pith. sign in

arxiv: 2606.00808 · v1 · pith:R4HSCIJEnew · submitted 2026-05-30 · 💻 cs.LG

Safe-Subspace Pseudo-Label Refinement for Source-Free Graph Domain Adaptation

Yingxu Wang , Xinwang Liu , Siyang Gao , Nan Yin This is my paper

Pith reviewed 2026-06-28 19:18 UTC · model grok-4.3

classification 💻 cs.LG
keywords source-free domain adaptationgraph domain adaptationpseudo-label refinementsafe subspacedomain shiftgraph neural networksself-training
0
0 comments X

The pith

A safe subspace lets source-free graph adaptation use reliable pseudo-labels by restricting hard supervision to samples with both semantic and structural support.

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

The paper establishes that source-free graph domain adaptation succeeds when pseudo-label supervision is limited to a confidence-consistent safe subspace rather than applied across the full target domain. This matters because feature and topological shifts often make source predictions confidently wrong, and message passing on graphs can then spread those errors during self-training. The authors derive a target-risk decomposition separating safe-subspace fitting error, selected-label noise, and uncertain-set risk. Guided by the decomposition they build S²PLR, which estimates semantic reliability with source-committee statistics, learns structural representations via contrastive learning, verifies labels by neighborhood consistency, and handles remaining samples with soft regularization.

Core claim

The paper claims that a confidence-consistent safe subspace exists on which pseudo-label noise can be controlled under restricted posterior discrepancy, and that applying hard supervision only inside this subspace while using noise-tolerant soft regularization on the uncertain remainder produces robust adaptation without any source-graph access.

What carries the argument

The confidence-consistent safe subspace, which isolates target samples supported by both semantic committee evidence and intrinsic structural consistency so that hard pseudo-labels remain reliable.

If this is right

  • Hard pseudo-label supervision is applied only where semantic reliability and neighborhood consistency both hold.
  • Source graphs remain inaccessible throughout adaptation.
  • Uncertain samples receive soft regularization instead of unreliable hard labels.
  • The resulting method reports competitive accuracy on image and real-world graph benchmarks under multiple domain-shift types.

Where Pith is reading between the lines

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

  • The safe-subspace selection principle could be tested in non-graph source-free adaptation settings.
  • The approach may reduce reliance on source data in privacy-restricted applications.
  • Dynamic re-estimation of the safe subspace across training epochs is a natural next measurement.

Load-bearing premise

A confidence-consistent safe subspace exists on which pseudo-label noise stays controllable under restricted posterior discrepancy.

What would settle it

An experiment in which no identifiable subspace yields lower adaptation error than applying hard pseudo-labels to the entire target set.

Figures

Figures reproduced from arXiv: 2606.00808 by Nan Yin, Siyang Gao, Xinwang Liu, Yingxu Wang.

Figure 1
Figure 1. Figure 1: T-SNE visualizations of S2PLR and baselines on the Mutagenicity dataset. TABLE III: Aggregate comparison across all reported transfer tasks. Bold results indicate the best performance. Method Node Avg. Edge Avg. Overall Avg. Avg. Rank S2PLR W/T/L WL 49.29 43.23 46.10 13.60 138/0/0 GCN 51.36 53.39 52.76 12.55 134/4/0 GIN 51.79 54.25 53.42 12.61 134/4/0 GMT 53.62 55.12 54.55 11.97 137/1/0 CIN 51.65 54.83 53.… view at source ↗
Figure 2
Figure 2. Figure 2: High-confidence false pseudo-labels on Mutagenicity. [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Sensitivity analysis of the confidence threshold [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of domain shifts across different types. (a,b) Node and edge distribution shifts between sub-datasets of [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Sensitivity analysis of the confidence threshold [PITH_FULL_IMAGE:figures/full_fig_p022_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Sensitivity analysis of the neighborhood-consistency threshold [PITH_FULL_IMAGE:figures/full_fig_p022_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Sensitivity analysis of the number of source experts [PITH_FULL_IMAGE:figures/full_fig_p023_7.png] view at source ↗
read the original abstract

