pith. machine review for the scientific record. sign in

arxiv: 2605.00731 · v1 · submitted 2026-05-01 · 💻 cs.SI · cs.AI

Recognition: unknown

Empowering Heterogeneous Graph Foundation Models via Decoupled Relation Alignment

Ziyu Zheng , Yaming Yang , Zhe Wang , Ziyu Guan , Wei Zhao
Authors on Pith no claims yet

Pith reviewed 2026-05-09 14:28 UTC · model grok-4.3

classification 💻 cs.SI cs.AI
keywords heterogeneous graphsgraph foundation modelsrelation alignmentsubspace projectiondecoupled representationcross-domain transferfew-shot learningmulti-domain graphs
0
0 comments X

The pith

Decoupled relation subspace alignment separates feature semantics from structures to fix type collapse and relation confusion in heterogeneous graph foundation models.

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

The paper argues that standard global feature alignment techniques like PCA and SVD distort type-specific meanings and break original graph topologies when applied to multi-domain heterogeneous graphs, producing type collapse and relation confusion that block effective transfer. DRSA instead decouples semantics from relations by projecting cross-type interactions into a shared low-rank subspace while splitting each aligned feature into a semantic part and a structural residual that absorbs domain-specific differences. This plug-and-play module is optimized with block coordinate descent and can be inserted before any existing graph foundation model. A sympathetic reader would care because it promises to let one pretrained model handle graphs from different domains and relation types without retraining from scratch or losing the original data geometry.

Core claim

DRSA fundamentally shifts the paradigm by decoupling feature semantics from relation structures. It introduces a dual-relation subspace projection mechanism to coordinate cross-type interactions within a shared low-rank relation subspace explicitly. Furthermore, a feature-structure decoupled representation is designed to decompose aligned features into a semantic projection component and a structural residual term, adaptively absorbing intra-domain variations. Optimized via a stable alternating minimization strategy based on Block Coordinate Descent, DRSA constructs a well-calibrated, structure-aware latent space.

What carries the argument

Dual-relation subspace projection combined with feature-structure decoupled representation, which projects cross-type interactions into a shared low-rank subspace while separating semantic projections from structural residuals.

If this is right

  • DRSA can be added as a universal preprocessing step to any state-of-the-art graph foundation model without changing its architecture.
  • The resulting latent space supports stronger cross-domain knowledge transfer across different heterogeneous graph domains.
  • Few-shot adaptation performance improves because intra-domain relation gaps are explicitly absorbed by the residual term.
  • Original topologies remain intact because alignment occurs only inside the low-rank relation subspace rather than in the full feature space.
  • Alternating minimization based on block coordinate descent yields stable optimization of the decoupled components.

Where Pith is reading between the lines

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

  • The same decoupling pattern could be tested on knowledge graphs with many relation types to see whether it reduces the need for relation-specific encoders.
  • If the residual term successfully captures variations, removing it should cause measurable drops in within-domain accuracy while preserving cross-domain gains.
  • The low-rank subspace size becomes a tunable knob that trades off alignment strength against preservation of type-specific detail.
  • Downstream tasks that rely on precise local neighborhoods may benefit more than global property prediction because topology is less distorted.

Load-bearing premise

Enforcing alignment by subspace projection and additive residuals will preserve necessary type semantics and intra-domain structures without creating new distortions in multi-domain heterogeneous graphs.

What would settle it

On a held-out multi-domain heterogeneous graph benchmark, a GFM preprocessed with DRSA shows no improvement in cross-domain few-shot accuracy over the same model without DRSA or over models using global PCA/SVD alignment.

Figures

Figures reproduced from arXiv: 2605.00731 by Wei Zhao, Yaming Yang, Zhe Wang, Ziyu Guan, Ziyu Zheng.

