Pith. sign in

REVIEW 4 major objections 5 minor 68 references

A unified 3×3 taxonomy of text, structure, and label failures yields nine stress-test scenarios for text-attributed graph learning, with label imbalance the harshest.

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

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

T0 review

2026-08-01 13:24 UTC pith:TWERMNHH

load-bearing objection A broad and well-organized TAG robustness benchmark that is currently not instantiable from the paper—no code/data, unreported perturbation strengths, and a persistence claim contradicted by its own Figure 2. the 4 major comments →

arxiv 2607.19108 v1 pith:TWERMNHH submitted 2026-07-21 cs.AI

OpenRTAG: A Comprehensive Benchmark for Robust Text-Attributed Graph Learning under Data Quality Degradation

Yuze Dai , Zhihan Zhang , Yan Zhao , Ruoyu Wu , Xunkai Li , Zekai Chen , Qiangqiang Dai , Hongchao Qin
show 1 more author
Ronghua Li
This is my paper
classification cs.AI
keywords text-attributed graphsgraph neural networksrobustness benchmarkdata quality degradationsparsity noise imbalancelabel imbalancegraph foundation modelscomposite degradation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to establish that the many ways real text-attributed graphs (TAGs) can be imperfect — missing or garbled text, broken or spurious edges, scarce, wrong, or skewed labels — can be organized into one 3×3 grid: three information sources (text, structure, labels) crossed with three failure modes (sparsity, noise, imbalance). It then argues that these nine scenarios are valid, non-collapsing stress tests: a standard graph convolutional model degrades under all nine, the damage persists across traditional GNNs, LLM-GNNs, and a graph foundation model, and label imbalance hurts most. The paper also studies whether existing repair methods help under the scenario they are designed for, and finds the gains are inconsistent across datasets and often disappear under composite degradation. If right, OpenRTAG gives the community a shared testbed so robustness results can be compared across datasets, tasks, and model families instead of the present fragmented evidence.

Core claim

On its own terms, OpenRTAG's central discovery is that a deliberately simple taxonomy — three modalities (text, structure, labels) times three degradation types (sparsity, noise, imbalance) — already generates nine scenarios that meaningfully separate the robustness of TAG models. The anchor model loses accuracy in all nine scenarios, with label imbalance the most punishing (about 21.8 percentage points average drop), and the pattern repeats across model families, though with notable exceptions. The paper further finds that text-, structure-, and label-oriented repair methods are highly scenario- and dataset-dependent: under text noise and label noise the untouched base model often remains c

What carries the argument

The load-bearing object is the 3×3 degradation taxonomy and its scenario constructors A_{s,α}(G), which map a clean text-attributed graph to a degraded one by applying sparsity, noise, or imbalance to text, structure, or labels with a controllable strength α. The constructors are intentionally compact — they state the target modality and direction of damage while leaving exact sampling rules to the benchmark constructor — and the clean/degraded paired evaluation is what turns the taxonomy into measurable robustness evidence.

Load-bearing premise

The benchmark's validity rests on treating a single anchor model's consistent degradation under all nine scenarios as evidence that the scenarios are broadly harmful stress tests; if that one-model evidence does not generalize to other backbones, the 'valid scenario' claim weakens, and the paper's own results already show one foundation model improving under two scenarios.

What would settle it

Run the nine scenarios on a broader panel of graph foundation models and count how many scenarios yield accuracy improvements instead of drops; if improvements are common rather than rare, the claim that the scenarios are broadly harmful stress tests would be contradicted.

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

If this is right

  • The nine degradation scenarios give researchers a shared grid for reporting TAG robustness, making results comparable across datasets and tasks.
  • Label imbalance is the strongest stress case among the nine, with an average accuracy drop of about 21.8 percentage points on the anchor model.
  • Existing repair methods are scenario- and dataset-dependent: under text noise and label noise, the unmodified base model often stays competitive, while matched repair methods can even hurt.
  • Composite degradation (two scenarios combined) reduces or reverses the gains of single-scenario repair methods, suggesting one-defect-at-a-time fixes are insufficient.
  • The benchmark framework extends the same taxonomy to additional TAG datasets beyond the nine used in the main evaluation.

