHoE introduces structure-conditioned experts with routing and semantic fusion to improve few-shot cross-domain heterogeneous graph prompt learning.

  2. CHoE: Cross-Domain Heterogeneous Graph Prompt Learning via Structure-Conditioned Experts

    cs.LG 2026-05 unverdicted novelty 5.0

    CHoE uses structure-conditioned experts, structure-aware routing with load balancing, and prompt-based semantic fusion to improve few-shot performance on cross-domain heterogeneous graph prompt learning tasks.

Reference graph

Works this paper leans on

70 extracted references · 13 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Hervé Abdi. 2007. Singular value decomposition (SVD) and generalized singular value decomposition.Encyclopedia of measurement and statistics907, 912 (2007), 44

    2007

  2. [2]

    Jiaoyan Chen, Yuxia Geng, Zhuo Chen, Jeff Z Pan, Yuan He, Wen Zhang, Ian Hor- rocks, and Huajun Chen. 2023. Zero-shot and few-shot learning with knowledge graphs: A comprehensive survey.Proc. IEEE111, 6 (2023), 653–685

    2023

  3. [3]

    Aaron Clauset, Cosma Rohilla Shalizi, and Mark EJ Newman. 2009. Power-law distributions in empirical data.SIAM review51, 4 (2009), 661–703

    2009

  4. [4]

    Yuxiao Dong, Nitesh V Chawla, and Ananthram Swami. 2017. metapath2vec: Scalable representation learning for heterogeneous networks. InProceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining. 135–144

    2017

  5. [5]

    Taoran Fang, Yunchao Zhang, Yang Yang, Chunping Wang, and Lei Chen. 2023. Universal prompt tuning for graph neural networks.Advances in Neural Infor- mation Processing Systems36 (2023), 52464–52489

    2023

  6. [6]

    Yin Fang, Qiang Zhang, Ningyu Zhang, Zhuo Chen, Xiang Zhuang, Xin Shao, Xiaohui Fan, and Huajun Chen. 2023. Knowledge graph-enhanced molecular contrastive learning with functional prompt.Nature Machine Intelligence5, 5 (2023), 542–553

    2023

  7. [7]

    Matthias Fey, Jinu Sunil, Akihiro Nitta, Rishi Puri, Manan Shah, Blaž Stojanovič, Ramona Bendias, Alexandria Barghi, Vid Kocijan, Zecheng Zhang, et al . 2025. PyG 2.0: Scalable learning on real world graphs.arXiv preprint arXiv:2507.16991 (2025)

    Pith/arXiv arXiv 2025

  8. [8]

    Xinyu Fu, Jiani Zhang, Ziqiao Meng, and Irwin King. 2020. Magnn: Metap- ath aggregated graph neural network for heterogeneous graph embedding. In Proceedings of the web conference 2020. 2331–2341

    2020

  9. [9]

    Victor Garcia and Joan Bruna. 2017. Few-shot learning with graph neural net- works.arXiv preprint arXiv:1711.04043(2017)

    Pith/arXiv arXiv 2017

  10. [10]

    Kaveh Hassani and Amir Hosein Khasahmadi. 2020. Contrastive multi-view representation learning on graphs. InInternational conference on machine learning. PMLR, 4116–4126

    2020

  11. [11]

    Zhenyu Hou, Yufei He, Yukuo Cen, Xiao Liu, Yuxiao Dong, Evgeny Kharlamov, and Jie Tang. 2023. Graphmae2: A decoding-enhanced masked self-supervised graph learner. InProceedings of the ACM web conference 2023. 737–746

    2023

  12. [12]

    Zhenyu Hou, Xiao Liu, Yukuo Cen, Yuxiao Dong, Hongxia Yang, Chunjie Wang, and Jie Tang. 2022. Graphmae: Self-supervised masked graph autoencoders. In Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining. 594–604

    2022

  13. [13]

    Weihua Hu, Bowen Liu, Joseph Gomes, Marinka Zitnik, Percy Liang, Vijay Pande, and Jure Leskovec. 2019. Strategies for pre-training graph neural networks.arXiv preprint arXiv:1905.12265(2019)