Source-free graph domain adaptation (SF-GDA) aims to adapt source-trained graph models to unlabeled target graphs when source graphs are no longer accessible. A central obstacle is pseudo-label reliability: under feature and topological shifts, source-induced predictions may become confidently wrong, and indiscriminate self-training can amplify systematic errors through graph message passing. This paper studies SF-GDA from a selective pseudo-labeling perspective. Instead of assuming globally bounded pseudo-label noise over the entire target domain, we identify a confidence-consistent safe subspace on which pseudo-label noise can be controlled under restricted posterior discrepancy, and derive a target-risk decomposition that separates safe-subspace fitting error, selected-label noise, and uncertain-set risk. Guided by this analysis, we propose SafeSubspace Pseudo-Label Refinement (S$^2$PLR), a source-free graph adaptation framework that applies hard pseudo-label supervision only to target graphs supported by both semantic and structural evidence. Specifically, S$^2$PLR estimates semantic reliability using source-committee confidence and disagreement, learns a targetintrinsic structural representation via graph contrastive learning, verifies pseudo-labels through neighborhood consistency, and exploits the remaining uncertain samples with noise-tolerant soft regularization rather than unreliable hard labels. Experiments on image and real-world graph benchmarks under different domain shifts demonstrate that S$^2$PLR achieves robust and competitive performance across diverse source-free transfer settings.

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

1 major / 2 minor

Summary. The paper proposes Safe-Subspace Pseudo-Label Refinement (S²PLR) for source-free graph domain adaptation. It identifies a confidence-consistent safe subspace on target graphs using source-committee confidence and disagreement together with target-intrinsic checks (graph contrastive learning and neighborhood consistency). A target-risk decomposition is derived that separates safe-subspace fitting error, selected-label noise, and uncertain-set risk; hard pseudo-label supervision is applied only inside the safe subspace while uncertain samples receive noise-tolerant soft regularization. Experiments on image and real-world graph benchmarks under various domain shifts report competitive performance.

Significance. If the risk decomposition and safe-subspace construction are valid, the work supplies a principled selective mechanism for controlling pseudo-label noise in SF-GDA without source data access. The explicit separation of risk terms and the combination of semantic and structural evidence are potentially useful contributions to graph domain adaptation.

major comments (1)
  1. [§3.2] §3.2 (target-risk decomposition): the claim that pseudo-label noise is controlled under restricted posterior discrepancy in the identified safe subspace is load-bearing for cleanly separating the three risk terms. The construction (committee confidence + contrastive learning + neighborhood consistency) does not include explicit bounds showing that topological shifts cannot produce large posterior discrepancy even on high-confidence nodes; without such bounds or a verification procedure the decomposition's separation guarantee is not established.
minor comments (2)
  1. [Abstract] Abstract: 'targetintrinsic' should be hyphenated as 'target-intrinsic'.
  2. [§3] Notation for the safe-subspace indicator and the posterior-discrepancy term should be introduced once with a clear definition before being used in the risk decomposition.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive comment on the target-risk decomposition. We address the concern point by point below.

read point-by-point responses

  1. Referee: [§3.2] §3.2 (target-risk decomposition): the claim that pseudo-label noise is controlled under restricted posterior discrepancy in the identified safe subspace is load-bearing for cleanly separating the three risk terms. The construction (committee confidence + contrastive learning + neighborhood consistency) does not include explicit bounds showing that topological shifts cannot produce large posterior discrepancy even on high-confidence nodes; without such bounds or a verification procedure the decomposition's separation guarantee is not established.

    Authors: The decomposition is explicitly derived under the assumption of restricted posterior discrepancy within the safe subspace, which enables separation of the three risk terms. The multi-signal construction (committee confidence/disagreement, contrastive learning, neighborhood consistency) is intended to identify nodes satisfying this assumption in practice. We do not provide explicit theoretical bounds proving that the criteria prevent large posterior discrepancy under arbitrary topological shifts; the separation guarantee therefore rests on the modeling assumption rather than a proven bound. In revision we will clarify this assumption in §3.2 and add an explicit statement that formal bounds are not derived. revision: partial

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The provided abstract and context describe a target-risk decomposition that separates safe-subspace fitting error, selected-label noise, and uncertain-set risk under an assumption of restricted posterior discrepancy on an identified subspace. The subspace is constructed explicitly via source-committee confidence/disagreement, graph contrastive learning, and neighborhood consistency checks. No equations or self-citations are available to exhibit a reduction of any prediction or result to its inputs by construction (e.g., no fitted parameter renamed as prediction, no self-definitional loop, no load-bearing self-citation chain). The derivation supplies independent analytical content guiding selective pseudo-labeling rather than tautologically restating the method inputs. This is the normal case of a self-contained framework.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the unverified existence of a safe subspace whose noise is bounded by restricted posterior discrepancy; this is treated as a domain assumption rather than derived. No free parameters or invented entities are explicitly listed in the abstract.

