Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

LLM4GNAS: A Large Language Model Based Toolkit for Graph Neural Architecture Search

T0 review · 5 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read Prompt edits alone let one toolkit search GNN architectures for new graph tasks.

desk verdict A useful toolkit idea, undermined by best-of-three reporting and an overbroad 'first work' claim; the empirical case is not currently credible. read the letter →

arxiv 2502.10459 v1 pith:T2I6MFO4 submitted 2025-02-12 cs.LG cs.AI

Yang Gao , Hong Yang , Yizhi Chen , Junxian Wu , Peng Zhang , Haishuai Wang This is my paper · ORCID
classification cs.LGcs.AI
keywords graphneuralarchitecturesearchlargelanguagemodelsAutoMLnetworkspromptengineeringheterogeneousgraphshyperparameteroptimizationnodeaugmentation
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

The paper presents LLM4GNAS, a toolkit that puts a large language model in charge of graph neural architecture search. The authors aim to show that one code base can handle feature augmentation, architecture search, and hyperparameter optimization across homogeneous and heterogeneous graphs, and that moving to a new search space requires only editing prompts rather than rewriting search code. They claim this is the first combination of LLMs with graph neural architecture search, and that experiments on node classification, link prediction, and graph classification show the toolkit matching or beating existing GNAS methods. The reported results include top accuracies on ACM (92.97%), DBLP (94.41%), and the NS link-prediction set (99.88%).

What carries the argument

The load-bearing mechanism is the LLM-as-controller loop: a prompt describing candidate operations, connections, and search strategy makes the LLM propose GNN architectures; those architectures are trained and scored; the score is appended to the prompt for the next round. A second mechanism, prompt-based extensibility, lets a user add a new search space by subclassing the SearchSpaceBase class and filling in operation, connection, and example prompts, then registering the class in a model factory—so no search-code rewrite is needed.

What would settle it

Run every baseline under the same best-of-three selection rule and validation-based model choice on the homogeneous graph benchmarks; if AutoGEL or GraphNAS still beats LLM4GNAS on Cora, Citeseer, Power, and PROTEINS under that protocol, the paper's claim of outperforming existing GNAS methods is falsified.

Watch

Extended reading notes

Core claim

The central claim is that an LLM can act as the controller of a complete graph architecture search pipeline, and that prompt text—not code—is the interface for adapting that pipeline to new search spaces. The search is formalized as $m^* = \arg\max_{m \in M(\Omega(P))} A(m(G))$, where $\Omega(P)$ is the set of architectures the LLM generates from prompt $P$ and $A$ is the validation metric. Each iteration feeds the validation accuracy back into a reward prompt, so the LLM converges on better architectures. The paper reports strong results on heterogeneous graphs and competitive results on homogeneous graphs, plus an ablation showing that LLM-based node augmentation improves GCN, GAT, and GCNII on PubMed, Citeseer, and Arxiv.

Load-bearing premise

The headline comparisons treat LLM4GNAS's best result over three runs as the reported number while baselines are reported as mean±std, and if the baselines were given the same best-of-N selection, the claimed outperformance could shrink or disappear.

Editorial extensions

If this is right

  • If the toolkit's claims hold, researchers can switch a GNAS pipeline to a new graph domain by editing a prompt instead of modifying search code.
  • The iterative LLM feedback loop can find strong architectures in 15 rounds with 10 candidates per round, suggesting low search cost relative to many GNAS methods.
  • The same toolkit covers node classification, link prediction, and graph classification on both homogeneous and heterogeneous datasets.
  • The ablation suggests LLM-generated node features can substantially raise accuracy on text-attributed graphs, with GCNII+NA jumping from 69.85% to 81.57% on Arxiv.
  • The choice of LLM matters: the paper shows GPT-4 outperforming GPT-3, LLAMA-2, and GLM-3 across all tested datasets.