Where Pith is reading between the lines

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

  • The 'imbalance' axis bundles very different statistical phenomena (long-tailed labels, degree imbalance, uneven text length), so a single scenario may mix several failure modes; a follow-up benchmark could split each imbalance cell into more specific interventions.
  • The paper's own figure shows negative accuracy drops (improvements) for a representative graph foundation model on some datasets under structure noise and label noise; that suggests the scenarios may be stress tests for some backbones but not others, and 'validity' should be reported per model family.
  • A testable extension would use the same 3×3 grid but perturb at multiple strengths and fit a response curve, then compare repair methods by the area under that curve rather than a single ratio.
  • Since label imbalance is the strongest stressor, methods that reweight or resample supervision could be prioritized in low-quality TAG settings.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. OpenRTAG is a proposed robustness benchmark for text-attributed graph learning. The paper defines a 3×3 degradation taxonomy (modality: text/structure/label; type: sparsity/noise/imbalance), constructs degraded TAGs through per-scenario constructors with perturbation strength α, and reports experiments across nine datasets and three tasks (node classification, clustering, link prediction). The evaluation covers GNN, LLM-GNN, and graph-foundation-model backbones, scenario-matched repair baselines for text/structure/label degradation, and two-factor composite degradations. The core claims are (i) the nine scenarios are valid, non-collapsing stress tests and (ii) degradation effects persist across model families, motivating OpenRTAG as a standardized testbed.

Significance. If fully specified and released, OpenRTAG would be a useful community resource: the 3×3 taxonomy is intuitive, the model and baseline coverage is broad (three GNN families, seven LLM-related backbones, OpenGraph, plus repair baselines), and the benchmark covers multiple tasks with runtime and OOM records and composite degradation. The paper explicitly separates scenario semantics from implementation, which is a sound design principle. However, the submitted text does not instantiate the benchmark: the scenario constructors are not defined algorithmically, no code, data, or parameter values are provided, and a key Q1 claim is internally contradicted by negative drops in Fig. 2. The contribution is therefore not presently usable or verifiable.

major comments (4)
  1. [Section 2 (Eqs. (2)–(5)) and Benchmark Design] The degraded TAGs are not defined to a reproducible degree. Eq. (3) for T-Imb says variance 'is large,' Eq. (4) requires 'sim(i,j) low' with no threshold, and the text explicitly delegates 'exact sampling rules and compatibility constraints' to an unreleased 'benchmark constructor'. Tables 2–4 report no α values for Eq. (2), and Fig. 10's composite scenarios are not constructed. Since a benchmark's central asset is the concrete instantiations, Q1–Q5 are not independently verifiable. Releasing the constructor and all generated data/parameter files is necessary.
  2. [Fig. 2 / Q1 Results] Negative drops contradict the claim of persistent cross-backbone degradation. OpenGraph on History improves by 12.6 pp under Struct-Noise and 3.5 pp under Label-Noise; GraphText on WikiCS improves by 1.1 pp under Text-Imb. A 12.6 pp improvement under an intended stress test is not consistent with 'valid, meaningful, non-collapsing' or with 'degradation effects persist across ... GFM-style representatives.' Please report per-backbone/per-scenario harm rates or qualify the claim to the GCN anchor.
  3. [Q1 Setup] The anchor validation is asserted for GCN but Fig. 2 shows GCN only for the Cora and Photo panels; no table lists GCN's clean/degraded scores for all nine datasets. Without such evidence, the claim that 'the GCN anchor degrades under all nine scenarios' is not verifiable. Add a full per-dataset/per-scenario table (with α and seeds) or an aggregate statistic.
  4. [Q5 Composite Scenarios] Composite-degradation results depend on an undefined construction procedure. Fig. 10 labels combinations such as T-Spa+S-Spa and reports drops for Base and matched baselines, but the paper does not state how two scenario constructors are composed (sequential, simultaneous, or conflict resolution) nor the α levels. Without this, Q5 is not reproducible.
minor comments (5)
  1. [Tables 2–4] These tables lack a clean-data BaseModel column. Adding clean accuracy would contextualize repair gains and make the total degradation visible across datasets.
  2. [Figure 2] Color scales are normalized per dataset. Add a color legend or use a shared scale so cross-dataset and cross-model drop magnitudes can be compared directly.
  3. [References] The IGL-Bench entry appears twice (2024 and 2025 versions). Consolidate or distinguish them explicitly.
  4. [Eq. (3)] The notation W(t'_i) and φ(t'_i) is used without definition. Define the token set and semantic-content function, even informally.
  5. [Figure 2 headings] Typos/formatting issues such as 'GA T' should be fixed; use consistent model names (GraphSAGE, etc.).

