pith. machine review for the scientific record. sign in

arxiv: 2605.05463 · v1 · submitted 2026-05-06 · 💻 cs.LG · cs.AI

Recognition: unknown

Robustness of Graph Self-Supervised Learning to Real-World Noise: A Case Study on Text-Driven Biomedical Graphs

Othmane Kabal , Mounira Harzallah , Fabrice Guillet , Hideaki Takeda , Ryutaro Ichise
Authors on Pith no claims yet

Pith reviewed 2026-05-08 17:10 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords graph self-supervised learningrobustness to noisetext-driven graphsbiomedical knowledge graphsGNN architecturespretext tasksrelation reconstructionfeature reconstruction
0
0 comments X

The pith

Real-world noise from text extraction makes some graph self-supervised methods and architectures far more reliable than others on biomedical data.

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

The paper tests graph self-supervised learning on automatically built biomedical graphs that contain extraction errors from medical text. It compares these noisy graphs against clean reference graphs that share the same evaluation labels. Feature reconstruction tasks keep their accuracy close to clean-graph levels, while relation reconstruction drops sharply unless the graph schema is tightly defined. Contrastive approaches hold up better overall but only when they match the final task. Bidirectional relational message passing works better on noisy graphs; unidirectional designs do better on clean ones. The results give concrete selection rules for applying GSSL to real extracted graphs and show gains over plain language-model baselines.

Core claim

GSSL pretext tasks and GNN architectures show uneven robustness to real-world noise in text-driven biomedical graphs. Feature reconstruction remains largely unaffected and matches clean-graph performance, relation reconstruction is highly sensitive unless schemas are well-defined, and contrastive objectives depend on downstream alignment. Bidirectional relational message-passing architectures suit noisy graphs, while unidirectional ones suit clean graphs. The NATD-GSSL framework demonstrates these patterns through a dual-graph protocol on MedMentions and UMLS data and reports up to 7% gains over pretrained language model baselines.

What carries the argument

The dual-graph protocol that pairs a noisy graph built from MedMentions text with a clean UMLS reference graph, aligned by shared gold-standard labels, to isolate the isolated effect of text-extraction noise on GSSL.

If this is right

  • Feature reconstruction can be applied directly to noisy text-derived graphs without major accuracy loss.
  • Relation reconstruction needs well-defined schemas to avoid large drops from noise.
  • Bidirectional relational message-passing GNNs are the safer choice when graphs come from automatic text extraction.
  • Contrastive GSSL works best only when the pretext task is aligned with the eventual downstream objective.
  • NATD-GSSL yields measurable gains over pretrained language models on unsupervised term typing.

Where Pith is reading between the lines

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

  • The same robustness pattern may appear in other domains that rely on automatic knowledge-graph construction from text, suggesting a need for tests outside biomedicine.
  • Hybrid designs that combine feature reconstruction with selective relation reconstruction could reduce noise sensitivity further.
  • The preference for bidirectional architectures on noisy graphs may generalize to any setting where missing or spurious edges are common.

Load-bearing premise

The noisy graph from MedMentions and the clean UMLS graph differ only in text-extraction noise once they are aligned on the gold standard.

What would settle it

Re-running the same GSSL methods on a second pair of noisy and clean graphs built with a different text extractor or in a non-biomedical domain and finding that the robustness rankings reverse or disappear.

read the original abstract

Graph Self-Supervised Learning (GSSL) offers a powerful paradigm for learning graph representations without labeled data. However, existing work assumes clean, manually curated graphs. Recent advances in NLP enable the large-scale automatic extraction of knowledge graphs from text, opening new opportunities for GSSL while introducing substantial real-world noise. This type of noise remains largely unexplored, as prior robustness studies typically rely on synthetic perturbations. To address this gap, we present the first comprehensive evaluation of GSSL methods on text-driven graphs for unsupervised term typing. We introduce Noise-Aware Text-Driven Graph GSSL (NATD-GSSL), a unified framework that combines automatic graph construction, graph refinement, and GSSL. Our evaluation follows a dual-graph protocol that contrasts a noisy graph derived from MedMentions with a clean Unified Medical Language System (UMLS) reference graph, aligned through a shared gold standard. Our results reveal variability in robustness across both pretext tasks and Graph Neural Network (GNN) architectures. Relation reconstruction is highly sensitive to noise and benefits from well-defined schemas, whereas feature reconstruction is considerably more robust, achieving performance comparable to clean-graph settings. Contrastive objectives are generally less affected by noise but depend strongly on alignment with downstream tasks. GNN architecture also plays a critical role: bidirectional relational message-passing designs are better suited to noisy, text-driven graphs, while unidirectional relational ones perform best on clean graphs. Overall, NATD-GSSL provides practical guidance for applying GSSL to real-world, noisy graphs and achieves up to a 7\% improvement over pretrained language model baselines. All code and benchmarks are publicly available at https://github.com/OthmaneKabal/MC2GAE.