Reading between the lines

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

  • The 'first work combining LLMs and GNAS' claim is undercut by the paper's own related-work section, which cites GNAS-LLM and GHGNAS as prior LLM-based GNAS methods; the contribution is better read as an integrated, extensible toolkit rather than the first use of LLMs in GNAS.
  • Because the paper reports best-of-three results for LLM4GNAS against mean±std baselines, the headline performance gaps may shrink under a uniform evaluation protocol; a follow-up should rerun baselines with the same best-of-N selection.
  • The prompt-as-interface design suggests a testable extension: if prompts are the only adaptation surface, then an automated prompt optimizer or meta-learned prompt could remove the last manual step.
  • The node-augmentation ablation points to a broader pattern: LLM-generated text features may matter more than architecture choices on text-attributed graphs, so future LLM-based GNAS systems should isolate those two effects.
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

5 major / 6 minor

Summary. The paper presents LLM4GNAS, a toolkit that uses an LLM (GPT-4 by default) as a controller to automate graph neural architecture search (GNAS). The toolkit combines three components: LLM-based node feature augmentation, LLM-based GNAS, and LLM-based hyperparameter optimization. The authors claim that new search spaces can be adapted by editing prompts, and they report experiments on homogeneous and heterogeneous graphs for node classification, link prediction, and graph classification, together with ablations on LLM choice and a case study on a new search space. The central empirical claim is that LLM4GNAS outperforms existing GNAS methods.

Significance. If the claims were fully supported, LLM4GNAS would be a practically valuable toolkit: a single LLM-driven system that adapts to heterogeneous and homogeneous graphs by prompt editing alone would reduce engineering overhead in GNAS. The paper covers a broad range of tasks and datasets and includes ablations on the main modules and on different LLMs. However, the significance is currently conditional because the empirical evaluation protocol is not apples-to-apples, the strongest headline claim is contradicted by the paper's own tables, the novelty claim is internally inconsistent, and the toolkit code is not released despite being a central part of the contribution. The core idea is plausible and worth pursuing, but the current manuscript does not substantiate the headline.