axioms (1)
  • domain assumption A confidence-consistent safe subspace exists on which pseudo-label noise can be controlled under restricted posterior discrepancy.
    Invoked to derive the target-risk decomposition that separates safe-subspace fitting error, selected-label noise, and uncertain-set risk.

pith-pipeline@v0.9.1-grok · 5781 in / 1360 out tokens · 14400 ms · 2026-06-28T19:18:52.257589+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

72 extracted references · 11 canonical work pages · 2 internal anchors

  1. [1]

    Transferable representation learning with deep adapta- tion networks,

    M. Long, Y . Cao, Z. Cao, J. Wang, and M. I. Jordan, “Transferable representation learning with deep adapta- tion networks,”IEEE transactions on pattern analysis and machine intelligence, vol. 41, no. 12, pp. 3071–3085, 2018

    2018

  2. [2]

    A comprehensive survey on graph neural networks,

    Z. Wu, S. Pan, F. Chen, G. Long, C. Zhang, and S. Y . Philip, “A comprehensive survey on graph neural networks,”IEEE Transactions on Neural Networks and Learning Systems, vol. 32, no. 1, pp. 4–24, 2020

    2020

  3. [3]

    Self- supervised learning of graph neural networks: A unified review,

    Y . Xie, Z. Xu, J. Zhang, Z. Wang, and S. Ji, “Self- supervised learning of graph neural networks: A unified review,”IEEE transactions on pattern analysis and ma- chine intelligence, vol. 45, no. 2, pp. 2412–2429, 2022

    2022

  4. [4]

    Empowering graph rep- resentation learning with test-time graph transformation

    W. Jin, T. Zhao, J. Ding, Y . Liu, J. Tang, and N. Shah, “Empowering graph representation learning with test-time graph transformation,”arXiv preprint arXiv:2210.03561, 2022

    work page arXiv 2022

  5. [5]

    Graph structure learning for robust graph neural networks,

    W. Jin, Y . Ma, X. Liu, X. Tang, S. Wang, and J. Tang, “Graph structure learning for robust graph neural networks,” inProceedings of the International ACM SIGKDD Conference on Knowledge Discovery & Data Mining, 2020, pp. 66–74

    2020

  6. [6]

    Degree distribution based spiking graph networks for domain adaptation,

    Y . Wang, S. Liu, M. Wang, S. Liang, and N. Yin, “Degree distribution based spiking graph networks for domain adaptation,”arXiv preprint arXiv:2410.06883, 2024

    work page arXiv 2024

  7. [7]

    Revisiting, benchmarking and understanding unsuper- vised graph domain adaptation,

    M. Liu, Z. Zhang, J. Tang, J. Bu, B. He, and S. Zhou, “Revisiting, benchmarking and understanding unsuper- vised graph domain adaptation,”Proceedings of the Conference on Neural Information Processing Systems, vol. 37, pp. 89 408–89 436, 2024

    2024

  8. [8]

    Coco: A coupled contrastive frame- work for unsupervised domain adaptive graph classifica- tion,

    N. Yin, L. Shen, M. Wang, L. Lan, Z. Ma, C. Chen, X.-S. Hua, and X. Luo, “Coco: A coupled contrastive frame- work for unsupervised domain adaptive graph classifica- tion,” inProceedings of the International Conference on Machine Learning, 2023, pp. 40 040–40 053

    2023

  9. [9]

    DSBD: Dual-Aligned Structural Basis Distillation for Graph Domain Adaptation

    Y . Wang, K. Zhang, J. Huang, M. Wang, M. Xiao, S. Gao, and N. Yin, “Dsbd: Dual-aligned structural basis distillation for graph domain adaptation,”arXiv preprint arXiv:2604.03154, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  10. [10]

    DisRFM: Polar Riemannian Flow Matching for Structure-Preserving Graph Domain Adaptation

    Y . Wang, X. Liu, M. Wang, S. Gao, and N. Yin, “Rie- mannian flow matching for disentangled graph domain adaptation,”arXiv preprint arXiv:2602.00656, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  11. [11]

