Pith. sign in

REVIEW 3 major objections 5 minor 77 references

Freeze, Prompt, and Adapt: A Framework for Source-free Unsupervised GNN Prompting

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a frozen graph neural network can be adapted to a shifted, unlabeled target domain by training only input-side prompts, using the network's own confident predictions as pseudo-labels.

desk verdict A genuinely new problem setting (UGPP) and a reasonable first unsupervised GNN prompting method, but the 'no labels' claim is weakened by labeled-validation hyperparameter tuning and the abstract's 'consistently outperforms' does not hold at higher label budgets. read the letter →

arxiv 2505.16903 v2 pith:WSBQ7RUZ submitted 2025-05-22 cs.LG

Peyman Baghershahi , Sourav Medya This is my paper · ORCID
classification cs.LG
keywords graphneuralnetworksprompttuningunsupervisedlearningpseudo-labelingconsistencyregularizationsource-freedomainadaptationcovariateshiftclassification
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 introduces the Unsupervised Graph Prompting Problem (UGPP), a setting in which a pretrained GNN must be adapted to a shifted target domain with all parameters frozen, no target labels, and no access to the source data. It then proposes UGPrompt, the first fully unsupervised GNN prompting framework, which trains a small set of prompt vectors by enforcing consistency between the frozen GNN's predictions on a weakly augmented graph and on a prompted, strongly augmented graph. The training signal comes from the network's own confident predictions, used as pseudo-labels, together with a diversity regularizer that avoids collapsing all predictions to a few classes and an adversarial discriminator that keeps prompted graphs close to the target input distribution. Across graph and node classification benchmarks under homophily, PageRank, density, and clustering-coefficient shifts, the paper reports that unsupervised UGPrompt improves the frozen base model in every setting and outperforms supervised GNN prompting baselines in most cases, even when those baselines use labeled data. If this holds, it establishes unsupervised prompting as a practical way to adapt frozen GNNs.

What carries the argument

The load-bearing object is the prompting function $f(.;\theta_f)$, instantiated as a small set of trainable prompt vectors $t^*_j$ added to node features through soft attention weights $\alpha_{i,j}=\exp(x_i^\top t^*_j)/\sum_l \exp(x_i^\top t^*_l)$, so each node receives a prompt vector $\sum_j \alpha_{i,j}t^*_j$. The GNN $\varphi = h\circ g$ stays frozen, and the prompt function is trained by the consistency loss $L_c$ in Eq. (3), which uses confident pseudo-labels from the weakly augmented graph; the diversity loss $L_{div}$ in Eq. (4); and the adversarial domain loss $L_{adv}$ in Eq. (6), with a discriminator $d$ over encoder embeddings. The prompted graph is deliberately treated as a strongly augmented instance, which is what lets pseudo-labeling and consistency regularization operate with no target labels and no source-data access.

What would settle it

Construct a target split where the source-trained GNN's confident predictions are known to be wrong, for example a homophily-reversed split in which high-confidence predictions are anti-correlated with true labels, and run UGPrompt; if F1 still improves over the frozen base model, the pseudo-label assumption is not load-bearing, and if it degrades, the claim that the method adapts under arbitrary covariate shift is refuted.

Watch

Extended reading notes

Core claim

The central claim is that prompting alone, training only input-side prompt vectors while the encoder and projection head of a pretrained GNN stay completely frozen, can adapt the model to a new unlabeled target distribution under covariate shift. The training signal is consistency: each target graph is weakly and strongly augmented, the strongly augmented version is passed through the learnable prompting function, and the prompt parameters are optimized so that the GNN's predictions for the prompted graph match the frozen GNN's high-confidence pseudo-label from the weak augmentation. Diversity regularization maximizes the entropy of the average batch prediction, and an adversarial discriminator over encoder representations penalizes prompted graphs that leave the distribution of non-prompted augmented graphs. In the paper's experiments on six datasets with GCN and GAT base models, this objective improves over the frozen base model in every configuration and, in most cases, achieves higher F1 than supervised prompting baselines that use 25% or more labeled target data.

Load-bearing premise

The load-bearing premise is that the frozen GNN's confident predictions on weakly augmented target graphs are correct enough to serve as pseudo-labels, and if covariate shift makes those confident predictions systematically wrong, the training signal amplifies the error instead of correcting it.