major comments (5)
  1. [Section 4.1.2, Tables 2 and 3] Section 4.1.2 states: 'we repeat our method three times and show the best results w.r.t. accuracy on a validation dataset.' Baselines are reported as mean±std in Tables 2 and 3. Selecting the best of three runs is an order-statistic estimator and is biased upward relative to the mean of the same runs. On many comparisons the reported gaps are smaller than the reported standard deviations, e.g. Table 2 Cora (LLM4GNAS 89.2±0.63 vs AutoGEL 89.89±0.62) and Table 3 DBLP (LLM4GNAS 94.41±0.35 vs DiffMG 94.45±0.15). The headline 'outperforms' claim is therefore not supported by the evidence as presented: either report the mean and standard deviation for LLM4GNAS, or apply the same best-of-N selection to the baselines.
  2. [Abstract and Section 4.2, Table 2] The abstract states that LLM4GNAS 'outperforms existing GNAS methods on tasks involving both homogeneous and heterogeneous graphs.' Table 2 itself shows AutoGEL achieving higher accuracy on Cora, Citeseer, Power, IMDB-B, and PROTEINS, and GraphNAS achieving higher accuracy on PubMed. Thus the data in the paper contradict the unqualified 'outperforms existing GNAS methods' claim even before the best-of-three issue is considered. The claim should be restricted to the settings where LLM4GNAS actually leads (link prediction, most heterogeneous datasets, and the new-search-space case study) or additional evidence should be provided.
  3. [Section 1 contributions and Section 5.1-5.2] The contribution list says 'this represents the first work that combines LLMs and graph neural architecture search.' This is contradicted by the paper's own references: Ref. [37] (GNAS-LLM) and Ref. [9] (GHGNAS) are described in Section 5.1 as works that 'explore the application of LLMs to enhance the GNAS search process.' The novelty claim must be rewritten, for example as the first LLM-based GNAS toolkit that integrates feature augmentation, search, and hyperparameter optimization, and it must clearly state what is added relative to GNAS-LLM and GHGNAS.
  4. [Table 1 vs Section 4.4.4] Table 1 reports the ogbn-products dataset as having 54,025 nodes and 74,420 edges, while Section 4.4.4 states that ogbn-products contains 2,449,029 nodes and 61,859,140 edges. These two descriptions cannot both be correct. Since Section 4.4.4 is the only evaluation on a truly large graph and the paper claims scalability, the dataset statistics and the experimental setting need to be reconciled and described precisely.
  5. [Sections 1, 3, and 4] The paper is a toolkit paper whose key promises are ease of use, adaptability via prompts, and extensibility. However, no code, no prompt templates, and no public repository are provided, and the prompt design is only described at a high level. Without these, the central claim that 'adaptation to new search spaces requires only prompt modification' cannot be independently verified, and the experiments cannot be reproduced even in principle. The paper should release the code and the full prompt templates (or place them in an appendix) as a condition of the claimed toolkit contribution.
minor comments (6)
  1. [Throughout] There are numerous typographical errors that should be corrected: 'PROTENS' in Table 1, 'dor GraphNAS' in Section 4.1.2, 'LLM4GANS' in Listing 2, and 'GNAS-LLM obtains the best GNN' in Section 2.3, which should refer to LLM4GNAS.
  2. [Section 2.3, Equation (1)] Equation (1) uses the notation M(Ω(𝑃)) both as a subscript and as the search space generated by the LLM; this is confusing and appears to conflate the search space with the set of generated architectures. Please rewrite the equation and define all symbols explicitly.
  3. [Section 4.2, Figures 2 and 3] The text says LLM4GNAS 'consistently surpasses' the other methods in Figure 2, but Table 2 shows several datasets where LLM4GNAS is not the best. Also, the caption says 'test accuracy' while Section 4.1.2 says best results are selected on the validation set; clarify whether the figure reports validation or test performance.
  4. [Section 4.4.3, Table 6] Table 6 reports 'three-time architecture search w.r.t. Accuracy' and shows LLM4GNAS with a standard deviation of 0.00 on Cora. Given three runs, a zero standard deviation is surprising and should be explained; more generally, the caption should state how the three runs are aggregated.
  5. [Section 4.1.2] The GPT-4 version (V20230314) is given, but the temperature and the prompt content are not. To make the LLM-based search reproducible, include at least the exact prompt templates or a pointer to them.
  6. [Section 4.3, Table 3] The most closely related LLM-based GNAS baseline, GHGNAS (Ref. [9]), is not included in the heterogeneous-graph comparisons, even though the approach builds on it. Adding this baseline would make the comparison more informative.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: LLM4GNAS is an empirical toolkit paper, its headline performance claim is checked against external baselines, and self-citations to prior GNAS-LLM/GHGNAS work are contextual rather than load-bearing.

full rationale

LLM4GNAS does not contain a derivation chain in which a predicted quantity is equivalent to an input by construction. The central claim is empirical: the toolkit's generated GNNs are compared with GCN, GAT, GraphNAS, AutoGEL, DiffMG, and HGNAS++ under experimental settings taken from AutoGEL and HGNAS++. The self-citations (GNAS-LLM [37], GHGNAS [9], HGNAS++ [17]) appear as related work, as baselines, or as internal search components; none is invoked as an unverified premise that forces the reported accuracies. Equation (1) merely defines the search objective m* = argmax A(m(G)) and is not used to derive any result. The protocol in Section 4.1.2 of repeating the method three times and reporting the best validation accuracy while baselines are reported as mean±std is a statistical fairness concern: it can inflate relative performance, but the reported test accuracy is not equal to the comparison by definition, so it is not a circular reduction. The 'first work' novelty claim is undercut by the paper's own citations of GNAS-LLM and GHGNAS, but that is a correctness/novelty issue, not a circularity issue. No circular step can be quoted and exhibited, so the score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper makes no mathematical derivation, so the ledger records hand-chosen experimental settings and domain assumptions. The central results depend on the search budget, the restricted candidate operation set, the specific GPT-4 version, and the untested assumption that validation accuracy reliably guides prompt feedback. No new physical or conceptual entities are introduced.