Figure 1
Figure 1. Figure 1: Multi-domain heterogeneous graph foundation models view at source ↗
Figure 2
Figure 2. Figure 2: Multi-domain heterogeneous graphs face two fundamen view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the DRSA framework. DRSA decouples node features into semantic projections and structural residuals, view at source ↗
Figure 4
Figure 4. Figure 4: Impact of shots number analysis. C. Performance on Few-Shot Classification To further evaluate the cross-domain generalization capability of DRSA under limited supervision, we conduct few-shot node classification experiments with varying numbers of labeled samples. We adopt GCOPE+DRSA as our representative model. The results are shown in view at source ↗
Figure 5
Figure 5. Figure 5: Ablation experiments. mechanism. Structure-driven estimation provides a stable target, while feature projection regularizes the solution and enforces consistency with the input space. This highlights the comple￾mentarity between structural and feature information, proving that decoupling their optimization is crucial for achieving stable and effective alignment. Removing the dual-relation operator causes t… view at source ↗
Figure 6
Figure 6. Figure 6: Sensitivity of the structural residual penalty view at source ↗
Figure 8
Figure 8. Figure 8: Relation structure error on ACM(paper-author) and IMDB(movie-actor). view at source ↗
read the original abstract

While Graph Foundation Models (GFMs) have achieved remarkable success in homogeneous graphs, extending them to multi-domain heterogeneous graphs (MDHGs) remains a formidable challenge due to cross-type feature shifts and intra-domain relation gaps. Existing global feature alignment methods (PCA or SVD) enforce a shared feature space blindly, which distorts type-specific semantics and disrupts original topologies, inevitably leading to "Type Collapse" and "Relation Confusion". To address these fundamental limitations, we propose Decoupled relation Subspace Alignment (DRSA), a novel, plug-and-play relation-driven alignment framework. DRSA fundamentally shifts the paradigm by decoupling feature semantics from relation structures. Specifically, it introduces a dual-relation subspace projection mechanism to coordinate cross-type interactions within a shared low-rank relation subspace explicitly. Furthermore, a feature-structure decoupled representation is designed to decompose aligned features into a semantic projection component and a structural residual term, adaptively absorbing intra-domain variations. Optimized via a stable alternating minimization strategy based on Block Coordinate Descent, DRSA constructs a well-calibrated, structure-aware latent space. Extensive experiments on multiple real-world benchmark datasets demonstrate that DRSA can be seamlessly integrated as a universal preprocessing module, significantly and consistently enhancing the cross-domain and few-shot knowledge transfer capabilities of state-of-the-art GFMs. The code is available at: https://github.com/zhengziyu77/DSRA.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 2 minor

Summary. The paper proposes Decoupled Relation Subspace Alignment (DRSA), a plug-and-play preprocessing framework for multi-domain heterogeneous graphs (MDHGs) to empower Graph Foundation Models (GFMs). It claims that existing global alignment methods like PCA/SVD cause 'Type Collapse' and 'Relation Confusion' by distorting type-specific semantics and topologies. DRSA addresses this by decoupling feature semantics from relation structures through a dual-relation subspace projection mechanism that coordinates cross-type interactions in a shared low-rank relation subspace, combined with a feature-structure decomposition (semantic projection plus structural residual) to absorb intra-domain variations. The method is optimized via block coordinate descent (BCD) alternating minimization and is shown to improve cross-domain and few-shot transfer when integrated with existing GFMs on benchmark datasets.

Significance. If the decoupling is rigorously achieved without reintroducing distortions, DRSA would represent a meaningful advance over global alignment baselines by preserving type-specific information while enabling structure-aware latent spaces. This could make it a standard preprocessing step for heterogeneous GFMs, particularly in multi-domain settings where current methods fail to transfer knowledge effectively. The plug-and-play design and code release are positive for reproducibility.