    A com- prehensive survey on source-free domain adaptation,

    J. Li, Z. Yu, Z. Du, L. Zhu, and H. T. Shen, “A com- prehensive survey on source-free domain adaptation,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 8, pp. 5743–5762, 2024

    2024

  12. [12]

    Stable neighbor denoising for source-free domain adaptive segmentation,

    D. Zhao, S. Wang, Q. Zang, L. Jiao, N. Sebe, and Z. Zhong, “Stable neighbor denoising for source-free domain adaptive segmentation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 23 416–23 427

    2024

  13. [13]

    Source-free domain adaptation with frozen multimodal foundation model,

    S. Tang, W. Su, M. Ye, and X. Zhu, “Source-free domain adaptation with frozen multimodal foundation model,” 13 inThe IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 23 711–23 720

    2024

  14. [14]

    Domain adaptation without source data,

    Y . Kim, D. Cho, K. Han, P. Panda, and S. Hong, “Domain adaptation without source data,”IEEE Transactions on Artificial Intelligence, vol. 2, no. 6, pp. 508–518, 2021

    2021

  15. [15]

    Source free graph unsupervised domain adaptation,

    H. Mao, L. Du, Y . Zheng, Q. Fu, Z. Li, X. Chen, S. Han, and D. Zhang, “Source free graph unsupervised domain adaptation,” inProceedings of the International ACM Conference on Web Search & Data Mining, 2024, pp. 520–528

    2024

  16. [16]

    Gala: Graph diffusion-based alignment with jigsaw for source-free domain adaptation,

    J. Luo, Y . Gu, X. Luo, W. Ju, Z. Xiao, Y . Zhao, J. Yuan, and M. Zhang, “Gala: Graph diffusion-based alignment with jigsaw for source-free domain adaptation,”IEEE Transactions on Pattern Analysis and Machine Intelli- gence, vol. 46, no. 12, pp. 9038–9051, 2024

    2024

  17. [17]

    Rank and align: Towards effec- tive source-free graph domain adaptation,

    J. Luo, Z. Xiao, Y . Wang, X. Luo, J. Yuan, W. Ju, L. Liu, and M. Zhang, “Rank and align: Towards effec- tive source-free graph domain adaptation,”arXiv preprint arXiv:2408.12185, 2024

    work page arXiv 2024

  18. [18]

    Nested graph pseudo-label refinement for noisy label domain adaptation learning,

    Y . Wang, M. Wang, Z. Huang, S. Liu, and N. Yin, “Nested graph pseudo-label refinement for noisy label domain adaptation learning,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 31, 2026, pp. 26 697–26 705

    2026

  19. [19]

    When source-free domain adap- tation meets learning with noisy labels,

    L. Yi, G. Xu, P. Xu, J. Li, R. Pu, C. Ling, A. I. McLeod, and B. Wang, “When source-free domain adap- tation meets learning with noisy labels,”arXiv preprint arXiv:2301.13381, 2023

    work page arXiv 2023

  20. [20]

    Unraveling the mysteries of label noise in source-free domain adaptation: Theory and practice,

    G. Xu, L. Yi, P. Xu, J. Li, R. Pu, C. Shui, C. Ling, A. I. McLeod, and B. Wang, “Unraveling the mysteries of label noise in source-free domain adaptation: Theory and practice,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    2025

  21. [21]

    Tackling noisy labels with network parameter additive decomposition,

    J. Wang, X. Xia, L. Lan, X. Wu, J. Yu, W. Yang, B. Han, and T. Liu, “Tackling noisy labels with network parameter additive decomposition,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 9, pp. 6341–6354, 2024

    2024

  22. [22]

    Ex- ploiting the intrinsic neighborhood structure for source- free domain adaptation,

    S. Yang, J. Van de Weijer, L. Herranz, S. Juiet al., “Ex- ploiting the intrinsic neighborhood structure for source- free domain adaptation,”Advances in neural information processing systems, vol. 34, pp. 29 393–29 405, 2021

    2021

  23. [23]

    Universal source-free domain adaptation,

    J. N. Kundu, N. Venkat, R. V . Babuet al., “Universal source-free domain adaptation,” inThe IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, 2020, pp. 4544–4553

    2020

  24. [24]

    Generalized source-free domain adaptation,

    S. Yang, Y . Wang, J. Van De Weijer, L. Herranz, and S. Jui, “Generalized source-free domain adaptation,” in The IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 8978–8987

    2021

  25. [25]

