Pith. sign in

REVIEW 2 major objections 7 minor 58 references

Semi-Supervised Text-Attributed Graph Distillation

T0 review · 2 major / 7 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read STAD compresses text-attributed graphs into a handful of readable nodes that match or beat full-data classifiers.

desk verdict First dedicated TAG distillation framework with code and a correct WSD-to-K-means lemma, but Algorithm 2 leaks test labels into the condensed graph, so the headline semi-supervised numbers are not supported as written. read the letter →

arxiv 2607.20477 v1 pith:PY76NM2O submitted 2026-05-27 cs.AI

Yurui Lai , Samir Moustafa , Renchi Yang , Tsz Nam Chan This is my paper
classification cs.AI
keywords text-attributedgraphsgraphdistillationcondensationsemi-supervisedlearningWassersteindistanceLLMtextsynthesispseudo-labelingnodeclassification
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to show that a large text-attributed graph — where nodes carry both links and text — can be distilled into a tiny graph of a few synthetic nodes, each with a short human-readable text summary, without losing classification accuracy, even when only a small fraction of nodes are labeled. The authors claim their framework, STAD, achieves this with a state-of-the-art performance-compression trade-off: at distillation ratio 0.05 on Cora, the condensed graph reaches 77.5% accuracy versus 77.7% for a GCN trained on the full graph, and on WikiCS the condensed graph at ratio 0.5 beats the full GCN (76.8% vs 75.9%). If true, this matters because it would make LLM-based analytics over graph data dramatically cheaper: downstream preprocessing and training times drop by one to two orders of magnitude, and the distilled graphs remain interpretable to humans.

What carries the argument

The 2-Wasserstein distance (WSD) — the optimal-transport cost between two distributions — is the guiding metric and theoretical backbone. Lemma 4.1 proves that for a sketching matrix S (one nonzero entry per row, a hard cluster assignment), minimizing WSD(M, S^T M) equals the K-Means within-cluster sum of squares on M. This turns distillation into clustering: K-Means on the fused soft-label distribution P, refined by a greedy class-affinity reassignment, yields the condensed adjacency, attribute, and label matrices. Collaborative self-training (CoST) uses consensus verification between GCN and MLP pathways to expand pseudo-labels and fuse complementary graph/text features.

What would settle it

Construct or find a TAG where the greedy reassignment is disabled (or run with only V_tr labels), and check whether cluster purity and downstream accuracy collapse relative to the reported numbers; if a version using only the small labeled set matches the full-label version, the dependence on Y is not load-bearing, but if accuracy drops significantly, the semi-supervised claim fails. A second check: verify Lemma 4.1's Voronoi optimality after the greedy reassignment — if reassignment breaks the Voronoi property, WSD is no longer minimized and the theoretical grounding of the sketching step is

Watch

Extended reading notes

Core claim

The central claim is that minimizing the 2-Wasserstein distance between original and condensed text-attributed graphs is a sound guide for distillation, and that doing so after fusing graph-aware (GCN) and graph-free (MLP) features via collaborative self-training yields condensed graphs whose classification accuracy matches or exceeds full-data training. The condensed graph is built from a sketching matrix that acts as a hard clustering: Lemma 4.1 proves that minimizing WSD(M, S^T M) equals the K-Means objective, so clustering the fused soft-label distribution produces the condensed adjacency, attribute, and label matrices. Each condensed node's text is generated from cluster keywords by an

Load-bearing premise

The load-bearing premise is that Algorithm 2 can compute the class-cluster affinities it needs — but the affinity formula uses the full ground-truth label matrix Y (Algorithm 2, lines 5–7), whereas the paper's semi-supervised setting assumes only a small labeled set V_tr; if full labels are unavailable, the reassignment step as written cannot run, and if only V_tr is available, the claimed semi-supervised regime is not what the algorithm actually consumes.

