Toward General and Robust LLM-enhanced Text-attributed Graph Learning

Bing Zhou; Guoren Wang; Rong-Hua Li; Xunkai Li; Zhenjun Li; Zihao Zhang

arxiv: 2504.02343 · v2 · submitted 2025-04-03 · 💻 cs.LG

Toward General and Robust LLM-enhanced Text-attributed Graph Learning

Zihao Zhang , Xunkai Li , Rong-Hua Li , Zhenjun Li , Bing Zhou , Guoren Wang This is my paper

Pith reviewed 2026-05-22 21:23 UTC · model grok-4.3

classification 💻 cs.LG

keywords LLM-enhanced graph learningText-attributed graphsGraph sparsity mitigationUnified frameworkText propagationPageRank node selectionEdge reconfigurationGraph neural networks

0 comments

The pith

UltraTAG unifies LLM and GNN methods for text-attributed graphs while its robust version UltraTAG-S uses text propagation, augmentation, PageRank selection, and edge reconfiguration to cut sparsity losses.

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

The paper seeks to create one pipeline that organizes the many existing ways of combining large language models with graph neural networks on graphs carrying text attributes. It then builds a concrete version called UltraTAG-S that adds LLM-driven text propagation and augmentation to fill missing node text, plus PageRank-guided node selection and edge changes to fix missing connections. If the approach holds, real-world graphs that suffer from incomplete text and links would yield better predictions without needing perfectly dense data. Experiments report gains of 2.12 percent in normal settings and 17.47 percent under sparsity, with the advantage widening as sparsity grows.

Core claim

UltraTAG supplies a unified comprehensive domain-adaptive framework that systematizes LLM-enhanced TAG learning and organizes prior methods. UltraTAG-S instantiates the framework by applying LLM-based text propagation and text augmentation against text sparsity together with LLM-augmented PageRank node selection and edge reconfiguration against edge sparsity, delivering 2.12 percent and 17.47 percent gains over baselines in ideal and sparse regimes respectively.

What carries the argument

The UltraTAG pipeline that organizes LLM-GNN interactions for TAGs, instantiated in UltraTAG-S through LLM text propagation, text augmentation, PageRank-based node selection, and edge reconfiguration to mitigate sparsity.

If this is right

TAG learning becomes feasible on the incomplete graphs common in practice rather than only on clean benchmark data.
Performance margins widen rather than shrink as the fraction of missing text and edges grows.
Existing LLM-GNN combinations can be placed inside one shared structure instead of remaining scattered.
New domain-specific variants can be added to the same pipeline without redesigning the overall flow.

Where Pith is reading between the lines

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

The same propagation-plus-reconfiguration steps could be tried on graphs that carry images or other modalities in addition to text.
Scaling the method to graphs with millions of nodes would test whether the LLM calls remain practical.
The observed robustness to increasing sparsity suggests the method may also help when graphs change over time rather than stay static.

Load-bearing premise

LLM-generated text, node selections, and edge changes reliably reduce sparsity without adding hallucinations or domain-specific biases that cancel the reported gains.

What would settle it

Running UltraTAG-S on a real sparse TAG dataset and observing either no accuracy lift over strong baselines or a measurable rise in biased or hallucinated predictions.

Figures

Figures reproduced from arXiv: 2504.02343 by Bing Zhou, Guoren Wang, Rong-Hua Li, Xunkai Li, Zhenjun Li, Zihao Zhang.

**Figure 1.** Figure 1: Performance of diffirent LLM-enhanced TAG learning methods in sparse scenarios. The horizontal axis represents the sparsity ratio of nodes and edges, while the vertical axis denotes classification accuracy. UltraTAG-S has the optimal robustness. In recent years, the advancements in large language models (LLMs) (Brown et al., 2020) have driven the evolution of graph ML, particularly in Text-Attributed Gra… view at source ↗

**Figure 2.** Figure 2: Overview of UltraTAG for LLM-Enhanced Text-Attributed Graph Learning, which is composed of three independent modules. Solution 1: UltraTAG: A Unified Pipeline toward General and Robust LLM-enhanced TAG Learning. To address Limitation 1, we propose UltraTAG, as detailed in Sec. 3. UltraTAG is composed of three key modules: Data Augmentation, Text Encoder, and Training Mechanism, as illustrated in [PITH_FU… view at source ↗