    Fg- plfa: Fine-grained pseudo-labeling and feature alignment for source-free unsupervised domain adaptation,

    Z. Wen, Q. Li, Y . Wang, H. Shao, and G. Sun, “Fg- plfa: Fine-grained pseudo-labeling and feature alignment for source-free unsupervised domain adaptation,”IEEE Transactions on Neural Networks and Learning Systems, 2025

    2025

  26. [26]

    Lead: Learning decomposition for source- free universal domain adaptation,

    S. Qu, T. Zou, L. He, F. R ¨ohrbein, A. Knoll, G. Chen, and C. Jiang, “Lead: Learning decomposition for source- free universal domain adaptation,” inThe IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 23 334–23 343

    2024

  27. [27]

    arXiv preprint arXiv:2602.08431 (2026)

    Y . Wang, K. Zhang, M. Wang, S. Gao, and N. Yin, “Usbd: Universal structural basis distillation for source-free graph domain adaptation,”arXiv preprint arXiv:2602.08431, 2026

    work page arXiv 2026

  28. [28]

    Sf(da) 2: Source-free domain adaptation through the lens of data augmentation,

    U. Hwang, J. Lee, J. Shin, and S. Yoon, “Sf(da) 2: Source-free domain adaptation through the lens of data augmentation,”arXiv preprint arXiv:2403.10834, 2024

    work page arXiv 2024

  29. [29]

    Adaptive adversar- ial network for source-free domain adaptation,

    H. Xia, H. Zhao, and Z. Ding, “Adaptive adversar- ial network for source-free domain adaptation,” inThe IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 9010–9019

    2021

  30. [30]

    Uncertainty-guided source-free domain adaptation,

    S. Roy, M. Trapp, A. Pilzer, J. Kannala, N. Sebe, E. Ricci, and A. Solin, “Uncertainty-guided source-free domain adaptation,” inEuropean conference on computer vision. Springer, 2022, pp. 537–555

    2022

  31. [31]

    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,” in Proceedings of the ACM Web Conference, 2025, pp. 1142–1153

    2025

  32. [32]

    Aggregate to adapt: Node-centric aggregation for multi-source-free graph domain adapta- tion,

    Z. Zhang and B. He, “Aggregate to adapt: Node-centric aggregation for multi-source-free graph domain adapta- tion,” inProceedings of the ACM Web Conference, 2025, pp. 4420–4431

    2025

  33. [33]

    Robust cross supervision with target mining for source-free graph domain adaptation,

    J. Luo, H. Tao, X. Luo, Y . Zhao, Z. Xiao, D. He, W. Ju, C. Chen, X.-S. Hua, and M. Zhang, “Robust cross supervision with target mining for source-free graph domain adaptation,”IEEE Transactions on Knowledge and Data Engineering, 2026

    2026

  34. [34]

    Source-free progressive graph learning for open-set domain adaptation,

    Y . Luo, Z. Wang, Z. Chen, Z. Huang, and M. Bak- tashmotlagh, “Source-free progressive graph learning for open-set domain adaptation,”IEEE Transactions on Pat- tern Analysis and Machine Intelligence, vol. 45, no. 9, pp. 11 240–11 255, 2023

    2023

  35. [35]

    To-ugda: target-oriented unsupervised graph domain adaptation,

    Z. Zeng, J. Xie, Z. Yang, T. Ma, and D. Chen, “To-ugda: target-oriented unsupervised graph domain adaptation,” Scientific Reports, vol. 14, no. 1, p. 9165, 2024

    2024

  36. [36]

    Smoothness really matters: A simple yet effective approach for unsupervised graph domain adaptation,

    W. Chen, G. Ye, Y . Wang, Z. Zhang, L. Zhang, D. Wang, Z. Zhang, and F. Zhuang, “Smoothness really matters: A simple yet effective approach for unsupervised graph domain adaptation,” inProceedings of the AAAI Confer- ence on Artificial Intelligence, vol. 39, no. 15, 2025, pp. 15 875–15 883

    2025

  37. [37]

    A parametrical model for instance-dependent label noise,

    S. Yang, S. Wu, E. Yang, B. Han, Y . Liu, M. Xu, G. Niu, and T. Liu, “A parametrical model for instance-dependent label noise,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 12, pp. 14 055–14 068, 2023

    2023

  38. [38]