    Pith/arXiv arXiv 2019

  14. [14]

    Ziniu Hu, Yuxiao Dong, Kuansan Wang, Kai-Wei Chang, and Yizhou Sun. 2020. Gpt-gnn: Generative pre-training of graph neural networks. InProceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining. 1857–1867

    2020

  15. [15]

    Yongqi Huang, Jitao Zhao, Dongxiao He, Xiaobao Wang, Yawen Li, Yuxiao Huang, Di Jin, and Zhiyong Feng. 2025. One prompt fits all: Universal graph adaptation for pretrained models.arXiv preprint arXiv:2509.22416(2025)

    arXiv 2025

  16. [16]

    Xunqiang Jiang, Tianrui Jia, Yuan Fang, Chuan Shi, Zhe Lin, and Hui Wang. 2021. Pre-training on large-scale heterogeneous graph. InProceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining. 756–766

    2021

  17. [17]

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

    Pith/arXiv arXiv 2016

  18. [18]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. 2017. Semi-Supervised Classification with Graph Convolutional Networks. InProceedings of the International Conference on Learning Representations (ICLR)

    2017

  19. [19]

    Xiao Li, Li Sun, Mengjie Ling, and Yan Peng. 2023. A survey of graph neural network based recommendation in social networks.Neurocomputing549 (2023), 126441

    2023

  20. [20]

    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)

    Pith/arXiv arXiv 2023

  21. [21]

    Jiawei Liu, Cheng Yang, Zhiyuan Lu, Junze Chen, Yibo Li, Mengmei Zhang, Ting Bai, Yuan Fang, Lichao Sun, Philip S Yu, et al. 2025. Graph foundation models: Concepts, opportunities and challenges.IEEE Transactions on Pattern Analysis and Machine Intelligence(2025)

    2025

  22. [22]

    Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. 2023. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing.ACM computing surveys55, 9 (2023), 1–35

    2023

  23. [23]

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

  24. [24]

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

    2023

  25. [25]

    Yuanfu Lu, Xunqiang Jiang, Yuan Fang, and Chuan Shi. 2021. Learning to pre- train graph neural networks. InProceedings of the AAAI conference on artificial intelligence, Vol. 35. 4276–4284

    2021

  26. [26]

    Yihong Ma, Ning Yan, Jiayu Li, Masood Mortazavi, and Nitesh V Chawla. 2024. Hetgpt: Harnessing the power of prompt tuning in pre-trained heterogeneous graph neural networks. InProceedings of the ACM Web Conference 2024. 1015– 1023

    2024

  27. [27]

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

  28. [28]

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

  29. [29]

    Chanyoung Park, Jiawei Han, and Hwanjo Yu. 2020. Deep multiplex graph infomax: Attentive multiplex network embedding using global information. Knowledge-Based Systems197 (2020), 105861

    2020

  30. [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. InProceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining. 1150–1160. Conference’17, July 2017, Washington, DC, USA Chundong Liang, Yongq...

    2020

  31. [31]

    Prithviraj Sen, Galileo Namata, Mustafa Bilgic, Lise Getoor, Brian Galligher, and Tina Eliassi-Rad. 2008. Collective classification in network data.AI magazine29, 3 (2008), 93–93

    2008

  32. [32]

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

    Pith/arXiv arXiv 2018

  33. [33]

    Chuan Shi, Yitong Li, Jiawei Zhang, Yizhou Sun, and Philip S Yu. 2016. A survey of heterogeneous information network analysis.IEEE Transactions on Knowledge and Data Engineering29, 1 (2016), 17–37

    2016

  34. [34]

    Guiquan Sun, Xikun Zhang, Jingchao Ni, and Dongjin Song. 2025. Towards Heterogeneous Continual Graph Learning via Meta-knowledge Distillation.arXiv preprint arXiv:2505.17458(2025)

    arXiv 2025

  35. [35]

    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 Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 1717–1727

    2022

  36. [36]

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

    2023

  37. [37]

    Xiangguo Sun, Jiawen Zhang, Xixi Wu, Hong Cheng, Yun Xiong, and Jia Li. 2023. Graph prompt learning: A comprehensive survey and beyond.arXiv preprint arXiv:2311.16534(2023)