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 manuscript introduces NATD-GSSL, a unified framework combining automatic construction of text-driven biomedical graphs, refinement, and graph self-supervised learning (GSSL) for unsupervised term typing. It evaluates GSSL robustness to real-world noise via a dual-graph protocol that contrasts a noisy graph derived from MedMentions with a clean UMLS reference graph, aligned through a shared gold-standard term typing task. The evaluation covers multiple pretext tasks (relation reconstruction, feature reconstruction, contrastive learning) and GNN architectures (bidirectional vs. unidirectional relational message passing). Key findings are that feature reconstruction is robust to noise and matches clean-graph performance, relation reconstruction is highly sensitive but benefits from well-defined schemas, contrastive objectives are less noise-affected but require downstream-task alignment, and bidirectional relational GNNs suit noisy graphs while unidirectional ones excel on clean graphs. The work reports up to 7% improvement over pretrained language model baselines and releases all code and benchmarks.

Significance. If the empirical comparisons hold after addressing protocol details, the work is significant for filling a gap between synthetic-noise robustness studies and real-world text-extraction noise in biomedical KGs. It supplies concrete, architecture- and pretext-specific guidance for deploying GSSL on automatically extracted graphs, which is practically relevant given the prevalence of noisy text-derived KGs. Public release of code and benchmarks is a clear strength that enables direct reproduction and extension.

major comments (2)
  1. [Abstract] Abstract (dual-graph protocol): The central attribution of robustness differences (feature vs. relation reconstruction, architecture effects) and the 7% gain to real-world text-extraction noise assumes the protocol isolates noise. No evidence is supplied that gold-standard alignment equalizes node/edge coverage, relation schemas, degree distributions, or subgraph connectivity between the MedMentions-derived and UMLS graphs; unaccounted structural differences could independently drive the observed GSSL performance gaps.
  2. [Results] Results (7% improvement): The claim of 'up to a 7% improvement over pretrained language model baselines' is stated without identifying the exact metric, the specific NATD-GSSL configuration that achieves it, standard deviations across runs, or statistical significance tests. This omission makes it impossible to judge whether the gain is reliable or practically meaningful.
minor comments (2)
  1. [Methods] The NATD-GSSL framework is presented as combining graph construction, refinement, and GSSL, but the manuscript would benefit from a concise algorithmic outline or pseudocode for the refinement step to clarify how noise is mitigated before pretext training.
  2. [Figures/Tables] Ensure all tables or figures comparing noisy vs. clean settings explicitly label the two graphs and report the number of nodes/edges after alignment so readers can verify comparability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. The comments highlight important aspects of protocol validation and result reporting that we will address to strengthen the work. We respond to each major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract (dual-graph protocol): The central attribution of robustness differences (feature vs. relation reconstruction, architecture effects) and the 7% gain to real-world text-extraction noise assumes the protocol isolates noise. No evidence is supplied that gold-standard alignment equalizes node/edge coverage, relation schemas, degree distributions, or subgraph connectivity between the MedMentions-derived and UMLS graphs; unaccounted structural differences could independently drive the observed GSSL performance gaps.