Editorial extensions

If this is right

  • UGPP removes the need for labeled target data and source-data access, so prompting can be applied to large unlabeled graph collections, including private or inaccessible source graphs.
  • A frozen GNN can be adapted by training only a small set of prompt vectors, preserving the parameter efficiency that motivates prompting in the first place.
  • In the reported experiments, supervised prompt baselines often hurt the frozen model under covariate shift, while the unsupervised method consistently improves it across datasets, GNN architectures, and shift types.
  • When a few labels do exist, the same consistency objective can be augmented with a supervised term, and the paper reports further gains with as little as 5% labeled data.
  • The framework is not tied to one prompting function: swapping in a different prefix-prompting module still yields gains, indicating the unsupervised training recipe rather than the specific prompt design is what carries the result.

Reading between the lines

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

  • A testable extension is to measure the accuracy of the confident pseudo-labels on a held-out labeled slice of the target set; if UGPrompt's gains shrink when pseudo-label accuracy is low, the method's robustness claim is bounded by the frozen GNN's calibration under shift.
  • The weak/strong augmentation pairing is matched to a feature-prefix prompt, which suggests the recipe would need a structurally aligned prompt (e.g., edge-level prompting) to handle pure structural shifts, a direction the paper leaves open.
  • The adversarial discriminator could be reused at inference as an out-of-distribution score for prompted graphs, letting users abstain on low-scoring inputs; the paper does not evaluate this use.
  • By analogy with source-free domain adaptation, the same input-prompt training loop might transfer to other frozen encoders (tabular, point-cloud, or time-series models) whose inputs can be weakly and strongly augmented.
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

3 major / 5 minor

Summary. The paper proposes a new problem setting, Unsupervised Graph Prompting Problem (UGPP), in which a pre-trained GNN is frozen, the source data is inaccessible, the target labels are unavailable, and the target distribution is shifted. The authors introduce UGPrompt, an unsupervised prompting framework that trains a feature-level prompting function by consistency regularization with confident pseudo-labels from the frozen GNN, plus diversity and domain-adversarial regularization. Experiments on graph and node classification under several covariate shifts claim that UGPrompt, with no labels, outperforms existing supervised GNN prompting methods in most cases at 25% label budgets, and remains competitive at higher label budgets.

Significance. If the main comparison is valid, UGPrompt would be the first unsupervised GNN prompting method to beat supervised GNN prompting in the low-label regime, and it would demonstrate a practical source-free, label-free adaptation paradigm for frozen GNNs. The paper ships code, runs a large benchmark across multiple base architectures (GCN and GAT), distribution shifts (homophily, PageRank, graph density, clustering coefficient), and includes several ablations and a few-shot extension. The framework is model-agnostic with respect to the prompting function, which is an interesting and useful property. The significance is, however, conditional on resolving two load-bearing issues: whether any target labels are used for hyperparameter selection, and whether the central claim of consistency across label budgets can be supported without overstatement.