Circularity Check

1 steps flagged

Scenario-validity confirmation is partly self-definitional: the 'intended directions' are written into the scenario definitions, and the validation then reports those directions as evidence; the remaining GCN-drop evidence lacks reported perturbation strengths and constructor code.

specific steps
  1. self definitional [Section 2 (Nine Quality-Degradation Scenarios, Eqs. 3–5) and Experiments Q1 (Scenario Validity results)]
    "These definitions are intentionally compact: they specify the target modality and the expected direction of degradation while leaving implementation details, such as exact sampling rules and compatibility constraints, to the benchmark constructor. ... The results (Fig. 2) confirm that the constructed scenarios are valid stress tests. The internal quality statistics move in the intended directions, and the GCN anchor degrades under all nine scenarios, showing that the perturbations are both effective and non-collapsing."

    The 'intended directions' are the constraints in Eqs. (3)–(5): T-Spa requires t'_i = ∅ or ℓ'_i ≪ ℓ̄; S-Spa requires |E^-|/|E| ≈ α; L-Noi fixes Pr(y_i=c|y*_i=c') = η_{cc'}. Because G_{s,α}=A_{s,α}(G) is defined to satisfy these inequalities, reporting that 'internal quality statistics move in the intended directions' only checks the constructor against its own specification. It is not independent evidence that the scenarios are valid stress tests. The only non-tautological evidence is the GCN accuracy drop, but the perturbation strength α and the constructor are not released, so that drop cannot be reproduced or shown to be non-collapsing across the nine scenarios. Thus the scenario-validity confirmation is partly by construction.

full rationale

OpenRTAG is an empirical benchmark rather than a formal derivation, so most of its content is not circular in the derivation sense. The dataset selection, task coverage, baseline comparisons, and the 3×3 taxonomy organization are independent contributions. The main circularity is localized to Q1's scenario-validity argument: the paper defines the nine scenarios via hard constraints on quality statistics and then cites the movement of those same statistics as confirmation that the scenarios are valid. That step is self-definitional. The GCN anchor degradation is a genuine empirical observation, but it is under-specified because α values in Eq. (2) and the benchmark constructor are deferred to an unavailable artifact, and Fig. 2 itself contains negative drops (e.g., OpenGraph on History under structure noise) that contradict the 'broadly harmful' claim. Self-citations such as UltraTAG from the authors' prior work appear only as baselines and are not load-bearing for the benchmark's validity, so they do not add circularity. Overall, the central benchmark asset is partially circular in its validation, but the surrounding evaluation framework retains independent content, giving a moderate score of 5.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The benchmark itself is a constructed artifact, not a scientific entity in the sense of a new particle or force. The ledger therefore records the design parameters (perturbation strengths, seeds, GCN hyperparameters) and the domain assumptions the empirical conclusions rest on. The most consequential is the use of a GCN anchor as the sole validity check.