    Authors: We agree that the manuscript would benefit from explicit evidence that the dual-graph protocol primarily isolates the effect of text-extraction noise. While the shared gold-standard term typing task ensures identical evaluation targets, we did not include a direct comparison of structural properties such as node/edge coverage, relation schemas, degree distributions, or subgraph connectivity. In the revised version, we will add a dedicated subsection (in the Experimental Setup) with tables and statistics comparing these properties between the MedMentions-derived graph and the UMLS reference graph. This addition will allow readers to evaluate the degree of structural alignment and support the attribution of performance differences to noise levels. revision: yes

  2. Referee: [Results] Results (7% improvement): The claim of 'up to a 7% improvement over pretrained language model baselines' is stated without identifying the exact metric, the specific NATD-GSSL configuration that achieves it, standard deviations across runs, or statistical significance tests. This omission makes it impossible to judge whether the gain is reliable or practically meaningful.

    Authors: The 7% figure refers to the improvement in micro-F1 score on the term typing task for the NATD-GSSL configuration using feature reconstruction with bidirectional relational message passing, relative to the strongest language-model baseline. These numbers appear in the results section (Table 2 and associated figures). We acknowledge that standard deviations across runs and statistical significance tests were not reported. We will revise the results section and figures to include standard deviations (computed over 5 random seeds), error bars, and paired t-test p-values for the key improvements. This will make the reported gains more transparent and allow proper assessment of reliability. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical comparison with no derivations or self-referential predictions

full rationale

The paper presents an empirical evaluation of GSSL methods on noisy text-derived graphs versus clean reference graphs using a dual-graph protocol. No mathematical derivations, first-principles results, fitted parameters renamed as predictions, or self-citation chains are described that would reduce claims to inputs by construction. The central claims concern observed robustness differences across pretext tasks and architectures, supported by experimental results rather than tautological definitions or imported uniqueness theorems. This is a standard experimental study design with no load-bearing circular steps.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

The abstract relies on standard GSSL assumptions and the premise that text-derived graphs contain substantial real-world noise; no explicit free parameters or new physical entities are introduced.

axioms (2)
  • domain assumption Graph self-supervised learning can produce useful representations without labeled data
    Core premise stated in the opening sentence of the abstract
  • domain assumption Automatically extracted text-driven graphs contain substantial real-world noise compared with manually curated graphs
    Explicitly contrasted in the abstract as the motivation for the study
invented entities (1)
  • NATD-GSSL framework no independent evidence
    purpose: Unified pipeline that combines automatic graph construction from text, graph refinement, and GSSL training
    Newly named and described in the abstract as the authors' contribution

pith-pipeline@v0.9.0 · 5630 in / 1664 out tokens · 26480 ms · 2026-05-08T17:10:33.967065+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

56 extracted references · 16 canonical work pages · 2 internal anchors

  1. [1]

    IEEE transactions on neural networks and learning systems32(1), 4–24 (2020)

    Wu, Z., Pan, S., Chen, F., Long, G., Zhang, C., Yu, P.S.: A comprehensive survey on graph neural networks. IEEE transactions on neural networks and learning systems32(1), 4–24 (2020)

    2020

  2. [2]

    Machine Intelligence Research21(6), 1011–1061 (2024)

    Dai, E., Zhao, T., Zhu, H., Xu, J., Guo, Z., Liu, H., Tang, J., Wang, S.: A comprehensive survey on trustworthy graph neural networks: Privacy, robust- ness, fairness, and explainability. Machine Intelligence Research21(6), 1011–1061 (2024)

    2024

  3. [3]

    In: Proceedings of the 26th International Joint Conference on Artificial Intelligence

    Xu, J., Qiu, X., Chen, K., Huang, X.: Knowledge graph representation with jointly structural and textual encoding. In: Proceedings of the 26th International Joint Conference on Artificial Intelligence. IJCAI’17, pp. 1318–1324. AAAI Press, ??? (2017)

    2017

  4. [4]

    Artificial intelligence review57(8), 190 (2024) 22

    Wang, K., Ding, Y., Han, S.C.: Graph neural networks for text classification: A survey. Artificial intelligence review57(8), 190 (2024) 22

    2024

  5. [5]

    IEEE Transactions on Knowledge and Data Engineering36(7), 3580–3599 (2024)

    Zhang, Z., Cui, P., Zhu, W.: Deep learning on graphs: A survey. IEEE Trans. on Knowl. and Data Eng.34(1), 249–270 (2022) https://doi.org/10.1109/TKDE. 2020.2981333

    work page doi:10.1109/tkde 2022

  6. [6]