free parameters (3)
  • Search budget (15 iterations, 10 candidates per iteration) = 150 architecture evaluations
    Chosen by hand in Section 4.1.2; no sensitivity analysis is given, and the comparison against baselines depends on this budget.
  • Candidate operation space (nine operations, one connection type) = 9 ops / 1 connection
    Section 4.1.2 states 'we only use one candidate operation connections and nine candidate operations'; this restricts the search space without justification.
  • GPT-4 model version = V20230314 (gpt-4-0314)
    Section 4.1.2; LLM generations and therefore the search results depend on the exact model snapshot.
assumptions (4)
  • domain assumption Validation accuracy of generated architectures is a reliable reward signal for the LLM's iterative prompt feedback.
    Used in Section 2.3; if validation accuracy is noisy or mis-specified, the iterative refinement loop cannot be expected to converge to good architectures.
  • ad hoc to paper Editing prompts is sufficient to adapt the search space to a new task or graph type without changing search algorithms.
    Stated in Sections 2.3 and 3.2 as the main extensibility mechanism; only one case study on NAS-Bench-Graph supports it.
  • domain assumption GPT-4 with temperature 0 yields sufficiently stable and reproducible architecture proposals across API calls.
    Section 4.1.2 sets temperature to 0 for reproducibility; API-level nondeterminism is still possible and not measured.
  • domain assumption Baseline GNAS methods are configured fairly and under comparable budgets.
    Section 4.1.3 says identical settings to AutoGEL and HGNAS++ are adopted, but detailed per-baseline configurations and code are not provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LLM4GNAS: A Large Language Model Based Toolkit for Graph Neural Architecture Search." pith.science (2026). https://pith.science/paper/T2I6MFO4

@misc{pith2026250210459,
  author       = {Pith},
  title        = {Pith review of: LLM4GNAS: A Large Language Model Based Toolkit for Graph Neural Architecture Search},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T2I6MFO4}},
  note         = {Machine review of arXiv:2502.10459}
}
read the original abstract

Graph Neural Architecture Search (GNAS) facilitates the automatic design of Graph Neural Networks (GNNs) tailored to specific downstream graph learning tasks. However, existing GNAS approaches often require manual adaptation to new graph search spaces, necessitating substantial code optimization and domain-specific knowledge. To address this challenge, we present LLM4GNAS, a toolkit for GNAS that leverages the generative capabilities of Large Language Models (LLMs). LLM4GNAS includes an algorithm library for graph neural architecture search algorithms based on LLMs, enabling the adaptation of GNAS methods to new search spaces through the modification of LLM prompts. This approach reduces the need for manual intervention in algorithm adaptation and code modification. The LLM4GNAS toolkit is extensible and robust, incorporating LLM-enhanced graph feature engineering, LLM-enhanced graph neural architecture search, and LLM-enhanced hyperparameter optimization. Experimental results indicate that LLM4GNAS outperforms existing GNAS methods on tasks involving both homogeneous and heterogeneous graphs.

Figures

Figures reproduced from arXiv: 2502.10459 by the authors.

Figure 1
Figure 1. The overall framework of LLM4GNAS. LLM4GNAS leverages Large Language Models (LLMs) as the controller to [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Variation in test accuracy of the best GNNs found at each iteration during the architecture search process for [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Iterations of the best GNN architectures generated by LLM4GNAS on homogeneous graphs. Colored blocks indicate [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Enhancing LLMs' Reasoning-Intensive Multimedia Search Capabilities through Fine-Tuning and Reinforcement Learning

    cs.IR 2025-05 conditional novelty 6.0 of 10

    A two-stage fine-tuning and reinforcement-learning method makes LLMs generate token-efficient natural-language search plans, reporting strong accuracy gains on financial and news search benchmarks.