major comments (2)
  1. [Optimization and BCD strategy] The central claim of clean decoupling relies on the BCD procedure separating the dual-relation subspace projection from the structural residual term. The manuscript must explicitly show (via the objective function and update rules) that no shared term (e.g., a joint Frobenius norm) causes the residual to couple back into the low-rank subspace parameters after each block update; otherwise the claimed separation from global PCA/SVD is not guaranteed.
  2. [Dual-relation subspace projection mechanism] The dual-relation subspace projection is described as coordinating cross-type interactions in a shared low-rank space, but the manuscript should provide the precise projection operator and orthogonality constraints (likely in the method equations) to demonstrate it does not collapse type-specific semantics, with a concrete comparison to standard SVD/PCA on the same MDHG inputs.
minor comments (2)
  1. [Abstract] The GitHub link in the abstract points to DSRA while the method is abbreviated DRSA; standardize the acronym throughout.
  2. [Experiments] The abstract claims 'extensive experiments' but the manuscript should include a table summarizing dataset statistics, baseline GFMs, and quantitative gains (e.g., accuracy/F1 deltas) to support the 'significantly and consistently enhancing' statement.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We sincerely thank the referee for the thorough and constructive review. The comments raise important points about the rigor of our optimization analysis and the explicit formulation of the projection mechanism. We address each major comment below and will revise the manuscript to incorporate the requested clarifications and derivations.

read point-by-point responses

  1. Referee: [Optimization and BCD strategy] The central claim of clean decoupling relies on the BCD procedure separating the dual-relation subspace projection from the structural residual term. The manuscript must explicitly show (via the objective function and update rules) that no shared term (e.g., a joint Frobenius norm) causes the residual to couple back into the low-rank subspace parameters after each block update; otherwise the claimed separation from global PCA/SVD is not guaranteed.

    Authors: We appreciate this observation, which strengthens the presentation of our method. The current manuscript describes the BCD alternating minimization at a high level but does not include the full expanded objective and per-block update derivations. In the revised version we will explicitly write the joint objective, derive the closed-form or iterative updates for each block, and prove that the subspace projection parameters are updated independently of the structural residual term (no shared Frobenius coupling term appears across blocks). This will formally establish the separation from global PCA/SVD and confirm that type-specific semantics are not re-coupled. revision: yes

  2. Referee: [Dual-relation subspace projection mechanism] The dual-relation subspace projection is described as coordinating cross-type interactions in a shared low-rank space, but the manuscript should provide the precise projection operator and orthogonality constraints (likely in the method equations) to demonstrate it does not collapse type-specific semantics, with a concrete comparison to standard SVD/PCA on the same MDHG inputs.

    Authors: We agree that the precise operator and constraints must be stated explicitly. In the revision we will add the mathematical definition of the dual-relation subspace projection (including the shared low-rank basis and the orthogonality constraints enforced on it) together with the feature-structure decomposition. We will also insert a direct side-by-side comparison—both algebraic and on the benchmark MDHG inputs—showing how the DRSA operator preserves type-specific semantics while standard SVD/PCA induces collapse. This comparison will be placed in the method section with supporting equations and a small illustrative table. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained construction

full rationale

The abstract and described framework present DRSA as an explicit construction: dual-relation subspace projection plus semantic-projection-plus-residual decomposition, optimized by alternating minimization. No equations or steps are shown reducing to fitted parameters renamed as predictions, self-definitions, or load-bearing self-citations. The central claim (decoupling avoids Type Collapse/Relation Confusion) is advanced as a design choice with empirical support, not derived by construction from its own inputs. This matches the default expectation of non-circularity for a methodological paper.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

From abstract only; no specific free parameters or invented entities are detailed beyond the proposed DRSA components which are methodological rather than new physical entities.

axioms (2)
  • domain assumption Cross-type feature shifts and intra-domain relation gaps are the primary challenges in MDHGs for GFMs
    Explicitly stated as the formidable challenge in the abstract.
  • domain assumption Global feature alignment methods like PCA or SVD distort type-specific semantics and disrupt topologies
    Claimed as leading to Type Collapse and Relation Confusion.

pith-pipeline@v0.9.0 · 5548 in / 1330 out tokens · 39877 ms · 2026-05-09T14:28:58.608523+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