Editorial extensions

If this is right

  • Compression ratios of hundreds to thousands of times are achievable with little or no accuracy loss: e.g., Cora shrinks from 2,708 nodes to 7 with accuracy 77.5% vs 77.7%.
  • Distilled human-readable graphs can replace full graphs for LLM-based node classification, with preprocessing speedups up to hundreds of times.
  • WSD acts as a reliable, theoretically grounded proxy for downstream performance, so distillation quality can be monitored without training a classifier.
  • Synthetic texts work as in-context-learning exemplars, improving training-free LLM classification (e.g., +17.7% on History with neighbor summarization).
  • Low-cost smaller LLMs match or outperform larger, more expensive ones for text synthesis, decoupling distillation quality from API budget.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper's Lemma 4.1 suggests a general recipe — sketch any feature space by K-Means when WSD correlates with task performance; this could extend to dynamic graphs or multimodal graphs beyond text, though those settings are not tested.
  • Algorithm 2's greedy reassignment computes class-cluster affinities from the full ground-truth label matrix Y, while the problem states only a small labeled set V_tr is available; if full labels are not actually accessible, this step must be driven by CoST pseudo-labels, and the paper does not isolate the impact of that substitution.
  • The WSD-validation-guided candidate selection uses validation accuracy on the original dataset; in a purely semi-supervised deployment with no validation labels, that selection would need an unsupervised proxy, which the paper does not provide.
  • A testable extension: measure whether the distilled texts preserve calibration or out-of-distribution robustness, not just accuracy, since the condensation sharpens class clusters and may over-clean the data.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 7 minor

Summary. The paper proposes STAD, a semi-supervised text-attributed graph (TAG) distillation framework. STAD has three main modules: (1) a graph-text collaborative encoding module that fuses GCN and MLP representations with a collaborative self-training scheme to generate pseudo-labels; (2) a Wasserstein-distance-based graph sketching algorithm that clusters nodes (via K-means on fused soft labels) and constructs a condensed graph (A', X', Y'); and (3) a keyword-based LLM text synthesis module that generates human-readable summaries for condensed nodes. Experiments on seven TAG datasets compare STAD against graph distillation, text distillation, and topology-removed baselines, reporting state-of-the-art accuracy under aggressive compression ratios (e.g., r=0.05), with STAD sometimes matching or exceeding full-data GCN/MLP performance. The paper also evaluates transfer to LLM4TAG methods and reports efficiency gains.

Significance. If the method worked as claimed, this would be a meaningful contribution to scalable TAG learning, especially for LLM-based downstream tasks, by producing tiny, interpretable condensed graphs that preserve task-relevant information. The paper ships a public code repository, includes extensive experiments, and provides a formal lemma (Appendix B) connecting the WSD sketching objective to K-means, which is mathematically sound under the stated one-hot/equal-weight sketch construction. The core idea of using WSD as a distillation-quality metric and generating readable texts via keyword-guided LLM summarization is interesting and potentially useful.

major comments (2)
  1. [§4.2 and Algorithm 2 (Appendix A)] The greedy reassignment step in Algorithm 2 uses the full ground-truth label matrix Y to compute class–cluster affinities: |{v_i in C_j : Y_{i,k}=1}| / |{v_i in V : Y_{i,k}=1}|. The problem formulation in §3.1 states that only a small labeled set V_tr is available under the semi-supervised setting. Thus, as written, the algorithm either leaks test labels into the construction of G' (and therefore into the reported accuracies) or is not executable from the allowed labeled data. The leak is load-bearing: Table 5 shows that removing reassignment drops accuracy substantially (e.g., WikiCS r=0.05 from 76.6% to 68.1%; Cora from 77.5% to 75.2%). The headline SOTA claims (e.g., WikiCS r=0.5: 76.8% vs. full GCN 75.9%) are therefore not supported by a valid semi-supervised evaluation. The authors should either replace Y with pseudo-labels derived from V_tr (e.g., from CoST) and rerun all experimen
  2. [§4.3 and §5.6] The WSD-validation-guided selection of synthetic texts (Algorithm 3, Step 3) and the checkpoint selection in CoST (§5.6) use 'validation accuracy' on the original dataset. The experimental setup (§5.1) only describes 'five random splits with 20 nodes per class' — it does not specify a separate validation split. If the validation accuracy is computed using the full ground-truth labels (including test nodes), this constitutes additional test-label leakage in model/text selection. The paper must specify how validation is constructed, ensure that it uses only the labeled training set (or a distinct held-out split) and rerun the selection procedures accordingly.
minor comments (7)
  1. [§2] Typo: 'We In this work, we define...' should be 'In this work, we define...'.
  2. [§5 (intro)] Typo: 'how can G' generated by STAD help LL4TAG methods?' should be 'LLM4TAG'.
  3. [Table 2] Formatting issues: many cells lack proper separators (e.g., '76.2±1.678.7±0.8'), making the table hard to read.
  4. [Figure 2] The caption contains garbled unicode sequences (e.g., '/uni00000013/uni00000011/...') and should be regenerated.
  5. [Algorithm 1 (Appendix A)] The algorithm input lists the full label matrix Y, but the algorithm's steps only appear to use the labels in V_tr. Clarify whether Y is actually needed and, if so, why.
  6. [§4.2] 'Lemma 4.11 implies' should be 'Lemma 4.1 implies'.
  7. [Eq. (1)] The bi-level formulation's constraint uses \phi^{G'} but \phi^{G'} is not defined; the notation should be clarified.