free parameters (4)
  • perturbation strength α for each scenario = unreported
    The paper describes constructors parameterized by ratio/strength/seed but does not report the specific α values used for the main tables (Tables 2-4). Figures show ratio=0.2 and 0.8, but the tables do not state which was used.
  • label noise rate η = unreported
    Eq. (5) defines label noise as Pr(y_i=c | y*_i=c')=η_c'c, but the noise rate used in Q1/Q4 experiments is not specified in the text or tables.
  • random seed for scenario generators = unreported
    The benchmark constructors take a random seed parameter, but no seeds are reported, so the exact degraded graphs cannot be reconstructed.
  • GCN hyperparameters = unreported
    Hidden size, number of layers, learning rate, weight decay, and training epochs for the shared GCN backbone are not reported anywhere in the manuscript.
axioms (5)
  • domain assumption The 3x3 taxonomy (modality × degradation type) spans the meaningful quality issues in real-world TAGs.
    The paper asserts this without evidence that other dimensions (e.g., temporal decay, adversarial manipulation) are less important. Invoked in Section 'Nine Quality-Degradation Scenarios'.
  • domain assumption The nine selected datasets (Cora, CiteSeer, Instagram, WikiCS, PubMed, Children, Photo, History, Arxiv) are representative of the TAG ecosystem.
    Stated in 'Benchmark Data and Scenario Space'; the selection is a convenience sample and may not cover all TAG application domains.
  • domain assumption GCN is a fair shared backbone for evaluating all repair methods, including LLM-based text repair.
    Used throughout Q2-Q4; an LLM-generated text may not couple well with a GCN that uses a fixed feature extractor, potentially disadvantaging text-repair methods.
  • domain assumption The scenario-matched baseline subsets are representative of each repair family.
    In Q3/Q4 the paper selects subsets of structure- and label-repair methods; the selection criteria are not fully specified, so the comparison may be unrepresentative.
  • domain assumption GCN's degradation under a scenario is a sufficient validity check for that scenario across all model families.
    Q1 relies on a GCN anchor to validate all nine scenarios, but Fig. 2 shows some models improve under degradation, so the anchor's behavior may not generalize.

pith-pipeline@v1.3.0-alltime-deepseek · 14527 in / 14923 out tokens · 128392 ms · 2026-08-01T13:24:57.054870+00:00 · methodology

0 comments
read the original abstract

Text-attributed graphs (TAGs) are an important graph data form that combine relational structure with rich node text. However, real-world TAGs are often imperfect, with quality issues arising from text, structure, and labels, and typically manifesting as sparsity, noise, and imbalance. These dimensions define nine representative degradation scenarios that can substantially affect TAG learning. Although prior studies have explored specific mitigation strategies, existing evidence remains fragmented across degradation types, datasets, tasks, and model families, leaving TAG robustness insufficiently understood. To address this gap, we present OpenRTAG, a robustness benchmark for text-attributed graph learning. OpenRTAG organizes TAG quality issues into a unified 3 * 3 taxonomy and supports standardized evaluation across nine TAG datasets and three downstream tasks. It systematically evaluates scenario validity and model sensitivity, compares traditional GNNs, LLM-GNNs, and a representative GFM, investigates the effectiveness, efficiency, and robustness of scenario-matched baselines, and further examines model behavior under composite degradation scenarios. OpenRTAG provides a standardized testbed for understanding robustness in TAG learning under realistic low-quality settings.

Figures

Figures reproduced from arXiv: 2607.19108 by Hongchao Qin, Qiangqiang Dai, Ronghua Li, Ruoyu Wu, Xunkai Li, Yan Zhao, Yuze Dai, Zekai Chen, Zhihan Zhang.