52 extracted references · 10 canonical work pages · 2 internal anchors

  1. [1]

    Interpretable and efficient heterogeneous graph convolutional network,

    Y . Yang, Z. Guan, J. Li, W. Zhao, J. Cui, and Q. Wang, “Interpretable and efficient heterogeneous graph convolutional network,”TKDE, 2021

    2021

  2. [2]

    Graph neural networks in recommender systems: a survey,

    S. Wu, F. Sun, W. Zhang, X. Xie, and B. Cui, “Graph neural networks in recommender systems: a survey,”ACM Computing Surveys, vol. 55, no. 5, pp. 1–37, 2022. 11

    2022

  3. [3]

    Truthsr: Trustworthy sequential recommender systems via user-generated multimodal content,

    M. Yan, H. Huang, Y . Liu, J. Zhao, X. Gao, C. Xu, Z. Guan, and W. Zhao, “Truthsr: Trustworthy sequential recommender systems via user-generated multimodal content,” inDatabase Systems for Advanced Applications. Singapore: Springer Nature Singapore, 2025, pp. 180–195

    2025

  4. [4]

    Graph neural networks for materials science and chemistry,

    P. Reiser, M. Neubert, A. Eberhard, L. Torresi, C. Zhou, C. Shao, H. Metni, C. van Hoesel, H. Schopmans, T. Sommeret al., “Graph neural networks for materials science and chemistry,”Communications Materials, vol. 3, no. 1, p. 93, 2022

    2022

  5. [5]

    Paaß and S

    G. Paaß and S. Giesselbach,Foundation models for natural language processing: Pre-trained language models integrating media. Springer Nature, 2023

    2023

  6. [6]

    Foundation models defining a new era in vision: a survey and outlook,

    M. Awais, M. Naseer, S. Khan, R. M. Anwer, H. Cholakkal, M. Shah, M.-H. Yang, and F. S. Khan, “Foundation models defining a new era in vision: a survey and outlook,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    2025

  7. [7]

    Graph foundation models: Concepts, opportunities and challenges,

    J. Liu, C. Yang, Z. Lu, J. Chen, Y . Li, M. Zhang, T. Bai, Y . Fang, L. Sun, P. S. Yuet al., “Graph foundation models: Concepts, opportunities and challenges,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    2025

  8. [8]

    Towards graph foundation models: A survey and beyond,

    ——, “Towards graph foundation models: A survey and beyond,”arXiv preprint arXiv:2310.11829, 2023

    work page arXiv 2023

  9. [9]

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

    X. Yu, Z. Gong, C. Zhou, Y . Fang, and H. Zhang, “Samgpt: Text-free graph foundation model for multi-domain pre-training and cross-domain adaptation,” inProceedings of the ACM on Web Conference 2025, 2025, pp. 1142–1153

    2025

  10. [10]

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

    H. Zhao, A. Chen, X. Sun, H. Cheng, and J. Li, “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, 2024, pp. 4443–4454

    2024

  11. [11]

    Multi-domain graph foundation models: Robust knowledge transfer via topology alignment,

    S. Wang, B. Wang, Z. Shen, B. Deng, and Z. Kang, “Multi-domain graph foundation models: Robust knowledge transfer via topology alignment,” arXiv preprint arXiv:2502.02017, 2025

    work page arXiv 2025

  12. [12]

    Graver: Generative graph vocabularies for robust graph foundation models fine-tuning,

    H. Yuan, Q. Sun, J. Shi, X. Fu, B. Hooi, J. Li, and P. S. Yu, “Graver: Generative graph vocabularies for robust graph foundation models fine- tuning,”arXiv preprint arXiv:2511.05592, 2025

    work page arXiv 2025

  13. [13]

    Overcoming in-memory bot- tlenecks in graph foundation models via retrieval-augmented generation,

    H. Yuan, Q. Sun, J. Tao, X. Fu, and J. Li, “Overcoming in-memory bot- tlenecks in graph foundation models via retrieval-augmented generation,” arXiv preprint arXiv:2601.15124, 2026

    work page arXiv 2026

  14. [14]