Reference graph

Works this paper leans on

52 extracted references · 43 canonical work pages · cited by 1 Pith paper

  1. [37]

    Graph neural architecture search with gpt-4

    Haishuai Wang, Yang Gao, Xin Zheng, Peng Zhang, Hongyang Chen, and Jiajun Bu. Graph neural architecture search with gpt-4. ArXiv preprint, 2023

    work page 2023

  2. [9]

    Het- erogeneous graph neural architecture search with gpt-4

    Haoyuan Dong, Yang Gao, Haishuai Wang, Hong Yang, and Peng Zhang. Het- erogeneous graph neural architecture search with gpt-4. ArXiv preprint, 2023

    work page 2023

  3. [1]

    Multi-view graph neural architecture search for biomedical entity and relation extraction

    Raeed Al-Sabri, Jianliang Gao, Jiamin Chen, Babatounde Moctard Oloulade, and Tengfei Lyu. Multi-view graph neural architecture search for biomedical entity and relation extraction. IEEE/ACM Transactions on Computational Biology and Bioinformatics, 20(2):1221–1233, 2022

    work page 2022

  4. [2]

    Palm 2 technical report

    Rohan Anil, Andrew M Dai, Orhan Firat, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. Palm 2 technical report. ArXiv preprint, 2023

    work page 2023

  5. [3]

    Beyond low-frequency infor- mation in graph convolutional networks

    Deyu Bo, Xiao Wang, Chuan Shi, and Huawei Shen. Beyond low-frequency infor- mation in graph convolutional networks. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 3950–3957, 2021

    work page 2021

  6. [4]

    How attentive are graph attention networks? arXiv preprint arXiv:2105.14491, 2021

    Shaked Brody, Uri Alon, and Eran Yahav. How attentive are graph attention networks? arXiv preprint arXiv:2105.14491, 2021

    arXiv 2021

  7. [5]

    Simple and deep graph convolutional networks

    Ming Chen, Zhewei Wei, Zengfeng Huang, Bolin Ding, and Yaliang Li. Simple and deep graph convolutional networks. CoRR, abs/2007.02133, 2020

    arXiv 2007

  8. [6]

    Palm: Scaling language modeling with pathways

    Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebas- tian Gehrmann, et al. Palm: Scaling language modeling with pathways. ArXiv preprint, abs/2204.02311, 2022

    arXiv 2022

Show all 52 references
  1. [7]

    BERT: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Langu...

    2019

  2. [8]

    Diffmg: Differentiable meta graph search for heterogeneous graph neural networks

    Yuhui Ding, Quanming Yao, Huan Zhao, and Tong Zhang. Diffmg: Differentiable meta graph search for heterogeneous graph neural networks. In Proc. of KDD, 2021

    2021

  3. [10]

    Matthias Fey and Jan E. Lenssen. Fast graph representation learning with PyTorch Geometric. In ICLR Workshop on Representation Learning on Graphs and Manifolds, 2019

    2019

  4. [11]

    Magnn: Metapath aggre- gated graph neural network for heterogeneous graph embedding

    Xinyu Fu, Jiani Zhang, Ziqiao Meng, and Irwin King. Magnn: Metapath aggre- gated graph neural network for heterogeneous graph embedding. In Proceedings of the web conference 2020 , pages 2331–2341, 2020

    2020

  5. [12]

    Gm2nas: multitask multiview graph neural architecture search

    Jianliang Gao, Raeed Al-Sabri, Babatounde Moctard Oloulade, Jiamin Chen, Tengfei Lyu, and Zhenpeng Wu. Gm2nas: multitask multiview graph neural architecture search. Knowledge and Information Systems , 65(10):4021–4054, 2023

    2023

  6. [13]