    Robust node classification on graph data with graph and label noise,

    Y . Zhu, L. Feng, Z. Deng, Y . Chen, R. Amor, and M. Witbrock, “Robust node classification on graph data with graph and label noise,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 38, 2024, pp. 17 220–17 227

    2024

  39. [39]

    Learning from noisy labels with deep neural networks: A survey,

    H. Song, M. Kim, D. Park, Y . Shin, and J.-G. Lee, “Learning from noisy labels with deep neural networks: A survey,”IEEE transactions on neural networks and 14 learning systems, vol. 34, no. 11, pp. 8135–8153, 2022

    2022

  40. [40]

    A survey of label-noise representa- tion learning: Past, present and future,

    B. Han, Q. Yao, T. Liu, G. Niu, I. W. Tsang, J. T. Kwok, and M. Sugiyama, “A survey of label-noise representa- tion learning: Past, present and future,”arXiv preprint arXiv:2011.04406, 2020

    work page arXiv 2011

  41. [41]

    Sport: A subgraph perspective on graph classification with label noise,

    N. Yin, L. Shen, C. Chen, X.-S. Hua, and X. Luo, “Sport: A subgraph perspective on graph classification with label noise,”ACM Transactions on Knowledge Discovery from Data, vol. 18, no. 9, pp. 1–20, 2024

    2024

  42. [42]

    Learning from noisy exam- ples,

    D. Angluin and P. Laird, “Learning from noisy exam- ples,”Machine learning, vol. 2, no. 4, pp. 343–370, 1988

    1988

  43. [43]

    Learning with noisy labels,

    N. Natarajan, I. S. Dhillon, P. K. Ravikumar, and A. Tewari, “Learning with noisy labels,”Advances in neural information processing systems, vol. 26, 2013

    2013

  44. [44]

    Gradient-based learning applied to document recogni- tion,

    Y . LeCun, L. Bottou, Y . Bengio, and P. Haffner, “Gradient-based learning applied to document recogni- tion,”Proceedings of the IEEE, 2002

    2002

  45. [45]

    Learning multiple layers of features from tiny images,

    A. Krizhevsky, G. Hintonet al., “Learning multiple layers of features from tiny images,” 2009

    2009

  46. [46]

    Dream: a dual variational framework for unsuper- vised graph domain adaptation,

    N. Yin, L. Shen, M. Wang, X. Liu, C. Chen, and X.-S. Hua, “Dream: a dual variational framework for unsuper- vised graph domain adaptation,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    2025

  47. [47]

    Weisfeiler-lehman graph kernels

    N. Shervashidze, P. Schweitzer, E. J. Van Leeuwen, K. Mehlhorn, and K. M. Borgwardt, “Weisfeiler-lehman graph kernels.”Journal of Machine Learning Research, vol. 12, no. 9, 2011

    2011

  48. [48]

    Semi-supervised classifica- tion with graph convolutional networks,

    T. N. Kipf and M. Welling, “Semi-supervised classifica- tion with graph convolutional networks,” inProceedings of the International Conference on Machine Learning, 2017

    2017

  49. [49]

    How powerful are graph neural networks?

    K. Xu, W. Hu, J. Leskovec, and S. Jegelka, “How powerful are graph neural networks?” inProceedings of the International Conference on Machine Learning, 2018

    2018

  50. [50]

    Weisfeiler and lehman go cellular: Cw networks,

    C. Bodnar, F. Frasca, N. Otter, Y . Wang, P. Lio, G. F. Montufar, and M. Bronstein, “Weisfeiler and lehman go cellular: Cw networks,”Proceedings of the Conference on Neural Information Processing Systems, vol. 34, pp. 2625–2640, 2021

    2021

  51. [51]

    Accurate learning of graph representations with graph multiset pooling,

    J. Baek, M. Kang, and S. J. Hwang, “Accurate learning of graph representations with graph multiset pooling,” arXiv preprint arXiv:2102.11533, 2021

    work page arXiv 2021

  52. [52]

    Path neural networks: Expressive and accurate graph neural networks,

    G. Michel, G. Nikolentzos, J. F. Lutzeyer, and M. Vazir- giannis, “Path neural networks: Expressive and accurate graph neural networks,” inProceedings of the Interna- tional Conference on Machine Learning. PMLR, 2023, pp. 24 737–24 755

    2023

  53. [53]

    Revisiting source-free domain adaptation: a new per- spective via uncertainty control,