    Domain adversarial neural networks for domain generalization: When it works and how to improve,

    A. Sicilia, X. Zhao, and S. J. Hwang, “Domain adversarial neural networks for domain generalization: When it works and how to improve,” Machine Learning, vol. 112, no. 7, pp. 2685–2721, 2023

    2023

  15. [15]

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

    Z. Liu, X. Yu, Y . Fang, and X. Zhang, “Graphprompt: Unifying pre- training and downstream tasks for graph neural networks,” inProceedings of the ACM web conference 2023, 2023, pp. 417–428

    2023

  16. [16]

    Beyond single-granularity prompts: A multi-scale chain-of-thought prompt learn- ing for graph,

    Z. Zheng, Y . Yang, Z. Guan, W. Zhao, X. Huang, and W. Lu, “Beyond single-granularity prompts: A multi-scale chain-of-thought prompt learn- ing for graph,” inProceedings of the ACM Web Conference 2026, 2026, pp. 547–558

    2026

  17. [17]

    Universal prompt tuning for graph neural networks,

    T. Fang, Y . Zhang, Y . Yang, C. Wang, and L. Chen, “Universal prompt tuning for graph neural networks,”Advances in Neural Information Processing Systems, vol. 36, pp. 52 464–52 489, 2023

    2023

  18. [18]

    Self-supervised heterogeneous graph pre-training based on structural clustering,

    Y . Yang, Z. Guan, Z. Wang, W. Zhao, C. Xu, W. Lu, and J. Huang, “Self-supervised heterogeneous graph pre-training based on structural clustering,”Advances in Neural Information Processing Systems, vol. 35, pp. 16 962–16 974, 2022

    2022

  19. [19]

    Enhancing homophily- heterophily separation: Relation-aware learning in heterogeneous graphs,

    Z. Zheng, Y . Yang, Z. Guan, W. Zhao, and W. Lu, “Enhancing homophily- heterophily separation: Relation-aware learning in heterogeneous graphs,” inProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V .2, 2025, p. 4050–4061

    2025

  20. [20]

    Heterogeneous graph attention network,

    X. Wang, H. Ji, C. Shi, B. Wang, Y . Ye, P. Cui, and P. S. Yu, “Heterogeneous graph attention network,” inThe World Wide Web Conference, 2019, pp. 2022–2032

    2019

  21. [21]

    Self-supervised heterogeneous graph neural network with co-contrastive learning,

    X. Wang, N. Liu, H. Han, and C. Shi, “Self-supervised heterogeneous graph neural network with co-contrastive learning,” inProceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, 2021, pp. 1726–1736

    2021

  22. [22]

    Self-supervised heterogeneous graph learning: a homophily and het- erogeneity view,

    Y . Mo, F. Nie, P. Hu, H. T. Shen, Z. Zhang, X. Wang, and X. Zhu, “Self-supervised heterogeneous graph learning: a homophily and het- erogeneity view,” inThe Twelfth International Conference on Learning Representations

  23. [23]

    Progressive feature alignment for unsupervised domain adaptation,

    C. Chen, W. Xie, W. Huang, Y . Rong, X. Ding, Y . Huang, T. Xu, and J. Huang, “Progressive feature alignment for unsupervised domain adaptation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 627–636

    2019

  24. [24]

    Principal component analysis,

    H. Abdi and L. J. Williams, “Principal component analysis,”Wiley interdisciplinary reviews: computational statistics, vol. 2, no. 4, pp. 433–459, 2010

    2010

  25. [25]

    On the early history of the singular value decomposition,

    G. W. Stewart, “On the early history of the singular value decomposition,” SIAM review, vol. 35, no. 4, pp. 551–566, 1993

    1993

  26. [26]

    Heteroge- neous graph neural network,

    C. Zhang, D. Song, C. Huang, A. Swami, and N. V . Chawla, “Heteroge- neous graph neural network,” inProceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, 2019, pp. 793–803