Circularity Check

1 steps flagged · score 6.0 of 10

Algorithm 2's greedy reassignment uses the full ground-truth label matrix Y to build the condensed graph, so the reported semi-supervised 'prediction' is partially fitted to the target labels.

  1. fitted input called prediction [§4.2 'WSD-based Graph Sketching' (greedy reassignment paragraph; Algorithm 2, Appendix A)]
    "we first evaluate the affinity between each class label y_k∈Y (1≤k≤K) and each cluster C_j∈{C_1,C_2,...,C_N'} by computing the fraction of the total population of label y_k that falls into C_j: |{v_i∈C_j | Y_{i,k}=1}| / |{v_i∈V | Y_{i,k}=1}|. We then pick, for each class, the class-cluster pair with the highest affinity value, and then select the remaining class-cluster pairs with the highest affinity values among all the K×N' pairs until obtaining top-N' pairs in total."

    §3.1 defines the semi-supervised setting with 'only a small set of nodes Vtr⊆V (|Vtr|≪N) are labeled' and defines Y as the full ground-truth label matrix. Algorithm 2 nevertheless takes Y as input and uses it to reassign all nodes to clusters; S then determines A', X', and the clusters whose keywords generate S'. Thus the condensed graph is constructed from the very labels the paper later reports predicting in Tables 2–4. Table 5 shows the reassignment is load-bearing (WikiCS 76.6→68.1, Cora 77.5→75.2 without it), so the semi-supervised SOTA accuracy is, in part, fitted to the target labels rather than independently predicted.

full rationale

Aside from the Algorithm 2 leak, the derivation is largely self-contained: Lemma 4.1 is proved in Appendix B (it correctly equates WSD(M,S^T M) minimization with the K-Means sum-of-squares), the WSD-accuracy link in §3.3 is an external empirical correlation, and the experimental comparisons are against external baselines. The self-citation to ClustGDD [14] is used as a baseline/motivation but is not load-bearing for STAD's central claim. However, the central semi-supervised claim is compromised by the full-Y reassignment: the greedy affinity computation requires labels for all nodes, not just V_tr, so the condensation G' encodes the target labels. Because the paper's own ablation shows this reassignment accounts for part of the reported accuracy, the SOTA result is partially a fitted-input-called-prediction. A further leakage risk (not counted separately) is §4.3's selection of final synthetic texts by 'validation accuracy on the original dataset' with no described held-out validation split. Score 6: one load-bearing prediction partially reduces to its own inputs.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The central method rests on an empirical distribution-shift assumption, standard K-means/OT mathematics, and semi-supervised pseudo-labeling. The most serious item is the implicit axiom that full ground-truth labels are available for sketching, which contradicts the paper's own problem definition. Hyperparameters such as κ and ξ are tuned per dataset and inflate reported performance.