    G. Xu, H. Guo, L. Yi, C. Ling, B. Wang, and G. Yi, “Revisiting source-free domain adaptation: a new per- spective via uncertainty control,” inProceedings of the International Conference on Learning Representations, vol. 2025, 2025, pp. 92 900–92 939

    2025

  54. [54]

    Collaborate to adapt: Source-free graph domain adaptation via bi-directional adaptation,

    Z. Zhang, M. Liu, A. Wang, H. Chen, Z. Li, J. Bu, and B. He, “Collaborate to adapt: Source-free graph domain adaptation via bi-directional adaptation,” inProceedings of the ACM Web Conference, 2024, pp. 664–675

    2024

  55. [55]

    Deal: An unsupervised domain adaptive framework for graph-level classification,

    N. Yin, L. Shen, B. Li, M. Wang, X. Luo, C. Chen, Z. Luo, and X.-S. Hua, “Deal: An unsupervised domain adaptive framework for graph-level classification,” in Proceedings of the ACM International Conference on Multimedia, 2022, pp. 3470–3479

    2022

  56. [56]

    Distinguishing enzyme structures from non-enzymes without alignments,

    P. D. Dobson and A. J. Doig, “Distinguishing enzyme structures from non-enzymes without alignments,”Jour- nal of molecular biology, vol. 330, no. 4, pp. 771–783, 2003

    2003

  57. [57]

    Comparison of descriptor spaces for chemical compound retrieval and classification,

    N. Wale, I. A. Watson, and G. Karypis, “Comparison of descriptor spaces for chemical compound retrieval and classification,”Knowledge and Information Systems, vol. 14, pp. 347–375, 2008

    2008

  58. [58]

    Derivation and validation of toxicophores for mutagenicity prediction,

    J. Kazius, R. McGuire, and R. Bursi, “Derivation and validation of toxicophores for mutagenicity prediction,” Journal of medicinal chemistry, vol. 48, no. 1, pp. 312– 320, 2005

    2005

  59. [59]

    Graph invari- ant kernels,

    F. Orsini, P. Frasconi, and L. De Raedt, “Graph invari- ant kernels,” inProceedings of the International Joint Conference on Artificial Intelligence, 2015

    2015

  60. [60]

    Ogb-lsc: A large-scale challenge for machine learning on graphs,

    W. Hu, M. Fey, H. Ren, M. Nakata, Y . Dong, and J. Leskovec, “Ogb-lsc: A large-scale challenge for machine learning on graphs,”arXiv preprint arXiv:2103.09430, 2021

    work page arXiv 2021

  61. [61]

    Benchmarking graph neural networks,

    V . P. Dwivedi, C. K. Joshi, A. T. Luu, T. Laurent, Y . Bengio, and X. Bresson, “Benchmarking graph neural networks,”The Journal of Machine Learning Research., 2023. 15 APPENDIX This appendix provides detailed proofs, implementation details, and additional experimental protocols for S 2PLR. The notation follows Sections III-E–IV. The purpose of the append...

    2023

  62. [62]

    By the definition of the safe subspace in Eq

    Proof of Theorem 2:Letx∈ H τ,ρ and denote its pseudo-label byc= ˜y(x). By the definition of the safe subspace in Eq. (1), we have pS(c|x) =s(x)≥τ.(21) Assumption 1 then gives pT (c|x)≥p S(c|x)−β τ(α)≥τ−β τ(α).(22) Sinceη(x) = Pr[˜y(x)̸=y|x] = 1−p T (c|x), we obtain η(x)≤1−τ+β τ(α),(23) which proves Eq. (2). We next prove the aggregate neighborhood certifi...

  63. [63]

    The empirical structural gate is ˆrτ(x) = 1 knn knnX i=1 Bi,(31) and its population counterpart isr τ(x) =E[B i]

    Proof of Proposition 1:For a fixed target representationx, define the Bernoulli variable Bi =I[˜y(xi) = ˜y(x), s(xi)≥τ],(30) wherex i is thei-th sampled neighbor ofx. The empirical structural gate is ˆrτ(x) = 1 knn knnX i=1 Bi,(31) and its population counterpart isr τ(x) =E[B i]. Hoeffding’s inequality gives Pr [|ˆrτ(x)−r τ(x)| ≥ϵ]≤2 exp(−2k nnϵ2).(32) Se...