**Figure 3.** Figure 3: Overview of UltraTAG-S for LLM-Enhanced Text-Attributed Graph Learning in Sparse Scenarios. 3.4. Training Mechanism After obtaining the nodes’ textual representations from LM or LLM H = {h1, h2, h3, ..., hN } and adjacency matrix A, input of them into a GNN will yield the final prediction. In terms of training mechanisms, we can use a simple GNN module, or combine GNN with LM for joint training. Simple GNN… view at source ↗

**Figure 4.** Figure 4: Robustness Comparison in Sparse Scenarios. The horizontal coordinate represents the sparse ratio of nodes and edges, and the vertical coordinate represents the accuracy of the node classification task. We conduct a comprehensive evaluation of UltraTAG-S by comparing with GNN-only, LM-only and LLM-GNN methods, as the results in [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

**Figure 5.** Figure 5: Robustness Comparison among All Datasets in Sparse Ratio of 20%, 50% and 80%. In order to simulate the challenges of the sparse scene, we randomly delete the texts and edges of nodes in a ratio of 20%, 50%, and 80% without considering data noise or additional constraints. As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

**Figure 6.** Figure 6: Ablation Study on PubMed and CiteSeer. The x-axis represents the modules in the ablation study, where ’w/o TA’, ’w/o SA’, ’w/o SL’ denote the removal of Text Augmentation module, Structure Augmentation module and Structure Learning module, respectively. The y-axis represents accuracy in different ratios. 5.4. Complexity Analysis and Hyperparameter Setting The computational complexity of our proposed method… view at source ↗

read the original abstract

Recent advancements in Large Language Models (LLMs) and the proliferation of Text-Attributed Graphs (TAGs) across various domains have positioned LLM-enhanced TAG learning as a critical research area. By utilizing rich graph descriptions, this paradigm leverages LLMs to generate high-quality embeddings, thereby enhancing the representational capacity of Graph Neural Networks (GNNs). However, the field faces significant challenges: (1) the absence of a unified framework to systematize the diverse optimization perspectives arising from the complex interactions between LLMs and GNNs, and (2) the lack of a robust method capable of handling real-world TAGs, which often suffer from texts and edge sparsity, leading to suboptimal performance. To address these challenges, we propose UltraTAG, a unified pipeline for LLM-enhanced TAG learning. UltraTAG provides a unified comprehensive and domain-adaptive framework that not only organizes existing methodologies but also paves the way for future advancements in the field. Building on this framework, we propose UltraTAG-S, a robust instantiation of UltraTAG designed to tackle the inherent sparsity issues in real-world TAGs. UltraTAG-S employs LLM-based text propagation and text augmentation to mitigate text sparsity, while leveraging LLM-augmented node selection techniques based on PageRank and edge reconfiguration strategies to address edge sparsity. Our extensive experiments demonstrate that UltraTAG-S significantly outperforms existing baselines, achieving improvements of 2.12\% and 17.47\% in ideal and sparse settings, respectively. Moreover, as the data sparsity ratio increases, the performance improvement of UltraTAG-S also rises, which underscores the effectiveness and robustness of UltraTAG-S.

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.

Desk Editor's Note private letter to a colleague

UltraTAG names a unified LLM-GNN pipeline for text graphs and adds sparsity fixes via text propagation plus PageRank edge tweaks, but the reported gains rest on unexamined LLM output quality.