    Graph neural architecture search

    Yang Gao, Hong Yang, Peng Zhang, Chuan Zhou, and Yue Hu. Graph neural architecture search. In Christian Bessiere, editor, Proceedings of the Twenty- Ninth International Joint Conference on Artificial Intelligence, IJCAI 2020 , pages 1403–1409. ijcai.org, 2020

    2020

  7. [14]

    Hetero- geneous graph neural architecture search

    Yang Gao, Peng Zhang, Zhao Li, Chuan Zhou, Yongchao Liu, and Yue Hu. Hetero- geneous graph neural architecture search. In 2021 IEEE International Conference on Data Mining (ICDM) , pages 1066–1071. IEEE, 2021

    2021

  8. [15]

    Het- erogeneous graph neural architecture search

    Yang Gao, Peng Zhang, Zhao Li, Chuan Zhou, Yongchao Liu, and Yue Hu. Het- erogeneous graph neural architecture search. In Proc. of ICDM, 2021

    2021

  9. [16]

    Graphnas++: Distributed architecture search for graph neural networks

    Yang Gao, Peng Zhang, Hong Yang, Chuan Zhou, Zhihong Tian, Yue Hu, Zhao Li, and Jingren Zhou. Graphnas++: Distributed architecture search for graph neural networks. IEEE Transactions on Knowledge and Data Engineering , 2022

    2022

  10. [17]

    Hgnas++: efficient architecture search for heterogeneous graph neural networks

    Yang Gao, Peng Zhang, Chuan Zhou, Hong Yang, Zhao Li, Yue Hu, and S Yu Philip. Hgnas++: efficient architecture search for heterogeneous graph neural networks. IEEE Transactions on Knowledge and Data Engineering , 2023

    2023

  11. [18]

    Large- scale graph neural architecture search

    Chaoyu Guan, Xin Wang, Hong Chen, Ziwei Zhang, and Wenwu Zhu. Large- scale graph neural architecture search. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvári, Gang Niu, and Sivan Sabato, editors, International Conference on Machine Learning, ICML 2022, 17-23 ...

    2022

  12. [19]

    Gpt4graph: Can large language models understand graph structured data? an empirical evaluation and benchmarking

    Jiayan Guo, Lun Du, and Hengyu Liu. Gpt4graph: Can large language models understand graph structured data? an empirical evaluation and benchmarking. ArXiv preprint, 2023

    2023

  13. [20]

    Deberta: decoding- enhanced bert with disentangled attention

    Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. Deberta: decoding- enhanced bert with disentangled attention. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021 . Open- Review.net, 2021

    2021

  14. [21]

    Harnessing explanations: Llm-to-lm interpreter for enhanced text- attributed graph representation learning, 2023

    Xiaoxin He, Xavier Bresson, Thomas Laurent, Adam Perold, Yann LeCun, and Bryan Hooi. Harnessing explanations: Llm-to-lm interpreter for enhanced text- attributed graph representation learning, 2023

    2023

  15. [22]

    Open graph benchmark: Datasets for machine learning on graphs

    Weihua Hu, Matthias Fey, Marinka Zitnik, Yuxiao Dong, Hongyu Ren, Bowen Liu, Michele Catasta, and Jure Leskovec. Open graph benchmark: Datasets for machine learning on graphs. Advances in neural information processing systems , 33:22118–22133, 2020

    2020

  16. [23]

    Heterogeneous graph transformer

    Ziniu Hu, Yuxiao Dong, Kuansan Wang, and Yizhou Sun. Heterogeneous graph transformer. In Proceedings of the web conference 2020 , pages 2704–2710, 2020

    2020

  17. [24]

    Search to aggregate neighborhood for graph neural network

    ZHAO Huan, YAO Quanming, and TU Weiwei. Search to aggregate neighborhood for graph neural network. In 2021 IEEE 37th International Conference on Data Engineering (ICDE), 2021

    2021

  18. [25]

    Semi-supervised classification with graph convolutional networks

    Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016

    2016 arXiv

  19. [26]

    ALBERT: A lite BERT for self-supervised learning of language representations

    Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. ALBERT: A lite BERT for self-supervised learning of language representations. In 8th International Conference on Learning Repre- sentations, ICLR 2020, Addis Ababa, Ethiopia, April 26...