free parameters (6)
  • keyword count ξ = not fixed; case study uses 256; ablated 64–512
    Chosen by hand per experiment; affects LLM input and downstream accuracy.
  • LLM output token limit κ = Cora κ=256, History κ=12 in Fig. 6b
    Selected per dataset based on validation accuracy, so it is fitted to the downstream task.
  • candidate summaries per cluster q = typically 3; ablated 2–5
    Manual hyperparameter; larger q increases selection pool.
  • candidate graph samples n_s = typically 100; ablated 10–200
    Manual hyperparameter controlling WSD-validation cost and selection quality.
  • top-k validation candidates k = often 5
    Manual hyperparameter for choosing the final synthetic text set.
  • pseudo-label ratio ρ increment = 0.1 per epoch until 1.0
    Borrowed from prior self-training work; not derived from the task.
assumptions (6)
  • domain assumption WSD between original and distilled distributions correlates with downstream classification accuracy.
    Used to justify the sketching objective; supported only by the paper's own Fig. 2 across four datasets.
  • standard math The sketching matrix is one-hot and cluster centroids are weighted by 1/|C_k|, so WSD minimization reduces to K-means.
    This is Lemma 4.1; the proof in Appendix B relies on the Voronoi property of K-means partitions.
  • domain assumption Consensus pseudo-labels from GCN and MLP pathways are reliable enough to expand the training set.
    Core to the semi-supervised mechanism; no independent evidence that error accumulation is avoided beyond the reported ablations.
  • domain assumption LLMs can generate coherent, label-consistent text summaries from cluster keywords.
    The entire text-synthesis module rests on this; only qualitative examples and downstream accuracy are provided as evidence.
  • domain assumption SBERT embeddings capture semantic similarity for keyword filtering and text-attribute encoding.
    Used for pruning keywords and encoding S' to X''; standard practice but load-bearing.
  • ad hoc to paper Full ground-truth labels Y are available for cluster reassignment in Algorithm 2.
    The semi-supervised setting says only V_tr is labeled, but Algorithm 2 computes affinities over all nodes' true labels. This is the leakage-enabling assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Semi-Supervised Text-Attributed Graph Distillation." pith.science (2026). https://pith.science/paper/PY76NM2O

@misc{pith2026260720477,
  author       = {Pith},
  title        = {Pith review of: Semi-Supervised Text-Attributed Graph Distillation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PY76NM2O}},
  note         = {Machine review of arXiv:2607.20477}
}
read the original abstract

{\em Text-Attributed Graphs} (TAGs) have emerged as an expressive data model for integrating graph topology with rich textual semantics. Existing representation learning methods over TAGs suffer from severe scalability bottlenecks, particularly together with {\em Large Language Models} (LLMs). While data distillation offers a promising data-centric solution, existing methods fail to capture the complex interplay between graph and text modalities, struggle with the label scarcity inherent in semi-supervised settings, and lack the ability to produce the human-readable textual attributes required for downstream LLM-based tasks. To address these challenges, we propose \algo{}, a unified semi-supervised framework guided by the {\em Wasserstein Distance} (WSD). Grounded in our empirical findings on real TAGs, \algo{} introduces a graph-text collaborative encoding module that utilizes dual-pathway encoders (graph-aware and -free) within a collaborative self-training scheme to harvest reliable pseudo-labels and fuse complementary graph-text features. Furthermore, we develop a theoretically grounded WSD-based graph sketching algorithm and a cost-effective LLM text synthesis module, which leverages cluster-based keyword extraction to generate coherent, human-readable summaries for condensed nodes. Extensive experiments on benchmark datasets demonstrate that \algo{} achieves a state-of-the-art performance-compression trade-off in terms of both GNN- and LLM-based downstream tasks, enabling effective and efficient TAG learning or analytics.

Figures

Figures reproduced from arXiv: 2607.20477 by the authors.

Figure 1
Figure 1. Complementary coverage of GA vs. GF on TAGs. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. WSD and node classification accuracy on condensed TAGs vs. distillation ratio. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The Overview of STAD true for the latter. The pseudo-labels of the nodes in R DS GA and R DS GF are thus determined by P GA and P GF, respectively. Accordingly, the labeled node sets are expanded as follows: U = U ∪ RDS GA ∪ RDS GF, UGA = UGA ∪ RDS GA, and UGF = UGF ∪ RDS GF . 4.2 WSD-based Graph Sketching After the obtainment of soft labels P, the next step in STAD is to con￾struct a sparse sketching matrix S ∈ R 𝑁… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Accuracy by P, P GA, and P GF when varying #iterations. Cora History Photo WikiCS 0.6 0.65 0.7 0.75 0.8 0.85 Purity Value Before After [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Cluster purity before/after greedy reassignment. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 7
Figure 7. Figure 7: t-SNE of node attributes: original (X) vs. synthetic (X ′ ) on Cora and History. potential direction for future work is to explore multi-modal graph distillation for continual learning. 6 Acknowledgments Renchi Yang is supported by the Guangdong and Hong Kong Uni￾versi…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