  64. [64]

    1 Ke KeX k=1 Zk ≥ 1 2 |x # ≤Pr

    Proof of Corollary 2:Fixx∈ H τ,ρ. LetZ k =I[f k S(x)̸=y]be the error indicator of thek-th source expert. Under the idealized condition in Corollary 2, the variablesZ 1, . . . , ZKe are conditionally independent givenx, andE[Z k |x]≤¯e <1/2. Majority voting is wrong only when 1 Ke KeX k=1 Zk ≥ 1 2 .(34) Hoeffding’s inequality gives Pr " 1 Ke KeX k=1 Zk ≥ 1...

  65. [65]

    Proof of Proposition 2:LetU=X T \ Hτ,ρ. The target risk can be decomposed as RT (h) = Pr[h(x)̸=y, x∈ H τ,ρ] + Pr[h(x)̸=y, x∈ U](37) =π H Pr[h(x)̸=y|x∈ H τ,ρ] +π U RU(h).(38) On the safe subspace, the eventh(x)̸=yis contained in the union of the two eventsh(x)̸= ˜y(x)and˜y(x)̸=y. Hence Pr[h(x)̸=y|x∈ H τ,ρ]≤Pr[h(x)̸= ˜y(x)|x∈ H τ,ρ] + Pr[˜y(x)̸=y|x∈ H τ,ρ](...

  66. [66]

    For completeness, we state a standard finite-sample form that explains the empirical objective in Eq

    A finite-sample variant:The previous proposition is a population decomposition. For completeness, we state a standard finite-sample form that explains the empirical objective in Eq. (20). Supposemselected target graphs are sampled fromH τ,ρ, and let bRH(h)be the empirical pseudo-label risk on them. For a bounded classification loss and a hypothesis class ...

  67. [67]

    For structure-based domain shifts, we use real-world graph benchmarks and image-derived graph benchmarks

    Dataset Description:We provide additional details of the datasets used in our experiments. For structure-based domain shifts, we use real-world graph benchmarks and image-derived graph benchmarks. For graph datasets, molecules are represented as graphs with atoms as nodes and chemical bonds as edges, while proteins are represented with amino acids as node...

  68. [68]

    Each graph is converted into a PyG data object, with the original node attributes or node labels used as input features when available

    Data Processing:For real-world graph datasets from TUDataset 4, including DD, PROTEINS, Mutagenicity, NCI1, FRANKENSTEIN, BZR, BZR MD, COX2, and COX2 MD, we follow the standard preprocessing pipeline provided by PyTorch Geometric5. Each graph is converted into a PyG data object, with the original node attributes or node labels used as input features when ...

  69. [69]

    More performance comparison:In this section, we present additional performance comparisons between S 2PLR and all baseline methods on more graph benchmarks, as summarized in Tables XVI–XXV. The results cover node-density and 21 Algorithm 1Training Procedure of S 2PLR 1:Input:Source expertsM, target dataD t, thresholdsζ, ρ min, variance percentileq u, neig...

  70. [70]

    Specifically, we evaluate five ablated variants of S2PLR, including S 2PLR w/o ME, S 2PLR w/o CF, S 2PLR w/o TS, S 2PLR w/o NC, and S 2PLR w/o SR

    More Ablation study:To further validate the effectiveness of each component in S 2PLR, we conduct additional ablation studies on the DD, NCI1, FRANKENSTEIN, and ogbg-molhiv datasets. Specifically, we evaluate five ablated variants of S2PLR, including S 2PLR w/o ME, S 2PLR w/o CF, S 2PLR w/o TS, S 2PLR w/o NC, and S 2PLR w/o SR. The corresponding experimen...

  71. [71]

    The results are shown in Fig

    More Sensitivity Analysis:We further analyze the sensitivity of S 2PLR to the confidence thresholdζand the neighborhood- consistency thresholdρ min on DD, FRANKENSTEIN, NCI1, and ogbg-molhiv. The results are shown in Fig. 5 and 6, and the observed trends are consistent with the analysis in Section V-G. For each dataset, we vary one threshold at a time ove...

  72. [72]

    The goal is to examine whether the performance gains of S 2PLR come from a substantially larger adaptation budget or from the proposed reliability-guided refinement mechanism

    Efficiency and Resource Consumption Analysis:In this section, we report the computational cost of S 2PLR and representative baselines, including GALA, GraphCTA, and SOGA, in Table XI. The goal is to examine whether the performance gains of S 2PLR come from a substantially larger adaptation budget or from the proposed reliability-guided refinement mechanis...