read the letter

The paper's core move is to give a name and structure to the mix of LLMs and GNNs on text-attributed graphs, then build UltraTAG-S on top of it to handle missing text and edges through LLM text propagation, augmentation, PageRank node picking, and edge reconfiguration. That framing pulls together scattered prior work and points to a concrete way to make the combination more usable on incomplete real graphs. The experiments show small lifts in clean settings and larger ones as sparsity grows, which lines up with the robustness story they want to tell. If the augmentations hold up, this could give practitioners a reusable template instead of reinventing combinations each time. The unification angle is the part that feels most organized and potentially reusable for others mapping the space. The sparsity techniques are straightforward extensions of existing LLM and graph ideas, but packaging them together under one pipeline is the practical step. The main gap is verification that the LLM-generated text and edges stay faithful. The abstract gives no fidelity metrics, human checks, or ablations on generation quality, so the bigger gains in sparse regimes could trace to particular LLM outputs rather than the pipeline itself. Without those controls, it's hard to separate real robustness from lucky or biased augmentations. Dataset and baseline details are also thin in the summary, which makes the percentage improvements harder to weigh. This is aimed at graph ML researchers who already work with text attributes and run into incomplete data. It is not a foundational shift, but the organization and the sparsity focus address a common pain point. The work shows clear thinking on the problem setup even if the empirical support needs tightening. I would send it to peer review so the experimental claims can be checked properly rather than desk-rejecting it outright.

Referee Report

2 major / 0 minor

Summary. The manuscript proposes UltraTAG, a unified comprehensive and domain-adaptive framework for LLM-enhanced text-attributed graph (TAG) learning that organizes existing methodologies, and UltraTAG-S, a robust instantiation addressing text and edge sparsity in real-world TAGs via LLM-based text propagation and augmentation, PageRank-based node selection, and edge reconfiguration strategies. Experiments are reported to show UltraTAG-S outperforming baselines by 2.12% in ideal settings and 17.47% in sparse settings, with gains increasing as the sparsity ratio rises.

Significance. If the empirical claims are substantiated, the unified framework could help systematize diverse LLM-GNN optimization perspectives in TAG learning, while UltraTAG-S could provide a practical method for improving performance and robustness under sparsity, a common issue in real-world TAG applications.

major comments (2)

[Abstract] Abstract: The central performance claims of 2.12% and 17.47% improvements (and the increasing-gains-with-sparsity observation) are stated without any information on datasets, baselines, statistical tests, ablation studies, or controls for LLM output variability, leaving the robustness claim without verifiable support.
[Abstract] Abstract: The robustness narrative in sparse regimes rests on the unverified assumption that LLM-based text propagation, augmentation, PageRank node selection, and edge reconfiguration mitigate sparsity without net negative effects from hallucinations or domain biases; no fidelity metrics, human validation of augmented content, or generation-quality ablations are described to support this load-bearing assumption.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment point by point below, providing the strongest honest defense of the manuscript while making revisions where the comments identify clear gaps in the abstract or supporting discussion.

read point-by-point responses

Referee: [Abstract] Abstract: The central performance claims of 2.12% and 17.47% improvements (and the increasing-gains-with-sparsity observation) are stated without any information on datasets, baselines, statistical tests, ablation studies, or controls for LLM output variability, leaving the robustness claim without verifiable support.

Authors: The abstract is intentionally concise as a high-level summary. All requested details—datasets (Cora, CiteSeer, PubMed, and others), baselines (including GNN and LLM-GNN methods), statistical reporting (means and standard deviations over multiple runs), ablation studies, and controls for LLM variability—are fully provided in Sections 4 and 5 of the manuscript. To improve verifiability at the abstract level, we have revised the abstract to briefly note the evaluation on standard TAG benchmarks with results averaged over runs. revision: yes
Referee: [Abstract] Abstract: The robustness narrative in sparse regimes rests on the unverified assumption that LLM-based text propagation, augmentation, PageRank node selection, and edge reconfiguration mitigate sparsity without net negative effects from hallucinations or domain biases; no fidelity metrics, human validation of augmented content, or generation-quality ablations are described to support this load-bearing assumption.

Authors: The increasing performance gains as sparsity rises provide direct empirical support that the net effect is positive. We acknowledge that dedicated fidelity metrics and human validation of LLM outputs are not present in the original submission. We have added a qualitative analysis subsection and discussion of potential hallucination risks in the revised experiments section, but comprehensive quantitative fidelity ablations would require new experiments beyond the current scope. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical framework with experimental validation

full rationale

The paper proposes UltraTAG as a unified pipeline and UltraTAG-S as a sparsity-handling instantiation, with all central claims resting on experimental comparisons (2.12% and 17.47% gains) rather than any derivation, equation, or first-principles result. No self-definitional steps, fitted inputs renamed as predictions, or load-bearing self-citations appear in the provided text; the work is self-contained as an empirical engineering contribution whose validity is externally falsifiable via replication on the reported datasets.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claims rest on standard machine-learning assumptions about LLM embedding quality and the effectiveness of the proposed sparsity mitigations; no new physical entities or mathematical axioms are introduced beyond domain conventions.