    Machine Vision and Applications33(1), 4 (2022)

    Xiao, S., Wang, S., Dai, Y., Guo, W.: Graph neural networks in node classification: survey and evaluation. Machine Vision and Applications33(1), 4 (2022)

    2022

  7. [7]

    IEEE transactions on knowledge and data engineering35(6), 5879–5900 (2022)

    Liu, Y., Jin, M., Pan, S., Zhou, C., Zheng, Y., Xia, F., Philip, S.Y.: Graph self-supervised learning: A survey. IEEE transactions on knowledge and data engineering35(6), 5879–5900 (2022)

    2022

  8. [8]

    arXiv preprint arXiv:2211.12875 , year=

    Liu, Y., Xia, J., Zhou, S., Yang, X., Liang, K., Fan, C., Zhuang, Y., Li, S.Z., Liu, X., He, K.: A survey of deep graph clustering: Taxonomy, challenge, application, and open resource. arXiv preprint arXiv:2211.12875 (2022)

    work page arXiv 2022

  9. [9]

    Kipf et al.Variational Graph Auto- Encoders

    Kipf, T.N., Welling, M.: Variational graph auto-encoders. arXiv preprint arXiv:1611.07308 (2016)

    work page arXiv 2016

  10. [10]

    In: Proceedings of the Sixteenth ACM International Conference on Web Search and Data Mining, pp

    Tan, Q., Liu, N., Huang, X., Choi, S.-H., Li, L., Chen, R., Hu, X.: S2gae: Self- supervised graph autoencoders are generalizable learners with graph masking. In: Proceedings of the Sixteenth ACM International Conference on Web Search and Data Mining, pp. 787–795 (2023)

    2023

  11. [11]

    In: Proceedings of the 34th International Conference on Neural Information Processing Systems

    You, Y., Chen, T., Sui, Y., Chen, T., Wang, Z., Shen, Y.: Graph contrastive learning with augmentations. In: Proceedings of the 34th International Conference on Neural Information Processing Systems. NIPS ’20. Curran Associates Inc., Red Hook, NY, USA (2020)

    2020

  12. [12]

    Graph Attention Networks

    Veliˇ ckovi´ c, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., Bengio, Y.: Graph attention networks. arXiv preprint arXiv:1710.10903 (2017)

    work page internal anchor Pith review arXiv 2017

  13. [13]

    https://arxiv.org/abs/ 1412.6575

    Yang, B., Yih, W.-t., He, X., Gao, J., Deng, L.: Embedding Entities and Relations for Learning and Inference in Knowledge Bases (2015). https://arxiv.org/abs/ 1412.6575

    work page arXiv 2015

  14. [14]

    In: Gangemi, A., Navigli, R., Vidal, M.-E., Hitzler, P., Troncy, R., Hollink, L., Tordai, A., Alam, M

    Schlichtkrull, M., Kipf, T.N., Bloem, P., Berg, R., Titov, I., Welling, M.: Modeling relational data with graph convolutional networks. In: Gangemi, A., Navigli, R., Vidal, M.-E., Hitzler, P., Troncy, R., Hollink, L., Tordai, A., Alam, M. (eds.) The Semantic Web, pp. 593–607. Springer, Cham (2018)

    2018

  15. [15]

    Vrandeˇ ci´ c, D., Kr¨ otzsch, M.: Wikidata: a free collaborative knowledgebase. Commun. ACM57(10), 78–85 (2014) https://doi.org/10.1145/2629489

    work page doi:10.1145/2629489 2014

  16. [16]

    In: Proceedings of the 3rd Workshop on Continuous Vector Space Models and Their Compositionality, pp

    Toutanova, K., Chen, D.: Observed versus latent features for knowledge base and text inference. In: Proceedings of the 3rd Workshop on Continuous Vector Space Models and Their Compositionality, pp. 57–66 (2015) 23

    2015

  17. [17]

    AI magazine29(3), 93–93 (2008)

    Sen, P., Namata, G., Bilgic, M., Getoor, L., Galligher, B., Eliassi-Rad, T.: Collective classification in network data. AI magazine29(3), 93–93 (2008)