major comments (3)
  1. [A.3.3] The statement 'We tune hyper-parameters based on the average F1 score on validation sets' is load-bearing for the central claim that UGPrompt is fully unsupervised. If the validation set is drawn from the target domain and its labels are used to select lambda_1, lambda_2, tau, p_s, p_w, and n_p, then UGPrompt is not label-free; it merely uses labels for model selection rather than parameter training. The manuscript must clarify whether these validation labels are from the target set. If they are, the comparison against baselines that use 25% labels for training is not a no-labels evaluation, and the authors should either re-tune on source validation only, use an unsupervised validation criterion, or explicitly report the sensitivity of the results to the validation labels.
  2. [Abstract, Tables 13 and 14] The abstract claims UGPrompt 'consistently outperforms state-of-the-art supervised prompting methods with access to labeled data', but the appendix results do not support 'consistently'. In Table 14 (50% label budget), UGPrompt is not the best method on Cora (57.3 vs GPF-Plus 58.2), CiteSeer (45.7 vs GPF-Plus 46.8), or PubMed (61.2 vs GraphPrompt+ 64.9). Figure 4 similarly shows that at 75% and 100% labels UGPrompt is often second-best rather than best on node classification. The claim should be made precise, e.g., 'in most cases' as in the introduction, or supported by a statistical significance test across the reported seeds that justifies 'consistently'.
  3. [Section 4.1, Eq. (3)] The consistency objective treats the frozen GNN's confident predictions on weakly augmented target graphs as pseudo-labels. The paper assumes P^t_{Y|X} = P^s_{Y|X} (Section 3.2) but provides no evidence about the correctness of these pseudo-labels under the constructed covariate shifts. If the frozen model is confidently wrong on a large fraction of target samples, the consistency, diversity, and domain losses would jointly reinforce those errors. The manuscript reports no pseudo-label accuracy, no sensitivity analysis for the confidence threshold tau, and no comparison with a baseline that minimizes predictive entropy without using pseudo-labels. I request at least one such diagnostic to support the claim that the label-free training signal is not self-reinforcingly biased.
minor comments (5)
  1. [Section 4.1, Eq. (1)] The denominator in the softmax expression for alpha_{i,j} is missing parentheses: it should read \sum_{l=1}^{n_t} \exp(x_i^T t_l^*).
  2. [Section 5.4.2] There is a typo: 'CireSeer' should be 'CiteSeer'.
  3. [Algorithm 2] In the Input line, 'freezed' should be 'frozen'.
  4. [References] Reference [31] appears to be a duplicate of reference [30]; the same paper by Liang et al. is listed twice.
  5. [Table 12 caption] The caption reads 'Statistic of the datasets'; it should be 'Statistics of the datasets'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; UGPrompt's self-referential pseudo-labeling is a bootstrap, not a construction that forces the claimed results.

full rationale

The paper's central claim is empirical: UGPrompt is evaluated with held-out labels on target graphs (Tables 1, 2, 13, 14), so the headline comparison against supervised prompting methods is not forced by the training objective. The training signal in Eq. 3 does use the frozen GNN's own confident predictions on weak augmentations as pseudo-labels, and the final objective (Eq. 7) only enforces consistency, diversity, and domain alignment; this is a self-referential bootstrap of the standard pseudo-labeling/consistency-regularization type, not a construction in which the predicted quantity equals the input quantity. No parameter is fitted to held-out labels and then renamed as a prediction; no load-bearing self-citation chain or imported uniqueness theorem is used. The assumption P^t_{Y|X}=P^s_{Y|X} (Sec. 3.2) and the absence of pseudo-label accuracy analysis are robustness and correctness concerns, but the paper states the assumption explicitly and does not derive its empirical gains from it by definition. The comparisons are against external baselines and a public benchmark [76], which makes the evaluation self-contained.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim rests on hyperparameters selected with target validation labels, which contradicts the no-label setting. The method's self-training signal is the model's own predictions, so there is no external supervision. The more free parameters tuned per dataset, the less 'unsupervised' the method truly is.

free parameters (6)
  • confidence threshold tau = selected from {0.1, 0.3, 0.5, 0.7}
    Filters pseudo-labels; tuned per dataset on labeled validation F1.
  • diversity weight lambda_1 = selected from {0.25, 0.5, 0.75, 1.0, 1.25, 1.5}
    Balances diversity regularization; tuned on validation labels.
  • domain adaptation weight lambda_2 = selected from {0.25, 0.5, 0.75, 1.0, 1.25, 1.5}
    Balances adversarial domain loss; tuned on validation labels.
  • strong augmentation probability p_s = values in {0.0, 0.1, 0.2, 0.3, 0.4} in ablation
    Controls strength of augmentation; performance varies by dataset.
  • weak augmentation probability p_w = from {0.05, 0.1, 0.2}
    Controls weak augmentation used to generate pseudo-labels.
  • number of prompting vectors n_p = from {10, 20, 30, 50, ENG}
    Capacity of the prompt; ablation shows marginal effect.
assumptions (4)
  • domain assumption P^t_{Y|X} = P^s_{Y|X} (conditional label distribution is preserved under covariate shift)
    Stated in Section 3.2; without this, pseudo-labels from the frozen GNN may be systematically wrong.
  • domain assumption The frozen GNN's confident predictions on weakly augmented target graphs are reliable enough to serve as pseudo-labels
    Section 4.1 uses thresholded predictions as ground truth for the prompting loss.
  • domain assumption Feature masking augmentation preserves the label of the graph or node
    Consistency regularization assumes the augmentation does not change the class; if masking removes class-discriminative features, pseudo-labels are corrupted.
  • ad hoc to paper The prompting function f has enough capacity to align target with source
    The additive feature prompt may be insufficient for complex shifts; no theoretical guarantee is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Freeze, Prompt, and Adapt: A Framework for Source-free Unsupervised GNN Prompting." pith.science (2026). https://pith.science/paper/WSBQ7RUZ