58 extracted references · 5 linked inside Pith

  1. [1]

    Bo, D., W ang, X., Shi, C., and Shen, H.Beyond low-frequency information in graph convolutional networks.Proceedings of the AAAI Conference on Artificial Intelligence 35, 5 (May 2021), 3950–3957

    2021

  2. [2]

    K., Shah, N., and W ang, Z.Llaga: Large language and graph assistant

    Chen, R., Zhao, T., Jaiswal, A. K., Shah, N., and W ang, Z.Llaga: Large language and graph assistant. InInternational Conference on Machine Learning(2024), PMLR, pp. 7809–7823

    2024

  3. [3]

    Chen, Z., Mao, H., Liu, J., Song, Y., Li, B., Jin, W., Fatemi, B., Tsitsulin, A., Perozzi, B., Liu, H., et al.Text-space graph foundation models: Comprehensive benchmarks and new insights.Advances in Neural Information Processing Systems 37(2024), 7464–7492

    2024

  4. [4]

    L., and Woodruff, D

    Clarkson, K. L., and Woodruff, D. P.Low-rank approximation and regression in input sparsity time.Journal of the ACM (JACM) 63, 6 (2017), 1–45

    2017

  5. [5]

    arXiv preprint arXiv:2406.14683(2024)

    Feng, J., Liu, H., Kong, L., Zhu, M., Chen, Y., and Zhang, M.Taglas: An atlas of text-attributed graph datasets in the era of large graph and language models. arXiv preprint arXiv:2406.14683(2024)

    arXiv 2024

  6. [6]

    In Proceedings of the ACM on Web Conference 2025(2025), pp

    Gao, X., Ye, G., Chen, T., Zhang, W., Yu, J., and Yin, H.Rethinking and accel- erating graph condensation: A training-free approach with class partition. In Proceedings of the ACM on Web Conference 2025(2025), pp. 4359–4373

    2025

  7. [7]

    Hamilton, W., Ying, Z., and Leskovec, J.Inductive representation learning on large graphs.Advances in neural information processing systems 30(2017)

    2017

  8. [8]

    InICLR(2024)

    He, X., Bresson, X., Laurent, T., Perold, A., LeCun, Y., and Hooi, B.Har- nessing explanations: Llm-to-lm interpreter for enhanced text-attributed graph representation learning. InICLR(2024)

    2024

Show all 58 references
  1. [9]

    InJoint European Conference on Machine Learning and Knowledge Discovery in Databases(2010), Springer, pp

    Ji, M., Sun, Y., Danilevsky, M., Han, J., and Gao, J.Graph regularized trans- ductive classification on heterogeneous information networks. InJoint European Conference on Machine Learning and Knowledge Discovery in Databases(2010), Springer, pp. 570–586

    2010

  2. [10]

    Jin, B., Liu, G., Han, C., Jiang, M., Ji, H., and Han, J.Large language models on graphs: A comprehensive survey.IEEE Transactions on Knowledge and Data Engineering(2024)

    2024

  3. [11]

    InInternational Conference on Learning Representations (2022)

    Jin, W., Zhao, L., Zhang, S., Liu, Y., Tang, J., and Shah, N.Graph condensation for graph neural networks. InInternational Conference on Learning Representations (2022)

    2022

  4. [12]

    InFindings of the Association for Computational Linguistics: ACL 2023(2023), pp

    Karisani, P.Neural networks against (and for) self-training: Classification with small labeled and large unlabeled sets. InFindings of the Association for Computational Linguistics: ACL 2023(2023), pp. 12148–12162

    2023

  5. [13]

    N., and Welling, M.Semi-supervised classification with graph con- volutional networks

    Kipf, T. N., and Welling, M.Semi-supervised classification with graph con- volutional networks. InInternational Conference on Learning Representations (2017)