    2008

  18. [18]

    ACM Computing Surveys56(4), 1–62 (2023)

    Zhong, L., Wu, J., Li, Q., Peng, H., Wu, X.: A comprehensive survey on automatic knowledge graph construction. ACM Computing Surveys56(4), 1–62 (2023)

    2023

  19. [19]

    Procedia Computer Science246, 2617–2626 (2024)

    Kabal, O., Harzallah, M., Guillet, F., Ichise, R.: Enhancing domain-independent knowledge graph construction through openie cleaning and llms validation. Procedia Computer Science246, 2617–2626 (2024)

    2024

  20. [20]

    Applied Network Science10(1), 64 (2025)

    Cai, E., O’Connor, B.: Understanding the effect of knowledge graph extraction error on downstream graph analyses: a case study on affiliation graphs. Applied Network Science10(1), 64 (2025)

    2025

  21. [21]

    arXiv preprint arXiv:2403.04468 , year=

    Ju, W., Yi, S., Wang, Y., Xiao, Z., Mao, Z., Li, H., Gu, Y., Qin, Y., Yin, N., Wang, S., Liu, X., Luo, X., Yu, P.S., Zhang, M.: A survey of graph neural networks in real world: Imbalance, noise, privacy and ood challenges. ArXivabs/2403.04468 (2024)

    work page arXiv 2024

  22. [22]

    In: Proceedings of the AAAI Conference on Artificial Intelligence, vol

    Zhuang, J., Al Hasan, M.: Defending graph convolutional networks against dynamic graph perturbations via bayesian self-supervision. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 36, pp. 4405–4413 (2022)

    2022

  23. [23]

    In: Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pp

    Hegde, M., Talukdar, P.: An entity-centric approach for overcoming knowledge graph sparsity. In: Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pp. 530–535 (2015)

    2015

  24. [24]

    Kggen: Extracting knowledge graphs from plain text with language models.arXiv preprint arXiv:2502.09956, 2025

    Mo, B., Yu, K., Kazdan, J., Cabezas, J., Mpala, P., Yu, L., Cundy, C., Kanat- soulis, C., Koyejo, S.: Kggen: Extracting knowledge graphs from plain text with language models. arXiv preprint arXiv:2502.09956 (2025)

    work page arXiv 2025

  25. [25]

    MedMentions: A Large Biomedical Corpus Annotated with UMLS Concepts

    Mohan, S., Li, D.: Medmentions: A large biomedical corpus annotated with umls concepts. arXiv preprint arXiv:1902.09476 (2019)

    work page Pith review arXiv 1902

  26. [26]

    Nucleic acids research32(suppl 1), 267–270 (2004)

    Bodenreider, O.: The unified medical language system (umls): integrating biomed- ical terminology. Nucleic acids research32(suppl 1), 267–270 (2004)

    2004

  27. [27]

    Graph neural networks: A review of methods and applications,

    Zhou, J., Cui, G., Hu, S., Zhang, Z., Yang, C., Liu, Z., Wang, L., Li, C., Sun, M.: Graph neural networks: A review of methods and applications. AI Open1, 57–81 (2020) https://doi.org/10.1016/j.aiopen.2021.01.001

    work page doi:10.1016/j.aiopen.2021.01.001 2020

  28. [28]

    In: Graph Neural Networks: Foundations, Frontiers, and Applications, pp

    Wang, Y., Jin, W., Derr, T.: Graph neural networks: Self-supervised learning. In: Graph Neural Networks: Foundations, Frontiers, and Applications, pp. 391–420. Springer, ??? (2022)

    2022

  29. [29]

    Semi-Supervised Classification with Graph Convolutional Networks

    Kipf, T.N., Welling, M.: Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907 (2016) 24

    work page internal anchor Pith review arXiv 2016

  30. [30]

    In: Proceedings of the 10th International Conference on Knowledge Capture, pp

    Cai, L., Yan, B., Mai, G., Janowicz, K., Zhu, R.: Transgcn: Coupling transfor- mation assumptions with graph convolutional networks for link prediction. In: Proceedings of the 10th International Conference on Knowledge Capture, pp. 131–138 (2019)