@misc{pith2026250516903,
  author       = {Pith},
  title        = {Pith review of: Freeze, Prompt, and Adapt: A Framework for Source-free Unsupervised GNN Prompting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WSBQ7RUZ}},
  note         = {Machine review of arXiv:2505.16903}
}
read the original abstract

Prompt tuning has become a key mechanism for adapting pre-trained Graph Neural Networks (GNNs) to new downstream tasks. However, existing approaches are predominantly supervised, relying on labeled data to optimize the prompting parameters and typically fine-tuning a task-specific prediction head -- practices that undermine the promise of parameter-efficient adaptation. We propose Unsupervised Graph Prompting Problem (UGPP), a challenging new setting where the pre-trained GNN is kept entirely frozen, labels on the target domain are unavailable, the source data is inaccessible, and the target distribution exhibits covariate shift. To address this, we propose UGPrompt, the first fully unsupervised GNN prompting framework. UGPrompt leverages consistency regularization and pseudo-labeling to train a prompting function, complemented with diversity and domain regularization to mitigate class imbalance and distribution mismatch. Our extensive experiments demonstrate that UGPrompt consistently outperforms state-of-the-art supervised prompting methods with access to labeled data, demonstrating the viability of unsupervised prompting as a practical adaptation paradigm for GNNs.

Figures

Figures reproduced from arXiv: 2505.16903 by the authors.

Figure 1
Figure 1. Overview of UGPROMPT. A) A non-parametric algorithm generates a weak augmentation Gw and a strong augmentation Gs from an unlabeled graph G. B) The learnable prompting function f generates a prompted graph Gp from Gs. C) The base GNN with frozen parameters scores Gp and Gw. A discriminator taking input from the latent representation (z G a for Gw and z G p for Gp) of the GNN’s encoder regularizes the model to adapt … view at source ↗
Figure 2
Figure 2. The effect of regularization objectives on [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Distribution of embeddings generated by the base [PITH_FULL_IMAGE:figures/full_fig_p020_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Performance gains for GCN and GAT base models on graph and node classification tasks [PITH_FULL_IMAGE:figures/full_fig_p024_4.png]
Figure 5
Figure 5. Figure 5: Performance gains for GCN as the base model on graph and node classification tasks in [PITH_FULL_IMAGE:figures/full_fig_p025_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

77 extracted references · 63 canonical work pages

  1. [1]

    Learning with pseudo-ensembles

    Philip Bachman, Ouais Alsharif, and Doina Precup. Learning with pseudo-ensembles. In Proceedings of the 27th International Conference on Neural Information Processing Systems - Volume 2, page 3365–3373, 2014

    work page 2014

  2. [2]

    Evaluating robustness and uncertainty of graph models under structural dis- tributional shifts

    Gleb Bazhenov, Denis Kuznedelev, Andrey Malinin, Artem Babenko, and Liudmila Prokhorenkova. Evaluating robustness and uncertainty of graph models under structural dis- tributional shifts. In Thirty-seventh Conference on Neural Information Processing Systems , 2023

    work page 2023

  3. [3]

    Borgwardt, Cheng Soon Ong, Stefan Schönauer, S

    Karsten M. Borgwardt, Cheng Soon Ong, Stefan Schönauer, S. V . N. Vishwanathan, Alex J. Smola, and Hans-Peter Kriegel. Protein function prediction via graph kernels. Bioinformatics, 21:i47–i56, 2005

    work page 2005

  4. [4]

    Introduction to Statistical Learning Theory, pages 169–207

    Olivier Bousquet, Stéphane Boucheron, and Gábor Lugosi. Introduction to Statistical Learning Theory, pages 169–207. Springer Berlin Heidelberg, 2004

    work page 2004

  5. [5]