Figure 1
Figure 1. Figure 1: Overview of the OpenRTAG benchmark framework, covering scenario construction, model-paradigm robustness, [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Q1 model-paradigm robustness. Heatmaps report clean-to-degraded node-classification accuracy drops under the nine [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Q2 text-repair methods’ robustness under low and [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Runtime of representative Q2 text-oriented base [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Runtime of representative Q3 structure-oriented [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Runtime of representative Q4 label-oriented base [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Q5 Performance under different downstream tasks. [PITH_FULL_IMAGE:figures/full_fig_p007_9.png] view at source ↗

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

68 extracted references · 1 canonical work pages

  1. [1]

    Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education

    Clancey, William J. Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education. Proceedings of the Eighth International Joint Conference on Artificial Intelligence (IJCAI-83)

  2. [2]

    Classification Problem Solving

    Clancey, William J. Classification Problem Solving. Proceedings of the Fourth National Conference on Artificial Intelligence

  3. [3]

    , title =

    Robinson, Arthur L. , title =. 1980 , doi =. https://science.sciencemag.org/content/208/4447/1019.full.pdf , journal =

    1980

  4. [4]

    New Ways to Make Microcircuits Smaller---Duplicate Entry

    Robinson, Arthur L. New Ways to Make Microcircuits Smaller---Duplicate Entry. Science

  5. [5]

    Clancey and Glenn Rennels , abstract =

    Diane Warner Hasling and William J. Clancey and Glenn Rennels , abstract =. Strategic explanations for a diagnostic consultation system , journal =. 1984 , issn =. doi:https://doi.org/10.1016/S0020-7373(84)80003-6 , url =

    doi:10.1016/s0020-7373(84)80003-6 1984

  6. [6]

    and Rennels, Glenn R

    Hasling, Diane Warner and Clancey, William J. and Rennels, Glenn R. and Test, Thomas. Strategic Explanations in Consultation---Duplicate. The International Journal of Man-Machine Studies

  7. [7]

    Poligon: A System for Parallel Problem Solving

    Rice, James. Poligon: A System for Parallel Problem Solving

  8. [8]

    Transfer of Rule-Based Expertise through a Tutorial Dialogue

    Clancey, William J. Transfer of Rule-Based Expertise through a Tutorial Dialogue

  9. [9]

    The Engineering of Qualitative Models

    Clancey, William J. The Engineering of Qualitative Models

  10. [10]

    2023 , eprint=

    Attention Is All You Need , author=. 2023 , eprint=

    2023

  11. [11]

    Pluto: The 'Other' Red Planet

    NASA. Pluto: The 'Other' Red Planet

  12. [12]

    Advances in Neural Information Processing Systems , volume=

    A Comprehensive Study on Text-Attributed Graphs: Benchmarking and Rethinking , author=. Advances in Neural Information Processing Systems , volume=

  13. [13]

    Graph Learning in the Era of

    Li, Xunkai and Wu, Zhengyu and Wu, Jiayi and Cui, Hanwen and Jia, Jishuo and Li, Rong-Hua and Wang, Guoren , journal=. Graph Learning in the Era of

  14. [14]

    arXiv preprint arXiv:2406.10727 , year=

    Text-Space Graph Foundation Models: Comprehensive Benchmarks and New Insights , author=. arXiv preprint arXiv:2406.10727 , year=

    Pith/arXiv arXiv

  15. [15]

    arXiv preprint arXiv:2103.03036 , year=

    Deep Graph Structure Learning for Robust Representations: A Survey , author=. arXiv preprint arXiv:2103.03036 , year=

    Pith/arXiv arXiv

  16. [16]

    arXiv preprint arXiv:2408.08685 , year=

    Can Large Language Models Improve the Adversarial Robustness of Graph Neural Networks? , author=. arXiv preprint arXiv:2408.08685 , year=

    Pith/arXiv arXiv

  17. [17]