    2019

  31. [31]

    IEEE transactions on cybernetics50(6), 2475–2487 (2019)

    Pan, S., Hu, R., Fung, S.-f., Long, G., Jiang, J., Zhang, C.: Learning graph embed- ding with adversarial training methods. IEEE transactions on cybernetics50(6), 2475–2487 (2019)

    2019

  32. [32]

    In: Proceedings of the 2017 ACM on Conference on Information and Knowledge Management, pp

    Wang, C., Pan, S., Long, G., Zhu, X., Jiang, J.: Mgae: Marginalized graph autoen- coder for graph clustering. In: Proceedings of the 2017 ACM on Conference on Information and Knowledge Management, pp. 889–898 (2017)

    2017

  33. [33]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp

    Park, J., Lee, M., Chang, H.J., Lee, K., Choi, J.Y.: Symmetric graph convolutional autoencoder for unsupervised graph representation learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 6519–6528 (2019)

    2019

  34. [34]

    Information Sciences632, 439–453 (2023)

    Li, J., Lu, G., Wu, Z., Ling, F.: Multi-view representation model based on graph autoencoder. Information Sciences632, 439–453 (2023)

    2023

  35. [35]

    Knowledge-Based Systems234, 107564 (2021)

    Sun, D., Li, D., Ding, Z., Zhang, X., Tang, J.: Dual-decoder graph autoencoder for unsupervised graph representation learning. Knowledge-Based Systems234, 107564 (2021)

    2021

  36. [36]

    https://arxiv.org/abs/2205

    Hou, Z., Liu, X., Cen, Y., Dong, Y., Yang, H., Wang, C., Tang, J.: GraphMAE: Self-Supervised Masked Graph Autoencoders (2022). https://arxiv.org/abs/2205. 10803

    2022

  37. [37]

    https://arxiv.org/abs/1911.08538

    Ren, Y., Liu, B., Huang, C., Dai, P., Bo, L., Zhang, J.: Heterogeneous Deep Graph Infomax (2020). https://arxiv.org/abs/1911.08538

    work page arXiv 2020

  38. [38]

    In2024 IEEE 40th International Conference on Data Engineering (ICDE)

    Wang, Y., Yan, X., Hu, C., Xu, Q., Yang, C., Fu, F., Zhang, W., Wang, H., Du, B., Jiang, J.: Generative and contrastive paradigms are complementary for graph self-supervised learning. In: 2024 IEEE 40th International Confer- ence on Data Engineering (ICDE), pp. 3364–3378. IEEE Computer Society, Los Alamitos, CA, USA (2024). https://doi.org/10.1109/ICDE601...

    work page doi:10.1109/icde60146.2024.00234 2024

  39. [39]

    Deep graph contrastive representation learning.arXiv:2006.04131,

    Zhu, Y., Xu, Y., Yu, F., Liu, Q., Wu, S., Wang, L.: Deep Graph Contrastive Representation Learning (2020). https://arxiv.org/abs/2006.04131

    work page arXiv 2020

  40. [40]

    (2019).https://openreview.net/forum?id=rklz9iAcKQ

    Veliˇ ckovi´ c, P., Fedus, W., Hamilton, W.L., Li` o, P., Bengio, Y., Hjelm, R.D.: Deep graph infomax. (2019).https://openreview.net/forum?id=rklz9iAcKQ

    2019

  41. [41]

    In: Proceedings of the AAAI Conference on Artificial Intelligence, vol

    Chen, J., Kou, G.: Attribute and structure preserving graph contrastive learning. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 37, pp. 25 7024–7032 (2023)

    2023

  42. [42]

    Semantic web8(3), 489–508 (2016)

    Paulheim, H.: Knowledge graph refinement: A survey of approaches and evalua- tion methods. Semantic web8(3), 489–508 (2016)

    2016

  43. [43]

    arXiv preprint arXiv:1907.10903 , year=

    Rong, Y., Huang, W., Xu, T., Huang, J.: Dropedge: Towards deep graph convolutional networks on node classification. arXiv preprint arXiv:1907.10903 (2019)

    work page arXiv 1907

  44. [44]

    Applied Intelligence 53(21), 25154–25170 (2023)