    Language models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-V oss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott G...

    work page 1901

  6. [6]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In Proceedings of the 37th International Conference on Machine Learning, 2020

    work page 2020

  7. [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 Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, 2019

    2019

  8. [8]

    A closer look at distribution shifts and out-of-distribution generalization on graphs

    Mucong Ding, Kezhi Kong, Jiuhai Chen, John Kirchenbauer, Micah Goldblum, David Wipf, Furong Huang, and Tom Goldstein. A closer look at distribution shifts and out-of-distribution generalization on graphs. In NeurIPS 2021 Workshop on Distribution Shifts: Connecting Methods and Applications, 2021

    work page 2021

Show all 77 references
  1. [9]

    Generalizing graph neural networks on out-of-distribution graphs

    Shaohua Fan, Xiao Wang, Chuan Shi, Peng Cui, and Bai Wang. Generalizing graph neural networks on out-of-distribution graphs. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(1):322–337, 2024

    2024

  2. [10]

    Universal prompt tuning for graph neural networks

    Taoran Fang, Yunchao Mercer Zhang, Yang Yang, Chunping Wang, and Lei CHEN. Universal prompt tuning for graph neural networks. In Neural Information Processing Systems, 2023

    2023

  3. [11]

    Abolfazl Farahani, Sahar V oghoei, Khaled Rasheed, and Hamid R. Arabnia. A brief review of domain adaptation. In Robert Stahlbock, Gary M. Weiss, Mahmoud Abou-Nasr, Cheng-Ying Yang, Hamid R. Arabnia, and Leonidas Deligiannidis, editors,Advances in Data Science and Information ...

    2021

  4. [12]

    Talk like a graph: Encoding graphs for large language models

    Bahare Fatemi, Jonathan Halcrow, and Bryan Perozzi. Talk like a graph: Encoding graphs for large language models. In The Twelfth International Conference on Learning Representations, 2024

    2024

  5. [13]

    Schoenholz, Patrick F

    Justin Gilmer, Samuel S. Schoenholz, Patrick F. Riley, Oriol Vinyals, and George E. Dahl. Neural message passing for quantum chemistry. In Proceedings of the 34th International Conference on Machine Learning - Volume 70, page 1263–1272, 2017

    2017

  6. [14]

    GOOD: A graph out-of-distribution benchmark

    Shurui Gui, Xiner Li, Limei Wang, and Shuiwang Ji. GOOD: A graph out-of-distribution benchmark. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 11

    2022

  7. [15]

    Parameter-efficient fine- tuning for large models: A comprehensive survey

    Zeyu Han, Chao Gao, Jinyang Liu, Jeff Zhang, and Sai Qian Zhang. Parameter-efficient fine- tuning for large models: A comprehensive survey. Transactions on Machine Learning Research, 2024

    2024

  8. [16]

    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. In Advances in Neural Information Processing Systems, volume 33, pages 22118–22133, 2020

    2020

  9. [17]

    Learning discrete representations via information maximizing self-augmented training

    Weihua Hu, Takeru Miyato, Seiya Tokui, Eiichi Matsumoto, and Masashi Sugiyama. Learning discrete representations via information maximizing self-augmented training. In Proceedings of the 34th International Conference on Machine Learning - Volume 70, page 1558–1567, 2017

    2017

  10. [18]

    Prodigy: Enabling in-context learning over graphs

    Qian Huang, Hongyu Ren, Peng Chen, Gregor Kržmanc, Daniel Zeng, Percy Liang, and Jure Leskovec. Prodigy: Enabling in-context learning over graphs. ArXiv, abs/2305.12600, 2023

    2023 arXiv

  11. [19]

    Domain adaptation without source data.IEEE Transactions on Artificial Intelligence, 2:508–518, 2021

    Youngeun Kim, Donghyeon Cho, Kyeongtak Han, Priyadarshini Panda, and Sungeun Hong. Domain adaptation without source data.IEEE Transactions on Artificial Intelligence, 2:508–518, 2021

    2021

  12. [20]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In International Conference on Learning Representations, 2017

    2017

  13. [21]