    Rethinking Graph Structure Learning in the Era of

    Zhang, Zhihan and Li, Xunkai and Lei, Zhu and Zeng, Guang and Li, Ronghua and Wang, Guoren , journal=. Rethinking Graph Structure Learning in the Era of

  18. [18]

    Qin, Jiawen and Yuan, Haonan and Sun, Qingyun and Xu, Lyujin and Yuan, Jiaqi and Huang, Pengfeng and Wang, Zhaonan and Fu, Xingcheng and Peng, Hao and Li, Jianxin and others , journal=

  19. [19]

    Harnessing Explanations:

    He, Xiaoxin and Bresson, Xavier and Laurent, Thomas and Perold, Adam and LeCun, Yann and Hooi, Bryan , journal=. Harnessing Explanations:

  20. [20]

    Companion Proceedings of the ACM Web Conference 2024 , year=

    Text-Attributed Graph Representation Learning , author=. Companion Proceedings of the ACM Web Conference 2024 , year=

    2024

  21. [21]

    2024 , eprint=

    TAGLAS: An atlas of text-attributed graph datasets in the era of large graph and language models , author=. 2024 , eprint=

    2024

  22. [22]

    Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , year=

    A Survey of Large Language Models for Graphs , author=. Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , year=

  23. [23]

    2023 , eprint=

    Towards Graph Foundation Models: A Survey and Beyond , author=. 2023 , eprint=

    2023

  24. [24]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

    Graph Foundation Models: Concepts, Opportunities and Challenges , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

  25. [25]

    Advances in Neural Information Processing Systems , year=

    Open Graph Benchmark: Datasets for Machine Learning on Graphs , author=. Advances in Neural Information Processing Systems , year=. 2005.00687 , archivePrefix=

    Pith/arXiv arXiv 2005

  26. [26]

    2021 , eprint=

    Benchmarking the Adversarial Robustness of Graph Machine Learning , author=. 2021 , eprint=

    2021

  27. [27]

    Advances in Neural Information Processing Systems Datasets and Benchmarks Track , year=

    NoisyGL: A Comprehensive Benchmark for Graph Neural Networks under Label Noise , author=. Advances in Neural Information Processing Systems Datasets and Benchmarks Track , year=. 2406.04299 , archivePrefix=

    Pith/arXiv arXiv

  28. [28]

    International Conference on Learning Representations , year=

    IGL-Bench: Establishing the Comprehensive Benchmark for Imbalanced Graph Learning , author=. International Conference on Learning Representations , year=. 2406.09870 , archivePrefix=

    Pith/arXiv arXiv

  29. [29]

    2025 , eprint=

    OpenGU: A Comprehensive Benchmark for Graph Unlearning , author=. 2025 , eprint=

    2025

  30. [30]

    International Conference on Learning Representations , year=

    Semi-Supervised Classification with Graph Convolutional Networks , author=. International Conference on Learning Representations , year=. 1609.02907 , archivePrefix=

    Pith/arXiv arXiv

  31. [31]

    International Conference on Learning Representations , year=

    Graph Attention Networks , author=. International Conference on Learning Representations , year=. 1710.10903 , archivePrefix=

    Pith/arXiv arXiv

  32. [32]

    Advances in Neural Information Processing Systems , year=

    Inductive Representation Learning on Large Graphs , author=. Advances in Neural Information Processing Systems , year=

  33. [33]

    International Conference on Learning Representations , year=

    Node Feature Extraction by Self-Supervised Multi-scale Neighborhood Prediction , author=. International Conference on Learning Representations , year=

  34. [34]

    Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence , year=

    Efficient Tuning and Inference for Large Language Models on Textual Graphs , author=. Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence , year=. 2401.15569 , archivePrefix=

    Pith/arXiv arXiv

  35. [35]

    Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval , year=

    GraphGPT: Graph Instruction Tuning for Large Language Models , author=. Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval , year=

  36. [36]