    2019

  27. [27]

    Heterogeneous graph transformer,

    Z. Hu, Y . Dong, K. Wang, and Y . Sun, “Heterogeneous graph transformer,” inProceedings of the web conference 2020, 2020, pp. 2704–2710

    2020

  28. [28]

    Magnn: Metapath aggregated graph neural network for heterogeneous graph embedding,

    X. Fu, J. Zhang, Z. Meng, and I. King, “Magnn: Metapath aggregated graph neural network for heterogeneous graph embedding,” inProceed- ings of the web conference 2020, 2020, pp. 2331–2341

    2020

  29. [29]

    Convergence of a block coordinate descent method for nondifferentiable minimization,

    P. Tseng, “Convergence of a block coordinate descent method for nondifferentiable minimization,”Journal of optimization theory and applications, vol. 109, no. 3, pp. 475–494, 2001

    2001

  30. [30]

    Unsupervised attributed multiplex network embedding,

    C. Park, D. Kim, J. Han, and H. Yu, “Unsupervised attributed multiplex network embedding,” inProceedings of the AAAI conference on artificial intelligence, vol. 34, no. 04, 2020, pp. 5371–5378

    2020

  31. [31]

    Hdmi: High-order deep multiplex infomax,

    B. Jing, C. Park, and H. Tong, “Hdmi: High-order deep multiplex infomax,” inProceedings of the Web Conference 2021, 2021, pp. 2414– 2424

    2021

  32. [32]

    Heterogeneous graph masked autoencoders,

    Y . Tian, K. Dong, C. Zhang, C. Zhang, and N. V . Chawla, “Heterogeneous graph masked autoencoders,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 8, 2023, pp. 9997–10 005

    2023

  33. [33]

    Mug: Meta- path-aware universal heterogeneous graph pre-training,

    L. Shan, J. Zhao, D. He, Y . Huang, Z. Feng, and W. Zhang, “Mug: Meta- path-aware universal heterogeneous graph pre-training,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 30, 2026, pp. 25 260–25 268

    2026

  34. [34]

    Revisiting self-supervised heterogeneous graph learning from spectral clustering perspective,

    Y . Mo, Z. Lu, R. Yu, X. Zhu, and X. Wang, “Revisiting self-supervised heterogeneous graph learning from spectral clustering perspective,” Advances in Neural Information Processing Systems, vol. 37, pp. 43 133– 43 163, 2024

    2024

  35. [35]

    Reserving-masking-reconstruction model for self-supervised heterogeneous graph representation,

    H. Duan, C. Xie, and L. Li, “Reserving-masking-reconstruction model for self-supervised heterogeneous graph representation,” inProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2024, pp. 689–700

    2024

  36. [36]

    Hetgpt: Harnessing the power of prompt tuning in pre-trained heterogeneous graph neural networks,

    Y . Ma, N. Yan, J. Li, M. Mortazavi, and N. V . Chawla, “Hetgpt: Harnessing the power of prompt tuning in pre-trained heterogeneous graph neural networks,” inProceedings of the ACM Web Conference 2024, 2024, pp. 1015–1023

    2024

  37. [37]

    Hgprompt: Bridging homogeneous and heterogeneous graphs for few-shot prompt learning,

    X. Yu, Y . Fang, Z. Liu, and X. Zhang, “Hgprompt: Bridging homogeneous and heterogeneous graphs for few-shot prompt learning,” inProceedings of the AAAI conference on artificial intelligence, vol. 38, no. 15, 2024, pp. 16 578–16 586

    2024

  38. [38]

    Edge prompt tuning for graph neural networks,

    X. Fu, Y . He, and J. Li, “Edge prompt tuning for graph neural networks,” arXiv preprint arXiv:2503.00750, 2025

    work page arXiv 2025

  39. [39]

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

    M. Sun, K. Zhou, X. He, Y . Wang, and X. 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, 2022, pp. 1717–1727

    2022

  40. [40]