    Understanding attention and gener- alization in graph neural networks

    Boris Knyazev, Graham W Taylor, and Mohamed Amer. Understanding attention and gener- alization in graph neural networks. In Advances in Neural Information Processing Systems, volume 32, 2019

    2019

  14. [22]

    Large language models are zero-shot reasoners

    Takeshi Kojima, Shixiang (Shane) Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. In Advances in Neural Information Processing Systems, volume 35, pages 22199–22213, 2022

    2022

  15. [23]

    Temporal ensembling for semi-supervised learning

    Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. In Interna- tional Conference on Learning Representations, 2017

    2017

  16. [24]

    Pseudo-label : The simple and efficient semi-supervised learning method for deep neural networks

    Dong-Hyun Lee. Pseudo-label : The simple and efficient semi-supervised learning method for deep neural networks. ICML 2013 Workshop : Challenges in Representation Learning (WREPL), 2013

    2013

  17. [25]

    The power of scale for parameter-efficient prompt tuning

    Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In Marie-Francine Moens, Xuanjing Huang, Lucia Specia, and Scott Wen-tau Yih, editors, Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processin...

    2021

  18. [26]

    Ood-gnn: Out-of-distribution gen- eralized graph neural network

    Haoyang Li, Xin Wang, Ziwei Zhang, and Wenwu Zhu. Ood-gnn: Out-of-distribution gen- eralized graph neural network. IEEE Trans. on Knowl. and Data Eng. , 35(7):7328–7340, 2023

    2023

  19. [27]

    Learning invariant graph representations for out-of-distribution generalization

    Haoyang Li, Ziwei Zhang, Xin Wang, and Wenwu Zhu. Learning invariant graph representations for out-of-distribution generalization. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022

    2022

  20. [28]

    A comprehensive survey on source-free domain adaptation

    Jingjing Li, Zhiqi Yu, Zhekai Du, Lei Zhu, and Heng Tao Shen. A comprehensive survey on source-free domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(8):5743–5762, 2024

    2024

  21. [29]

    Prefix-tuning: Optimizing continuous prompts for generation

    Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4582–4597, 2021

    2021

  22. [30]

    Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation

    Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International Conference on Machine Learning, pages 6028–6039, 2020. 12

    2020

  23. [31]

    Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation

    Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International Conference on Machine Learning (ICML), pages 6028–6039, 2020

    2020

  24. [32]

    Large scale learning on non-homophilous graphs: New bench- marks and strong simple methods

    Derek Lim, Felix Matthew Hohne, Xiuyu Li, Sijia Linda Huang, Vaishnavi Gupta, Omkar Prasad Bhalerao, and Ser-Nam Lim. Large scale learning on non-homophilous graphs: New bench- marks and strong simple methods. In Advances in Neural Information Processing Systems , 2021

    2021

  25. [33]

    Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing

    Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Comput. Surv., 55(9), 2023

    2023

  26. [34]

    Graphprompt: Unifying pre-training and downstream tasks for graph neural networks

    Zemin Liu, Xingtong Yu, Yuan Fang, and Xinming Zhang. Graphprompt: Unifying pre-training and downstream tasks for graph neural networks. In Proceedings of the ACM Web Conference 2023, page 417–428. Association for Computing Machinery, 2023

    2023

  27. [35]

    Position: Graph foundation models are already here, 2024

    Haitao Mao, Zhikai Chen, Wenzhuo Tang, Jianan Zhao, Yao Ma, Tong Zhao, Neil Shah, Mikhail Galkin, and Jiliang Tang. Position: Graph foundation models are already here, 2024

    2024

  28. [36]

    Refram- ing instructional prompts to GPTk’s language

    Swaroop Mishra, Daniel Khashabi, Chitta Baral, Yejin Choi, and Hannaneh Hajishirzi. Refram- ing instructional prompts to GPTk’s language. In Findings of the Association for Computational Linguistics: ACL 2022, pages 589–612, 2022

    2022

  29. [37]

    Future directions in the theory of graph machine learning, 2024

    Christopher Morris, Fabrizio Frasca, Nadav Dym, Haggai Maron, ˙Ismail ˙Ilkan Ceylan, Ron Levie, Derek Lim, Michael Bronstein, Martin Grohe, and Stefanie Jegelka. Future directions in the theory of graph machine learning, 2024