    Pith/arXiv arXiv 2023

  38. [38]

    Jie Tang, Jimeng Sun, Chi Wang, and Zi Yang. 2009. Social influence analysis in large-scale networks. InProceedings of the 15th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 807–816

    2009

  39. [39]

    Yijun Tian, Kaiwen Dong, Chunhui Zhang, Chuxu Zhang, and Nitesh V Chawla

  40. [40]

    Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. 2018. Graph Attention Networks. In6th International Conference on Learning Representations, ICLR 2018

    2018

  41. [41]

    Petar Veličković, William Fedus, William L Hamilton, Pietro Liò, Yoshua Bengio, and R Devon Hjelm. 2018. Deep graph infomax.arXiv preprint arXiv:1809.10341 (2018)

    Pith/arXiv arXiv 2018

  42. [42]

    Haotao Wang, Ziyu Jiang, Yuning You, Yan Han, Gaowen Liu, Jayanth Srinivasa, Ramana Kompella, Zhangyang Wang, et al . 2023. Graph mixture of experts: Learning on large-scale graphs with explicit diversity modeling.Advances in Neural Information Processing Systems36 (2023), 50825–50837

    2023

  43. [43]

    Shuo Wang, Bokui Wang, Zhixiang Shen, Boyan Deng, and Zhao Kang. 2025. Multi-domain graph foundation models: Robust knowledge transfer via topology alignment.arXiv preprint arXiv:2502.02017(2025)

    Pith/arXiv arXiv 2025

  44. [44]

    Xiao Wang, Houye Ji, Chuan Shi, Bai Wang, Yanfang Ye, Peng Cui, and Philip S Yu

  45. [45]

    Xiao Wang, Nian Liu, Hui Han, and Chuan Shi. 2021. Self-supervised heteroge- neous graph neural network with co-contrastive learning. InProceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining. 1726–1736

    2021

  46. [46]

    Felix Wu, Amauri Souza, Tianyi Zhang, Christopher Fifty, Tao Yu, and Kilian Weinberger. 2019. Simplifying graph convolutional networks. InInternational conference on machine learning. 6861–6871

    2019

  47. [47]

    Xuansheng Wu, Kaixiong Zhou, Mingchen Sun, Xin Wang, and Ninghao Liu. 2023. A survey of graph prompting methods: techniques, applications, and challenges. arXiv preprint arXiv:2303.07275(2023)

    Pith/arXiv arXiv 2023

  48. [48]

    Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and Philip S Yu. 2020. A comprehensive survey on graph neural networks.IEEE transactions on neural networks and learning systems32, 1 (2020), 4–24

    2020

  49. [49]

    Yaochen Xie, Zhao Xu, Jingtun Zhang, Zhengyang Wang, and Shuiwang Ji. 2022. Self-supervised learning of graph neural networks: A unified review.IEEE trans- actions on pattern analysis and machine intelligence45, 2 (2022), 2412–2429

    2022

  50. [50]

    Carl Yang, Yuxin Xiao, Yu Zhang, Yizhou Sun, and Jiawei Han. 2020. Heteroge- neous network representation learning: A unified framework with survey and benchmark.IEEE Transactions on Knowledge and Data Engineering34, 10 (2020), 4854–4873

    2020

  51. [51]

    Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. 2020. Graph contrastive learning with augmentations.Advances in neural information processing systems33 (2020), 5812–5823

    2020

  52. [52]

    Xingtong Yu, Yuan Fang, Zemin Liu, and Xinming Zhang. 2024. Hgprompt: Bridging homogeneous and heterogeneous graphs for few-shot prompt learning. InProceedings of the AAAI conference on artificial intelligence, Vol. 38. 16578– 16586

    2024

  53. [53]

    Xingtong Yu, Zechuan Gong, Chang Zhou, Yuan Fang, and Hui Zhang. 2025. Samgpt: Text-free graph foundation model for multi-domain pre-training and cross-domain adaptation. InProceedings of the ACM on Web Conference 2025. 1142–1153

    2025

  54. [54]