    2020

  20. [27]

    One-shot graph neural archi- tecture search with dynamic search space

    Yanxi Li, Zean Wen, Yunhe Wang, and Chang Xu. One-shot graph neural archi- tecture search with dynamic search space. In Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Applica- tions of Artificial Intelligence, IAAI 202...

    2021

  21. [28]

    Autograph: Automated graph neural network

    Yaoman Li and Irwin King. Autograph: Automated graph neural network. InNeu- ral Information Processing: 27th International Conference, ICONIP 2020, Bangkok, Thailand, November 23–27, 2020, Proceedings, Part II 27 , 2020

    2020

  22. [29]

    Roberta: A robustly optimized BERT pretraining approach

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized BERT pretraining approach. CoRR, 2019

    2019

  23. [30]

    Graph neural architecture search: A survey

    Babatounde Moctard Oloulade, Jianliang Gao, Jiamin Chen, Tengfei Lyu, and Raeed Al-Sabri. Graph neural architecture search: A survey. Tsinghua Science and Technology, 2021

    2021

  24. [31]

    Gpt-4 technical report, 2023

    OpenAI. Gpt-4 technical report, 2023

    2023

  25. [32]

    Graph neural architecture search under distribution shifts

    Yijian Qin, Xin Wang, Ziwei Zhang, Pengtao Xie, and Wenwu Zhu. Graph neural architecture search under distribution shifts. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvári, Gang Niu, and Sivan Sabato, editors, Inter- national Conference on Machine Learning, I...

    2022

  26. [33]

    Modeling relational data with graph convolutional net- works

    Michael Schlichtkrull, Thomas N Kipf, Peter Bloem, Rianne Van Den Berg, Ivan Titov, and Max Welling. Modeling relational data with graph convolutional net- works. In The semantic web: 15th international conference, ESWC 2018, Heraklion, Crete, Greece, June 3–7, 2018, proceedin...

    2018

  27. [34]

    Genetic-gnn: Evolutionary architecture search for graph neural networks

    Min Shi, Yufei Tang, Xingquan Zhu, Yu Huang, David Wilson, Yuan Zhuang, and Jianxun Liu. Genetic-gnn: Evolutionary architecture search for graph neural networks. Knowledge-Based Systems, 2022

    2022

  28. [35]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Isabelle Guyon, Ulrike von Luxburg, Samy Bengio, Hanna M. Wallach, Rob Fergus, S. V. N. Vishwanathan, and Roman Garnett, ...

    2017

  29. [36]

    Graph attention networks

    Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, Yoshua Bengio, et al. Graph attention networks. stat, 1050(20):10–48550, 2017

    2017

  30. [38]

    Heterogeneous graph attention network

    Xiao Wang, Houye Ji, Chuan Shi, Bai Wang, Yanfang Ye, Peng Cui, and Philip S Yu. Heterogeneous graph attention network. In The world wide web conference , pages 2022–2032, 2019

    2022

  31. [39]

    Autogel: An automated graph neural net- work with explicit link information

    Zhili Wang, Shimin Di, and Lei Chen. Autogel: An automated graph neural net- work with explicit link information. In Marc’Aurelio Ranzato, Alina Beygelzimer, Yann N. Dauphin, Percy Liang, and Jennifer Wortman Vaughan, editors,Advances in Neural Information Processing Systems 3...