axioms (2)

domain assumption LLMs can generate high-quality embeddings from graph descriptions that improve GNN representational capacity
Explicitly stated in the abstract as the basis for leveraging LLMs.
ad hoc to paper LLM-based text propagation, augmentation, PageRank node selection, and edge reconfiguration mitigate sparsity without net negative effects
This premise is required for the robustness claim and performance gains to hold.

pith-pipeline@v0.9.0 · 5835 in / 1496 out tokens · 29333 ms · 2026-05-22T21:23:17.119840+00:00 · methodology

discussion (0)

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

When LLM Agents Meet Graph Optimization: An Automated Data Quality Improvement Approach
cs.LG 2025-10 unverdicted novelty 7.0

LAGA is a unified multi-agent LLM framework that automates comprehensive quality optimization for text-attributed graphs by running detection, planning, action, and evaluation agents in a closed loop.

Reference graph

Works this paper leans on

21 extracted references · 21 canonical work pages · cited by 1 Pith paper · 5 internal anchors

[1]

Language Models are Few-Shot Learners

URL https:// arxiv.org/abs/2005.14165. Chen, M., Wei, Z., Huang, Z., Ding, B., and Li, Y . Simple and deep graph convolutional networks. In International Conference on Machine Learning, ICML,

work page internal anchor Pith review Pith/arXiv arXiv 2005
[2]

Chien, E., Chang, W.-C., Hsieh, C.-J., Yu, H.-F., Zhang, J., Milenkovic, O., and Dhillon, I

URL https://arxiv.org/abs/2310.04668. Chien, E., Chang, W.-C., Hsieh, C.-J., Yu, H.-F., Zhang, J., Milenkovic, O., and Dhillon, I. S. Node feature extrac- tion by self-supervised multi-scale neighborhood predic- tion,

work page arXiv
[3]

BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding

URL https://arxiv. org/abs/1810.04805. Duan, K., Liu, Q., Chua, T.-S., Yan, S., Ooi, W. T., Xie, Q., and He, J. Simteg: A frustratingly simple approach improves textual graph learning,

work page internal anchor Pith review Pith/arXiv arXiv
[4]

Giles, C

URL https: //arxiv.org/abs/2308.02565. Giles, C. L., Bollacker, K. D., and Lawrence, S. Citeseer: An automatic citation indexing system. In Proceedings of the third ACM conference on Digital libraries,

work page arXiv
[5]

Hamilton, W., Ying, Z., and Leskovec, J

URL https: //arxiv.org/abs/2302.12977. Hamilton, W., Ying, Z., and Leskovec, J. Inductive repre- sentation learning on large graphs. Advances in Neural Information Processing Systems, NeurIPS,

work page arXiv
[6]

DeBERTa: Decoding-enhanced BERT with Disentangled Attention

URL https://arxiv.org/abs/2006.03654. He, X., Bresson, X., Laurent, T., Perold, A., LeCun, Y ., and Hooi, B. Harnessing explanations: Llm-to-lm inter- preter for enhanced text-attributed graph representation learning, 2024a. URL https://arxiv.org/abs/ 2305.19523. He, Y ., Sui, Y ., He, X., and Hooi, B. Unigraph: Learning a unified cross-domain foundation ...

work page internal anchor Pith review Pith/arXiv arXiv 2006
[7]

URL https://arxiv.org/abs/2402.12984. Kipf, T. N. and Welling, M. Semi-supervised classifica- tion with graph convolutional networks. In International Conference on Learning Representations, ICLR,

work page arXiv
[8]

Lightdic: A simple yet effective approach for large-scale digraph representation learning