    Proceedings of the ACM Web Conference 2024 , year=

    Can GNN be Good Adapter for LLMs? , author=. Proceedings of the ACM Web Conference 2024 , year=. 2402.12984 , archivePrefix=

    Pith/arXiv arXiv 2024

  37. [37]

    2023 , eprint=

    GraphText: Graph Reasoning in Text Space , author=. 2023 , eprint=

    2023

  38. [38]

    International Conference on Learning Representations , year=

    One for All: Towards Training One Graph Model for All Classification Tasks , author=. International Conference on Learning Representations , year=. 2310.00149 , archivePrefix=

    Pith/arXiv arXiv

  39. [39]

    Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

    ZeroG: Investigating Cross-dataset Zero-shot Transferability in Graphs , author=. Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=. 2024 , eprint=

    2024

  40. [40]

    International Conference on Learning Representations , year=

    Harnessing Explanations: LLM-to-LM Interpreter for Enhanced Text-Attributed Graph Representation Learning , author=. International Conference on Learning Representations , year=. 2305.19523 , archivePrefix=

    Pith/arXiv arXiv

  41. [41]

    Findings of the Association for Computational Linguistics: EMNLP 2024 , pages=

    OpenGraph: Towards Open Graph Foundation Models , author=. Findings of the Association for Computational Linguistics: EMNLP 2024 , pages=. 2024 , doi=

    2024

  42. [42]

    2025 , eprint=

    Toward General and Robust LLM-enhanced Text-attributed Graph Learning , author=. 2025 , eprint=

    2025

  43. [43]

    Proceedings of the 5th Workshop on Noisy User-generated Text , year=

    Contextual Text Denoising with Masked Language Model , author=. Proceedings of the 5th Workshop on Noisy User-generated Text , year=. 1910.14080 , archivePrefix=

    Pith/arXiv arXiv 1910

  44. [44]

    Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing , year=

    Denoising based Sequence-to-Sequence Pre-training for Text Generation , author=. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing , year=. 1908.08206 , archivePrefix=

    Pith/arXiv arXiv 2019

  45. [45]

    A Text Normalisation System for Non-Standard

    Flint, Emma and Ford, Elliot and Thomas, Olivia and Caines, Andrew and Buttery, Paula , booktitle =. A Text Normalisation System for Non-Standard. 2017 , pages =. doi:10.18653/v1/W17-4414 , url =

    work page doi:10.18653/v1/w17-4414 2017

  46. [46]

    Natural Language Generation , year =

    The Practical Value of N-Grams Is in Generation , author =. Natural Language Generation , year =

  47. [47]

    2019 , pages =

    Devlin, Jacob and Chang, Ming-Wei and Lee, Kenton and Toutanova, Kristina , booktitle =. 2019 , pages =. doi:10.18653/v1/N19-1423 , url =

    doi:10.18653/v1/n19-1423 2019

  48. [48]

    Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , year=

    GAugLLM: Improving Graph Contrastive Learning for Text-Attributed Graphs with Large Language Models , author=. Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , year=. 2406.11945 , archivePrefix=

    Pith/arXiv arXiv

  49. [49]

    IEEE Transactions on Neural Networks and Learning Systems , year=

    Homophily-Enhanced Self-Supervision for Graph Structure Learning: Insights and Directions , author=. IEEE Transactions on Neural Networks and Learning Systems , year=

  50. [50]

    Proceedings of the ACM Web Conference 2023 , year=

    SE-GSL: A General and Effective Graph Structure Learning Framework through Structural Entropy Optimization , author=. Proceedings of the ACM Web Conference 2023 , year=. 2303.09778 , archivePrefix=

    Pith/arXiv arXiv 2023

  51. [51]

    2022 , eprint=

    Reliable Representations Make A Stronger Defender: Unsupervised Structure Refinement for Robust GNN , author=. 2022 , eprint=

    2022

  52. [52]