    2021

  32. [40]

    Adversarially robust neural architecture search for graph neural networks

    Beini Xie, Heng Chang, Ziwei Zhang, Xin Wang, Daixin Wang, Zhiqiang Zhang, Rex Ying, and Wenwu Zhu. Adversarially robust neural architecture search for graph neural networks. In Proc. of CVPR, 2023

    2023

  33. [41]

    Large language models as optimizers

    Chengrun Yang, Xuezhi Wang, Yifeng Lu, Hanxiao Liu, Quoc V Le, Denny Zhou, and Xinyun Chen. Large language models as optimizers. ArXiv reprint, 2023

    2023

  34. [42]

    Graph-toolformer: To empower llms with graph reasoning ability via prompt augmented by chatgpt

    Jiawei Zhang. Graph-toolformer: To empower llms with graph reasoning ability via prompt augmented by chatgpt. ArXiv preprint, 2023

    2023

  35. [43]

    Using large language models for hyperparameter optimization

    Michael R Zhang, Nishkrit Desai, Juhan Bae, Jonathan Lorraine, and Jimmy Ba. Using large language models for hyperparameter optimization. In NeurIPS 2023 Foundation Models for Decision Making Workshop , 2023

    2023

  36. [44]

    Automl-gpt: Automatic machine learning with gpt

    Shujian Zhang, Chengyue Gong, Lemeng Wu, Xingchao Liu, and Mingyuan Zhou. Automl-gpt: Automatic machine learning with gpt. ArXiv preprint, 2023

    2023

  37. [45]

    Pasca: A graph neural architecture search system under the scalable paradigm

    Wentao Zhang, Yu Shen, Zheyu Lin, Yang Li, Xiaosen Li, Wen Ouyang, Yangyu Tao, Zhi Yang, and Bin Cui. Pasca: A graph neural architecture search system under the scalable paradigm. In WWW, 2022

    2022

  38. [46]

    Dynamic heterogeneous graph attention neural architecture search

    Zeyang Zhang, Ziwei Zhang, Xin Wang, Yijian Qin, Zhou Qin, and Wenwu Zhu. Dynamic heterogeneous graph attention neural architecture search. In Proc. of AAAI, 2023

    2023

  39. [47]

    Automated machine learning on graphs: A survey

    Ziwei Zhang, Xin Wang, and Wenwu Zhu. Automated machine learning on graphs: A survey. In Proc. of IJCAI, 2021. Conference acronym ’XX, April 28–May 2 , 2025, Sydney, Australia Yang Gao and Peng Zhang, et al

    2021

  40. [48]

    Autogt: Automated graph transformer architecture search

    Zizhao Zhang, Xin Wang, Chaoyu Guan, Ziwei Zhang, Haoyang Li, and Wenwu Zhu. Autogt: Automated graph transformer architecture search. In Proc. of ICLR, 2022

    2022

  41. [49]

    Can gpt-4 perform neural architecture search? ArXiv preprint, 2023

    Mingkai Zheng, Xiu Su, Shan You, Fei Wang, Chen Qian, Chang Xu, and Samuel Albanie. Can gpt-4 perform neural architecture search? ArXiv preprint, 2023

    2023

  42. [50]

    Auto-heg: Automated graph neural network on heterophilic graphs

    Xin Zheng, Miao Zhang, Chunyang Chen, Qin Zhang, Chuan Zhou, and Shirui Pan. Auto-heg: Automated graph neural network on heterophilic graphs. In Proceedings of the ACM Web Conference 2023 , 2023

    2023

  43. [51]

    Multi-relational graph neural architecture search with fine-grained message passing

    Xin Zheng, Miao Zhang, Chun cheng Jason Chen, Chaojie Li, Chuan Zhou, and Shirui Pan. Multi-relational graph neural architecture search with fine-grained message passing. 2022 IEEE International Conference on Data Mining (ICDM) , 2022

    2022

  44. [52]

    Auto-gnn: Neural architecture search of graph neural networks

    Kaixiong Zhou, Xiao Huang, Qingquan Song, Rui Chen, and Xia Hu. Auto-gnn: Neural architecture search of graph neural networks. Frontiers in big Data , 2022. Received 20 February 2007; revised 12 March 2009; accepted 5 June 2009

    2022

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.