Li, X., Liao, M., Wu, Z., Su, D., Zhang, W., Li, R.-H., and Wang, G. Lightdic: A simple yet effective approach for large-scale digraph representation learning. arXiv preprint arXiv:2401.11772, 2024a. Li, Z., Li, R.-H., Liao, M., Jin, F., and Wang, G. Privacy- preserving graph embedding based on local differential privacy, 2024b. URL https://arxiv.org/abs/...

work page arXiv
[9]

RoBERTa: A Robustly Optimized BERT Pretraining Approach

URL https://arxiv.org/abs/1907.11692. Mernyei, P. and Cangea, C. Wiki-cs: A wikipedia-based benchmark for graph neural networks. arXiv preprint arXiv:2007.02901,

work page internal anchor Pith review Pith/arXiv arXiv 1907
[10]

Distributed Representations of Words and Phrases and their Compositionality

URL https://arxiv. org/abs/1310.4546. Ni, J., Li, J., and McAuley, J. Justifying recommendations using distantly-labeled reviews and fine-grained aspects. In Proc. of EMNLP,

work page internal anchor Pith review Pith/arXiv arXiv
[11]

Rossi, E., Kenlay, H., Gorinova, M

URL https://arxiv.org/abs/ 2402.12022. Rossi, E., Kenlay, H., Gorinova, M. I., Chamberlain, B. P., Dong, X., and Bronstein, M. On the unrea- sonable effectiveness of feature propagation in learn- ing on graphs with missing node features,

work page arXiv
[12]

9 Rethinking Graph Structure Learning in the Era of LLMs Sen, P., Namata, G., Bilgic, M., Getoor, L., Galligher, B., and Eliassi-Rad, T

URL https://arxiv.org/abs/2111.12128. 9 Rethinking Graph Structure Learning in the Era of LLMs Sen, P., Namata, G., Bilgic, M., Getoor, L., Galligher, B., and Eliassi-Rad, T. Collective classification in network data. AI magazine,

work page arXiv
[13]

and Sachan, M

Singh, G. and Sachan, M. Multi-layer perceptron (mlp) neural network technique for offline handwrit- ten gurmukhi character recognition. In 2014 IEEE International Conference on Computational Intelligence and Computing Research, pp. 1–5,

work page 2014
[14]

Veliˇckovi´c, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., and Bengio, Y

doi: 10.1109/ ICCIC.2014.7238334. Veliˇckovi´c, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., and Bengio, Y . Graph attention networks. In International Conference on Learning Representations, ICLR,

work page arXiv 2014
[15]

URL https://arxiv.org/ abs/2406.12608. Wen, Z. and Fang, Y . Augmenting low-resource text classi- fication with graph-grounded pre-training and prompt- ing. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’23, pp. 506–516. ACM, July

work page arXiv
[16]

URL http: //dx.doi.org/10.1145/3539618.3591641

doi: 10.1145/3539618.3591641. URL http: //dx.doi.org/10.1145/3539618.3591641. Yan, H., Li, C., Long, R., Yan, C., Zhao, J., Zhuang, W., Yin, J., Zhang, P., Han, W., Sun, H., et al. A comprehen- sive study on text-attributed graphs: Benchmarking and rethinking. In Proc. of NeurIPS,

work page doi:10.1145/3539618.3591641
[17]

URL http://dx.doi

doi: 10.1145/3534678.3539121. URL http://dx.doi. org/10.1145/3534678.3539121. 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,

work page doi:10.1145/3534678.3539121
[18]

Zhu, Y ., Wang, Y ., Shi, H., and Tang, S

URL https: //arxiv.org/abs/2210.14709. Zhu, Y ., Wang, Y ., Shi, H., and Tang, S. Efficient tun- ing and inference for large language models on tex- tual graphs,

work page arXiv
[19]

10 Rethinking Graph Structure Learning in the Era of LLMs A

URL https://arxiv.org/abs/ 2401.15569. 10 Rethinking Graph Structure Learning in the Era of LLMs A. Datasets This section provides a detailed introduction to the datasets used in the main content. The statistics of the TAG datasets we use is as shown in Table

work page arXiv
[20]