    Proceedings of the ACM Web Conference 2022 , year=

    Towards Unsupervised Deep Graph Structure Learning , author=. Proceedings of the ACM Web Conference 2022 , year=. 2201.06367 , archivePrefix=

    Pith/arXiv arXiv 2022

  53. [53]

    Advances in Neural Information Processing Systems , year=

    Mitigating Degree Bias for Graph Neural Networks via Test-Time Augmentation , author=. Advances in Neural Information Processing Systems , year=. 2310.00800 , archivePrefix=

    Pith/arXiv arXiv

  54. [54]

    Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , year=

    GraphSHA: Synthesizing Harder Samples for Class-Imbalanced Node Classification , author=. Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , year=. 2306.09612 , archivePrefix=

    Pith/arXiv arXiv

  55. [55]

    Proceedings of the 31st ACM International Conference on Information and Knowledge Management , year=

    LTE4G: Long-Tail Experts for Graph Neural Networks , author=. Proceedings of the 31st ACM International Conference on Information and Knowledge Management , year=. 2208.10205 , archivePrefix=

    Pith/arXiv arXiv

  56. [56]

    Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , year=

    Tail-GNN: Tail-Node Graph Neural Networks , author=. Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , year=

  57. [57]

    International Conference on Machine Learning , year=

    TAM: Topology-Aware Margin Loss for Class-Imbalanced Node Classification , author=. International Conference on Machine Learning , year=. 2206.12917 , archivePrefix=

    Pith/arXiv arXiv

  58. [58]

    Proceedings of the 30th ACM International Conference on Multimedia , pages=

    A Unified Framework against Topology and Class Imbalance , author=. Proceedings of the 30th ACM International Conference on Multimedia , pages=

  59. [59]

    Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval , year=

    GraFN: Semi-Supervised Node Classification on Graph with Few Labels via Non-Parametric Distribution Assignment , author=. Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval , year=. 2204.01303 , archivePrefix=

    Pith/arXiv arXiv

  60. [60]

    2021 , eprint=

    GraphHop: An Enhanced Label Propagation Method for Node Classification , author=. 2021 , eprint=

    2021

  61. [61]

    Transactions on Machine Learning Research , year=

    Noise-robust Graph Learning by Estimating and Leveraging Pairwise Interactions , author=. Transactions on Machine Learning Research , year=. 2106.07451 , archivePrefix=

    Pith/arXiv arXiv

  62. [62]

    Proceedings of the 16th ACM International Conference on Web Search and Data Mining , year=

    Robust Training of Graph Neural Networks via Noise Governance , author=. Proceedings of the 16th ACM International Conference on Web Search and Data Mining , year=. 2211.06614 , archivePrefix=

    Pith/arXiv arXiv

  63. [63]

    Proceedings of the AAAI Conference on Artificial Intelligence , year=

    Robust Node Classification on Graph Data with Graph and Label Noise , author=. Proceedings of the AAAI Conference on Artificial Intelligence , year=

  64. [64]

    International Conference on Machine Learning , year=

    Revisiting Semi-Supervised Learning with Graph Embeddings , author=. International Conference on Machine Learning , year=

  65. [65]

    arXiv preprint arXiv:2007.02901 , year=

    Wiki-CS: A Wikipedia-Based Benchmark for Graph Neural Networks , author=. arXiv preprint arXiv:2007.02901 , year=

    Pith/arXiv arXiv 2007

  66. [66]

    Proceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Retrieval , year=

    Image-based Recommendations on Styles and Substitutes , author=. Proceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Retrieval , year=

  67. [67]

    Advances in Neural Information Processing Systems Datasets and Benchmarks Track , year=

    GLBench: A Comprehensive Benchmark for Graph with Large Language Models , author=. Advances in Neural Information Processing Systems Datasets and Benchmarks Track , year=

  68. [68]

    2024 , eprint=

    GraphEdit: Large Language Models for Graph Structure Learning , author=. 2024 , eprint=

    2024

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.