    2017

  6. [14]

    InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V

    Lai, Y., Zhang, T., and Yang, R.Simple yet effective graph distillation via clustering. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2(2025), pp. 1229–1240

    2025

  7. [15]

    InWorkshop on challenges in representation learning, ICML(2013), vol

    Lee, D.-H., et al.Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. InWorkshop on challenges in representation learning, ICML(2013), vol. 3, Atlanta, p. 896

    2013

  8. [16]

    Lei, S., and Tao, D.A comprehensive survey of dataset distillation.IEEE Transactions on Pattern Analysis and Machine Intelligence 46, 1 (2023), 17–32

    2023

  9. [17]

    InProceedings of the AAAI conference on artificial intelligence(2018), vol

    Li, Q., Han, Z., and Wu, X.-M.Deeper insights into graph convolutional networks for semi-supervised learning. InProceedings of the AAAI conference on artificial intelligence(2018), vol. 32

    2018

  10. [18]

    Li, T., Fang, Z., Zhang, X., Tang, K., Chen, H., Jiang, Z., Zhao, T., Xu, R., Cheng, F., Li, X., et al.Druglm: A unified framework to enhance drug-target interaction predictions by incorporating textual embeddings via language models.bioRxiv (2025), 2025–07

    2025

  11. [19]

    Li, Y., and Li, W.Data distillation for text classification.arXiv preprint arXiv:2104.08448(2021)

    2021 arXiv

  12. [20]

    X.A survey of graph meets large language model: progress and future directions

    Li, Y., Li, Z., W ang, P., Li, J., Sun, X., Cheng, H., and Yu, J. X.A survey of graph meets large language model: progress and future directions. InProceedings of the Thirty-Third International Joint Conference on Artificial Intelligence(2024), pp. 8123–8131

    2024

  13. [21]

    S., et al.Towards graph foundation models: A survey and beyond.arXiv preprint arXiv:2310.11829(2023)

    Liu, J., Y ang, C., Lu, Z., Chen, J., Li, Y., Zhang, M., Bai, T., Fang, Y., Sun, L., Yu, P. S., et al.Towards graph foundation models: A survey and beyond.arXiv preprint arXiv:2310.11829(2023)

    2023 arXiv

  14. [22]

    Liu, M., Li, S., Chen, X., and Song, L.Graph condensation via receptive field distribution matching.arXiv preprint arXiv:2206.13697(2022)

    2022 arXiv

  15. [23]

    In International Conference on Machine Learning(2024), PMLR, pp

    Liu, Y., Bo, D., and Shi, C.Graph distillation with eigenbasis matching. In International Conference on Machine Learning(2024), PMLR, pp. 30702–30717

    2024

  16. [24]

    Lloyd, S.Least squares quantization in pcm.IEEE transactions on information theory 28, 2 (1982), 129–137

    1982

  17. [25]

    InAdvances in Neural Information Processing Systems(2022), S

    Luan, S., Hua, C., Lu, Q., Zhu, J., Zhao, M., Zhang, S., Chang, X.-W., and Precup, D.Revisiting heterophily for graph neural networks. InAdvances in Neural Information Processing Systems(2022), S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, Eds., vol. 35, C...

    2022

  18. [26]

    InACL Short Papers(Toronto, Canada, July 2023), A

    Maekawa, A., Kobayashi, N., Funakoshi, K., and Okumura, M.Dataset distil- lation with attention labels for fine-tuning BERT. InACL Short Papers(Toronto, Canada, July 2023), A. Rogers, J. Boyd-Graber, and N. Okazaki, Eds., Association for Computational Linguistics, pp. 119–127

    2023

  19. [27]

    Maekawa, A., Kosugi, S., Funakoshi, K., and Okumura, M.Dilm: Distilling dataset into language model for text-level dataset distillation.Journal of Natural Language Processing 32, 1 (2025), 252–282

    2025

  20. [28]

    M.Text embeddings reveal (almost) as much as text

    Morris, J., Kuleshov, V., Shmatikov, V., and Rush, A. M.Text embeddings reveal (almost) as much as text. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing(2023), pp. 12448–12460

    2023

  21. [29]

    J.Exploring the limits of transfer learning with a unified text-to-text transformer.Journal of machine learning research 21, 140 (2020)

    Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., and Liu, P. J.Exploring the limits of transfer learning with a unified text-to-text transformer.Journal of machine learning research 21, 140 (2020)

    2020

  22. [30]

    InEMNLP-IJCNLP(2019), pp

    Reimers, N., and Gurevych, I.Sentence-bert: Sentence embeddings using siamese bert-networks. InEMNLP-IJCNLP(2019), pp. 3982–3992

    2019

  23. [31]

    Survey Certification

    Sachdeva, N., and McAuley, J.Data distillation: A survey.Transactions on Machine Learning Research(2023). Survey Certification

    2023

  24. [32]

    Solomon, J., de Goes, F., Peyré, G., Cuturi, M., Butscher, A., Nguyen, A., Du, T., and Guibas, L.Convolutional wasserstein distances: efficient optimal transportation on geometric domains.ACM Trans. Graph. 34, 4 (July 2015)

    2015

  25. [33]

    Solomon, J., De Goes, F., Peyré, G., Cuturi, M., Butscher, A., Nguyen, A., Du, T., and Guibas, L.Convolutional wasserstein distances: Efficient optimal transportation on geometric domains.ACM Transactions on Graphics (ToG) 34, 4 (2015), 1–11

    2015

  26. [34]

    Su, G., Wang, H., Wang, J., Zhang, W., Zhang, Y., and Pei, J.Large language models meet text-attributed graphs: A survey of integration frameworks and applications.arXiv preprint arXiv:2510.21131(2025)

    2025

  27. [35]

    InIJCNN(2021), IEEE, pp

    Sucholutsky, I., and Schonlau, M.Soft-label dataset distillation and text dataset distillation. InIJCNN(2021), IEEE, pp. 1–8

    2021

  28. [36]

    InInternational Conference on Database Systems for Advanced Applications(2025), Springer, pp

    Sun, S., Ren, Y., Chen, J., and Ma, C.Large language models as topological structure enhancers for text-attributed graphs. InInternational Conference on Database Systems for Advanced Applications(2025), Springer, pp. 106–122

    2025

  29. [37]

    Graphgpt: Graph instruction tuning for large language models

    Tang, J., Yang, Y., Wei, W., Shi, L., Su, L., Cheng, S., Yin, D., and Huang, C. Graphgpt: Graph instruction tuning for large language models. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval(2024), pp. 491–500

    2024

  30. [38]

    InFindings of the Association for Computational Linguistics: EMNLP 2024(2024), pp

    Tao, Y., Kong, L., Kan, A., and Callot, L.Textual dataset distillation via language model embedding. InFindings of the Association for Computational Linguistics: EMNLP 2024(2024), pp. 12557–12569

    2024

  31. [39]

    Graph attention networks

    Veličković, P., Cucurull, G., Casanova, A., Romero, A., Liò, P., and Bengio, Y. Graph attention networks. InInternational Conference on Learning Representations (2018)

    2018

  32. [40]

    1412–1432

    Wang, Z., Liu, S., Zhang, Z., Ma, T., Zhang, C., and Ye, Y.Can llms convert graphs to text-attributed graphs? InNAACL(2025), pp. 1412–1432

    2025

  33. [41]

    InForty-second International Conference on Machine Learning(2025)

    Wu, X., Shen, Y., Ge, F., Shan, C., Jiao, Y., Sun, X., and Cheng, H.When do llms help with node classification? a comprehensive analysis. InForty-second International Conference on Machine Learning(2025)

    2025

  34. [42]

    Y an, H., Li, C., Long, R., Y an, C., Zhao, J., Zhuang, W., Yin, J., Zhang, P., Han, W., Sun, H., et al.A comprehensive study on text-attributed graphs: Benchmarking and rethinking.Advances in Neural Information Processing Systems 36(2023)

    2023

  35. [43]

    Y ang, A., Li, A., Y ang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Gao, C., Huang, C., Lv, C., et al.Qwen3 technical report.arXiv preprint arXiv:2505.09388(2025)

    2025 arXiv

  36. [44]

    Y ang, B., W ang, K., Sun, Q., Ji, C., Fu, X., Tang, H., You, Y., and Li, J.Does graph distillation see like vision dataset counterpart?Advances in Neural Information Processing Systems 36(2023), 53201–53226