In this dataset, nodes represent electronics-related products, and edges sig- nify frequent co-purchases or co-views between products

dataset is derived from the Amazon-Electronics dataset (Ni et al., 2019). In this dataset, nodes represent electronics-related products, and edges sig- nify frequent co-purchases or co-views between products. Each node is labeled based on a three-level classification scheme for electronics products. User reviews serve as the textual attributes for the nod...

work page 2019
[21]

It consists of multiple layers of neurons, where each layer is fully connected to the previous one

is a simple feed-forward neural network model, commonly used for baseline classifi- cation tasks. It consists of multiple layers of neurons, where each layer is fully connected to the previous one. The model is trained via backpropagation, with the final output layer producing predictions. GCN (Kipf & Welling, 2017)is a graph-based neural net- work model ...

work page 2017

[1] [1]

Language Models are Few-Shot Learners

URL https:// arxiv.org/abs/2005.14165. Chen, M., Wei, Z., Huang, Z., Ding, B., and Li, Y . Simple and deep graph convolutional networks. In International Conference on Machine Learning, ICML,

work page internal anchor Pith review Pith/arXiv arXiv 2005

[2] [2]

Chien, E., Chang, W.-C., Hsieh, C.-J., Yu, H.-F., Zhang, J., Milenkovic, O., and Dhillon, I

URL https://arxiv.org/abs/2310.04668. Chien, E., Chang, W.-C., Hsieh, C.-J., Yu, H.-F., Zhang, J., Milenkovic, O., and Dhillon, I. S. Node feature extrac- tion by self-supervised multi-scale neighborhood predic- tion,

work page arXiv

[3] [3]

BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding

URL https://arxiv. org/abs/1810.04805. Duan, K., Liu, Q., Chua, T.-S., Yan, S., Ooi, W. T., Xie, Q., and He, J. Simteg: A frustratingly simple approach improves textual graph learning,

work page internal anchor Pith review Pith/arXiv arXiv

[4] [4]

Giles, C

URL https: //arxiv.org/abs/2308.02565. Giles, C. L., Bollacker, K. D., and Lawrence, S. Citeseer: An automatic citation indexing system. In Proceedings of the third ACM conference on Digital libraries,

work page arXiv

[5] [5]

Hamilton, W., Ying, Z., and Leskovec, J

URL https: //arxiv.org/abs/2302.12977. Hamilton, W., Ying, Z., and Leskovec, J. Inductive repre- sentation learning on large graphs. Advances in Neural Information Processing Systems, NeurIPS,

work page arXiv

[6] [6]

DeBERTa: Decoding-enhanced BERT with Disentangled Attention

URL https://arxiv.org/abs/2006.03654. He, X., Bresson, X., Laurent, T., Perold, A., LeCun, Y ., and Hooi, B. Harnessing explanations: Llm-to-lm inter- preter for enhanced text-attributed graph representation learning, 2024a. URL https://arxiv.org/abs/ 2305.19523. He, Y ., Sui, Y ., He, X., and Hooi, B. Unigraph: Learning a unified cross-domain foundation ...

work page internal anchor Pith review Pith/arXiv arXiv 2006

[7] [7]

URL https://arxiv.org/abs/2402.12984. Kipf, T. N. and Welling, M. Semi-supervised classifica- tion with graph convolutional networks. In International Conference on Learning Representations, ICLR,

work page arXiv

[8] [8]

Lightdic: A simple yet effective approach for large-scale digraph representation learning

Li, X., Liao, M., Wu, Z., Su, D., Zhang, W., Li, R.-H., and Wang, G. Lightdic: A simple yet effective approach for large-scale digraph representation learning. arXiv preprint arXiv:2401.11772, 2024a. Li, Z., Li, R.-H., Liao, M., Jin, F., and Wang, G. Privacy- preserving graph embedding based on local differential privacy, 2024b. URL https://arxiv.org/abs/...

work page arXiv

[9] [9]

RoBERTa: A Robustly Optimized BERT Pretraining Approach

URL https://arxiv.org/abs/1907.11692. Mernyei, P. and Cangea, C. Wiki-cs: A wikipedia-based benchmark for graph neural networks. arXiv preprint arXiv:2007.02901,

work page internal anchor Pith review Pith/arXiv arXiv 1907

[10] [10]