    Text-free multi-domain graph pre-training: Toward graph foundation models,

    X. Yu, C. Zhou, Y . Fang, and X. Zhang, “Text-free multi-domain graph pre-training: Toward graph foundation models,”arXiv preprint arXiv:2405.13934, 2024

    work page arXiv 2024

  41. [41]

    Saˆ 2gfm: Enhancing robust graph foundation models with structure-aware semantic augmentation,

    J. Shi, Q. Sun, H. Yuan, and X. Fu, “Saˆ 2gfm: Enhancing robust graph foundation models with structure-aware semantic augmentation,”arXiv preprint arXiv:2512.07857, 2025

    work page arXiv 2025

  42. [42]

    How much can transfer? bridge: Bounded multi-domain graph foundation model with generalization guarantees,

    H. Yuan, Q. Sun, J. Shi, X. Fu, B. Hooi, J. Li, and P. S. Yu, “How much can transfer? bridge: Bounded multi-domain graph foundation model with generalization guarantees,” inForty-second International Conference on Machine Learning

  43. [43]

    Random projection in dimensionality reduction: applications to image and text data,

    E. Bingham and H. Mannila, “Random projection in dimensionality reduction: applications to image and text data,” inProceedings of the seventh ACM SIGKDD international conference on Knowledge discovery and data mining, 2001, pp. 245–250

    2001

  44. [44]

    Random features for large-scale kernel machines,

    A. Rahimi and B. Recht, “Random features for large-scale kernel machines,”Advances in neural information processing systems, vol. 20, 2007

    2007

  45. [45]

    Extensions of lipschitz mappings into a hilbert space,

    W. B. Johnson, J. Lindenstrausset al., “Extensions of lipschitz mappings into a hilbert space,”Contemporary mathematics, vol. 26, no. 189-206, p. 1, 1984

    1984

  46. [46]

    Bridging input feature spaces towards graph foundation models,

    M. Eliasof, K. S. I. Mantri, B. Bevilacqua, B. Ribeiro, and C.-B. Schönlieb, “Bridging input feature spaces towards graph foundation models,” inThe Fourteenth International Conference on Learning Representations

  47. [47]

    Network schema preserving heterogeneous information network embedding,

    J. Zhao, X. Wang, C. Shi, Z. Liu, and Y . Ye, “Network schema preserving heterogeneous information network embedding,” inInternational joint conference on artificial intelligence (IJCAI), 2020. 12

    2020

  48. [48]

    Rhine: Relation structure-aware heterogeneous information network embedding,

    C. Shi, Y . Lu, L. Hu, Z. Liu, and H. Ma, “Rhine: Relation structure-aware heterogeneous information network embedding,”IEEE Transactions on Knowledge and Data Engineering, vol. 34, no. 1, pp. 433–447, 2020

    2020

  49. [49]

    metapath2vec: Scalable representation learning for heterogeneous networks,

    Y . Dong, N. V . Chawla, and A. Swami, “metapath2vec: Scalable representation learning for heterogeneous networks,” inSIGKDD. ACM, 2017, pp. 135–144

    2017

  50. [50]

    Graph Attention Networks

    P. Veli ˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P. Lio, and Y . Bengio, “Graph attention networks,”arXiv preprint arXiv:1710.10903, 2017

    work page internal anchor Pith review arXiv 2017

  51. [51]

    Semi-Supervised Classification with Graph Convolutional Networks

    T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,”arXiv preprint arXiv:1609.02907, 2016

    work page internal anchor Pith review arXiv 2016

  52. [52]

    arXiv preprint arXiv:1803.00676 , year=

    M. Ren, E. Triantafillou, S. Ravi, J. Snell, K. Swersky, J. B. Tenenbaum, H. Larochelle, and R. S. Zemel, “Meta-learning for semi-supervised few-shot classification,”arXiv preprint arXiv:1803.00676, 2018

    work page arXiv 2018