    Yuan, J., Yu, H., Cao, M., Song, J., Xie, J., Wang, C.: Self-supervised robust graph neural networks against noisy graphs and noisy labels. Applied Intelligence 53(21), 25154–25170 (2023)

    2023

  45. [45]

    Fox, J., Rajamanickam, S.: How robust are graph neural networks to structural noise? arXiv preprint arXiv:1912.10206 (2019)

    work page arXiv 1912

  46. [46]

    In: Proceedings of the AAAI Conference on Artificial Intelligence, vol

    Wang, Y., Wang, Y., Zhang, Z., Yang, S., Zhao, K., Liu, J.: User: Unsupervised structural entropy-based robust graph neural network. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 37, pp. 10235–10243 (2023)

    2023

  47. [47]

    In: Proceedings of the AAAI Conference on Artificial Intelligence, vol

    Ennadir, S., Abbahaddou, Y., Lutzeyer, J.F., Vazirgiannis, M., Bostr¨ om, H.: A simple and yet fairly effective defense for graph neural networks. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, pp. 21063–21071 (2024)

    2024

  48. [48]

    In: International Conference on Database Systems for Advanced Applications, pp

    Guo, J., Li, S., Zhao, Y., Zhang, Y.: Learning robust representation through graph adversarial contrastive learning. In: International Conference on Database Systems for Advanced Applications, pp. 682–697 (2022). Springer

    2022

  49. [49]

    In: Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pp

    Jin, W., Ma, Y., Liu, X., Tang, X., Wang, S., Tang, J.: Graph structure learning for robust graph neural networks. In: Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pp. 66–74 (2020)

    2020

  50. [50]

    In: IJCAI, pp

    Wang, Z., Ren, Z., He, C., Zhang, P., Hu, Y.: Robust embedding with multi-level structures for link prediction. In: IJCAI, pp. 5240–5246 (2019)

    2019

  51. [51]

    In: Gesese, G.A., Sack, H., Paul- heim, H., Merono-Penuela, A., Chen, L

    Dong, N., Kertkeidkachorn, N., Liu, X., Shirai, K.: Refining noisy knowl- edge graph with large language models. In: Gesese, G.A., Sack, H., Paul- heim, H., Merono-Penuela, A., Chen, L. (eds.) Proceedings of the Workshop on Generative AI and Knowledge Graphs (GenAIK), pp. 78–86. Interna- tional Committee on Computational Linguistics, Abu Dhabi, UAE (2025)...

    2025

  52. [52]

    In: International Semantic Web Conference, pp

    Auer, S., Bizer, C., Kobilarov, G., Lehmann, J., Cyganiak, R., Ives, Z.: Dbpedia: A nucleus for a web of open data. In: International Semantic Web Conference, pp. 722–735 (2007). Springer 26

    2007

  53. [53]

    In: Proceedings of the 2008 ACM SIGMOD International Conference on Management of Data, pp

    Bollacker, K., Evans, C., Paritosh, P., Sturge, T., Taylor, J.: Freebase: a collabora- tively created graph database for structuring human knowledge. In: Proceedings of the 2008 ACM SIGMOD International Conference on Management of Data, pp. 1247–1250 (2008)

    2008

  54. [54]

    In: ISWC (Posters, Demos & Industry Tracks) (2017)

    Mihindukulasooriya, N., Hassanzadeh, O., Dash, S., Gliozzo, A.: Towards com- prehensive noise detection in automatically created knowledge graphs. In: ISWC (Posters, Demos & Industry Tracks) (2017)

    2017

  55. [55]

    Journal of Artificial Intelligence Research78, 37–68 (2023)

    Faralli, S., Lenzi, A., Velardi, P.: A benchmark study on knowledge graphs enrich- ment and pruning methods in the presence of noisy relationships. Journal of Artificial Intelligence Research78, 37–68 (2023)

    2023

  56. [56]

    Hamilton, W., Ying, Z., Leskovec, J.: Inductive representation learning on large graphs. Advances in neural information processing systems30(2017) A PLM Initial Embeddings This appendix reports the evaluation of initial embeddings obtained from different categories of pre-trained language models on typing task, in order to identify the most effective repr...

    2017