Distributed Representations of Words and Phrases and their Compositionality

URL https://arxiv. org/abs/1310.4546. Ni, J., Li, J., and McAuley, J. Justifying recommendations using distantly-labeled reviews and fine-grained aspects. In Proc. of EMNLP,

work page internal anchor Pith review Pith/arXiv arXiv

[11] [11]

Rossi, E., Kenlay, H., Gorinova, M

URL https://arxiv.org/abs/ 2402.12022. Rossi, E., Kenlay, H., Gorinova, M. I., Chamberlain, B. P., Dong, X., and Bronstein, M. On the unrea- sonable effectiveness of feature propagation in learn- ing on graphs with missing node features,

work page arXiv

[12] [12]

9 Rethinking Graph Structure Learning in the Era of LLMs Sen, P., Namata, G., Bilgic, M., Getoor, L., Galligher, B., and Eliassi-Rad, T

URL https://arxiv.org/abs/2111.12128. 9 Rethinking Graph Structure Learning in the Era of LLMs Sen, P., Namata, G., Bilgic, M., Getoor, L., Galligher, B., and Eliassi-Rad, T. Collective classification in network data. AI magazine,

work page arXiv

[13] [13]

and Sachan, M

Singh, G. and Sachan, M. Multi-layer perceptron (mlp) neural network technique for offline handwrit- ten gurmukhi character recognition. In 2014 IEEE International Conference on Computational Intelligence and Computing Research, pp. 1–5,

work page 2014

[14] [14]

Veliˇckovi´c, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., and Bengio, Y

doi: 10.1109/ ICCIC.2014.7238334. Veliˇckovi´c, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., and Bengio, Y . Graph attention networks. In International Conference on Learning Representations, ICLR,

work page arXiv 2014

[15] [15]

URL https://arxiv.org/ abs/2406.12608. Wen, Z. and Fang, Y . Augmenting low-resource text classi- fication with graph-grounded pre-training and prompt- ing. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’23, pp. 506–516. ACM, July

work page arXiv

[16] [16]

URL http: //dx.doi.org/10.1145/3539618.3591641

doi: 10.1145/3539618.3591641. URL http: //dx.doi.org/10.1145/3539618.3591641. Yan, H., Li, C., Long, R., Yan, C., Zhao, J., Zhuang, W., Yin, J., Zhang, P., Han, W., Sun, H., et al. A comprehen- sive study on text-attributed graphs: Benchmarking and rethinking. In Proc. of NeurIPS,

work page doi:10.1145/3539618.3591641

[17] [17]

URL http://dx.doi

doi: 10.1145/3534678.3539121. URL http://dx.doi. org/10.1145/3534678.3539121. 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,

work page doi:10.1145/3534678.3539121

[18] [18]

Zhu, Y ., Wang, Y ., Shi, H., and Tang, S

URL https: //arxiv.org/abs/2210.14709. Zhu, Y ., Wang, Y ., Shi, H., and Tang, S. Efficient tun- ing and inference for large language models on tex- tual graphs,

work page arXiv

[19] [19]

10 Rethinking Graph Structure Learning in the Era of LLMs A

URL https://arxiv.org/abs/ 2401.15569. 10 Rethinking Graph Structure Learning in the Era of LLMs A. Datasets This section provides a detailed introduction to the datasets used in the main content. The statistics of the TAG datasets we use is as shown in Table

work page arXiv

[20] [20]

In this dataset, nodes represent electronics-related products, and edges sig- nify frequent co-purchases or co-views between products

dataset is derived from the Amazon-Electronics dataset (Ni et al., 2019). In this dataset, nodes represent electronics-related products, and edges sig- nify frequent co-purchases or co-views between products. Each node is labeled based on a three-level classification scheme for electronics products. User reviews serve as the textual attributes for the nod...

work page 2019

[21] [21]

It consists of multiple layers of neurons, where each layer is fully connected to the previous one

is a simple feed-forward neural network model, commonly used for baseline classifi- cation tasks. It consists of multiple layers of neurons, where each layer is fully connected to the previous one. The model is trained via backpropagation, with the final output layer producing predictions. GCN (Kipf & Welling, 2017)is a graph-based neural net- work model ...

work page 2017