    2023

  37. [45]

    InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V

    Y ang, C., Liu, H., W ang, D., Zhang, Z., Y ang, C., and Shi, C.Flag: Fraud detection with llm-enhanced graph neural network. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2(2025)

    2025

  38. [46]

    InInternational conference on machine learning(2016), PMLR, pp

    Y ang, Z., Cohen, W., and Salakhudinov, R.Revisiting semi-supervised learning with graph embeddings. InInternational conference on machine learning(2016), PMLR, pp. 40–48

    2016

  39. [47]

    C., Yang, M., Ying, R., and Lauw, H

    Zhang, D. C., Yang, M., Ying, R., and Lauw, H. W.Text-attributed graph representation learning: Methods, applications, and challenges. InCompanion Proceedings of the ACM Web Conference 2024(2024), pp. 1298–1301

    2024

  40. [48]

    S., and Zhang, J.A systematic survey of text summarization: From statistical methods to large language models.ACM Computing Surveys 57, 11 (2025), 1–41

    Zhang, H., Yu, P. S., and Zhang, J.A systematic survey of text summarization: From statistical methods to large language models.ACM Computing Surveys 57, 11 (2025), 1–41

    2025

  41. [49]

    Zhang, J., Chen, J., Y ang, M., Feng, A., Liang, S., Shao, J., and Ying, R.Dtgb: A comprehensive benchmark for dynamic text-attributed graphs.Advances in Neural Information Processing Systems 37(2024), 91405–91429

    2024

  42. [50]

    Zhang, Y., Jin, H., Meng, D., W ang, J., and Tan, J.A comprehensive survey on automatic text summarization with exploration of llm-based methods.Neuro- computing(2025), 131928

    2025

  43. [51]

    InThe Eleventh International Conference on Learning Representations(2023)

    Zhao, J., Qu, M., Li, C., Yan, H., Liu, Q., Li, R., Xie, X., and Tang, J.Learning on large-scale text-attributed graphs via variational inference. InThe Eleventh International Conference on Learning Representations(2023)

    2023

  44. [52]

    Zheng, X., Zhang, M., Chen, C., Nguyen, Q. V. H., Zhu, X., and Pan, S.Structure- free graph condensation: From large-scale graphs to condensed graph-free data. Advances in Neural Information Processing Systems 36(2023), 6026–6047

    2023

  45. [53]

    Output:Soft labelsPand labeled setsU,U GA,U GF

    Zhou, C., Du, J., Zhou, H., Chen, H., Huang, F., and Huang, X.Text-attributed Semi-Supervised Text-Attributed Graph Distillation Conference’17, July 2017, Washington, DC, USA Algorithm 1:Graph-Text Collaborative Encoding Input:TAGGwithA,X,Y, and initial labeled setV tr. Output...

    2017

  46. [54]

    {dataset_name }

    Zhu, Y., Wang, Y., Shi, H., and Tang, S.Efficient tuning and inference for large language models on textual graphs. InProceedings of the Thirty-Third International Joint Conference on Artificial Intelligence(2024), IJCAI ’24. A Complete Algorithm and Analysis This section prov...

    2024

  47. [55]

    Token Limit: Total output (including [[LABEL]], [[ SUMMARY]], spaces, punctuation) must NOT exceed {tok_lim} tokens

  48. [56]

    Output Format (EXACTLY as shown, NO extra lines/ words/symbols/line breaks/markdown): [[LABEL]] <selected category name> [[SUMMARY]] <single short paragraph summary>

  49. [57]

    - Output ONLY the category name (no additional text, punctuation, or explanations)

    [[LABEL]] Rule: - Select ONLY ONE category from the given category ({center_label1}). - Output ONLY the category name (no additional text, punctuation, or explanations)

  50. [58]

    "" Candidate 2:

    [[SUMMARY]] Rules (ALL must be satisfied): - Length: Max {tok_lim} tokens (after [[SUMMARY ]]:). - Content: - Integrates the first 3-5 core keywords naturally into the summary (avoid keyword stacking); incorporate other keywords only if they fit the context without exceeding t...

    2017

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.