    Xingtong Yu, Chang Zhou, Yuan Fang, and Xinming Zhang. 2024. Multigprompt for multi-task pre-training and prompting on graphs. InProceedings of the ACM Web Conference 2024. 515–526

    2024

  55. [55]

    Xingtong Yu, Chang Zhou, Yuan Fang, and Xinming Zhang. 2024. Text-free multi- domain graph pre-training: Toward graph foundation models.arXiv preprint arXiv:2405.13934(2024)

    Pith/arXiv arXiv 2024

  56. [56]

    Yao Zhang, Yun Xiong, Yun Ye, Tengfei Liu, Weiqiang Wang, Yangyong Zhu, and Philip S Yu. 2020. SEAL: Learning heuristics for community detection with generative adversarial networks. InProceedings of the 26th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 1103–1113

    2020

  57. [57]

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

    2024

  58. [58]

    Jitao Zhao, Di Jin, Meng Ge, Lianze Shan, Xin Wang, Dongxiao He, and Zhiyong Feng. 2024. FUG: Feature-universal graph contrastive pre-training for graphs with diverse node features.Advances in Neural Information Processing Systems37 (2024), 4003–4034

    2024

  59. [59]

    Yizhen Zheng, Shirui Pan, Vincent Lee, Yu Zheng, and Philip S Yu. 2022. Rethink- ing and scaling up graph contrastive learning: An extremely efficient approach with group discrimination.Advances in Neural Information Processing Systems35 (2022), 10809–10820

    2022

  60. [60]

    Jie Zhou, Ganqu Cui, Shengding Hu, Zhengyan Zhang, Cheng Yang, Zhiyuan Liu, Lifeng Wang, Changcheng Li, and Maosong Sun. 2020. Graph neural networks: A review of methods and applications.AI open1 (2020), 57–81

    2020

  61. [61]

    Yun Zhu, Jianhao Guo, and Siliang Tang. 2023. Sgl-pt: A strong graph learner with graph prompt tuning.arXiv preprint arXiv:2302.12449(2023)

    Pith/arXiv arXiv 2023

  62. [62]

    Yanqiao Zhu, Yichen Xu, Feng Yu, Qiang Liu, Shu Wu, and Liang Wang. 2020. Deep graph contrastive representation learning.arXiv preprint arXiv:2006.04131 (2020)

    Pith/arXiv arXiv 2020

  63. [63]

    Yanqiao Zhu, Yichen Xu, Feng Yu, Qiang Liu, Shu Wu, and Liang Wang. 2021. Graph contrastive learning with adaptive augmentation. InProceedings of the web conference 2021. 2069–2080

    2021

  64. [64]

    Chenyi Zi, Haihong Zhao, Xiangguo Sun, Yiqing Lin, Hong Cheng, and Jia Li

  65. [2000]

    Information Retrieval3, 2 (2000), 127–163

    Automating the Construction of Internet Portals with Machine Learning. Information Retrieval3, 2 (2000), 127–163

    2000

  66. [2015]

    InProceedings of the 38th international ACM SIGIR conference on research and development in information retrieval

    Image-based recommendations on styles and substitutes. InProceedings of the 38th international ACM SIGIR conference on research and development in information retrieval. 43–52

  67. [2019]

    InThe world wide web conference

    Heterogeneous graph attention network. InThe world wide web conference. 2022–2032

    2022

  68. [2022]

    Graph self-supervised learning: A survey.IEEE transactions on knowledge and data engineering35, 6 (2022), 5879–5900

    2022

  69. [2023]

    InProceedings of the AAAI conference on artificial intelligence, Vol

    Heterogeneous graph masked autoencoders. InProceedings of the AAAI conference on artificial intelligence, Vol. 37. 9997–10005

  70. [2024]

    Prog: A graph prompt learning benchmark.Advances in Neural Information Processing Systems37 (2024), 95406–95437. GPH2 Conference’17, July 2017, Washington, DC, USA A Experiment Details A.1 Datasets We evaluatedGPH2 on9datasets covering multiple domains, includ- ing academic citation networks, product networks, and knowledge graphs. The statistics are summ...

    2024