    2024

  30. [38]

    Automatic differentiation in pytorch

    Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017

    2017

  31. [39]

    Geom-gcn: Geo- metric graph convolutional networks

    Hongbin Pei, Bingzhe Wei, Kevin Chen-Chuan Chang, Yu Lei, and Bo Yang. Geom-gcn: Geo- metric graph convolutional networks. In International Conference on Learning Representations, 2020

    2020

  32. [40]

    Improving language understanding by generative pre-training

    Alec Radford and Karthik Narasimhan. Improving language understanding by generative pre-training. 2018

    2018

  33. [41]

    Language models are unsupervised multitask learners

    Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019

    2019

  34. [42]

    Peters, Swabha Swayamdipta, and Thomas Wolf

    Sebastian Ruder, Matthew E. Peters, Swabha Swayamdipta, and Thomas Wolf. Transfer learning in natural language processing. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Tutorials, pages 15–18, 2019

    2019

  35. [43]

    Regularization with stochastic trans- formations and perturbations for deep semi-supervised learning

    Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic trans- formations and perturbations for deep semi-supervised learning. In Proceedings of the 30th International Conference on Neural Information Processing Systems, page 1171–1179, 2016

    2016

  36. [44]

    Brenda, the enzyme database: updates and major new developments

    Ida Schomburg, Antje Chang, Christian Ebeling, Marion Gremse, Christian Heldt, Gregor Huhn, and Dietmar Schomburg. Brenda, the enzyme database: updates and major new developments. Nucleic Acids Research, 32:D431–D433, 2004

    2004

  37. [45]

    Understanding Machine Learning: From Theory to Algorithms

    Shai Shalev-Shwartz and Shai Ben-David. Understanding Machine Learning: From Theory to Algorithms. Cambridge University Press, 2014

    2014

  38. [46]

    Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel

    Kihyuk Sohn, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin D. Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel. Fixmatch: simplifying semi-supervised learning with consistency and confidence. In Proceedings of the 34th International Conference on Neural ...

    2020

  39. [47]

    Unleashing the power of graph data augmentation on covariate distribution shift

    Yongduo Sui, Qitian Wu, Jiancan Wu, Qing Cui, Longfei Li, Jun Zhou, Xiang Wang, and Xiangnan He. Unleashing the power of graph data augmentation on covariate distribution shift. In Advances in Neural Information Processing Systems, volume 36, pages 18109–18131, 2023. 13

    2023

  40. [48]

    Gppt: Graph pre-training and prompt tuning to generalize graph neural networks

    Mingchen Sun, Kaixiong Zhou, Xin He, Ying Wang, and Xin Wang. Gppt: Graph pre-training and prompt tuning to generalize graph neural networks. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, page 1717–1727. Association for Computing Mach...

    2022

  41. [49]

    All in one: Multi-task prompting for graph neural networks

    Xiangguo Sun, Hong Cheng, Jia Li, Bo Liu, and Jihong Guan. All in one: Multi-task prompting for graph neural networks. In ACM SIGKDD Conference on Knowledge Discovery and Data Mining, page 2120–2131, 2023

    2023

  42. [50]

    Sutherland, Lee A

    Jeffrey J. Sutherland, Lee A. O’Brien, and Donald F. Weaver. Spline-fitting with a genetic algorithm: A method for developing classification structure-activity relationships. Journal of Chemical Information and Computer Sciences, 43(6):1906–1915, 2003

    1906

  43. [51]

    Representation learning with contrastive predictive coding

    Aäron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. ArXiv, abs/1807.03748, 2018

    2018 arXiv

  44. [52]

    Graph attention networks

    Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. Graph attention networks. In International Conference on Learning Representations, 2018

    2018

  45. [53]

    Freematch: Self- adaptive thresholding for semi-supervised learning

    Yidong Wang, Hao Chen, Qiang Heng, Wenxin Hou, Yue Fan, Zhen Wu, Jindong Wang, Marios Savvides, Takahiro Shinozaki, Bhiksha Raj, Bernt Schiele, and Xing Xie. Freematch: Self- adaptive thresholding for semi-supervised learning. In The Eleventh International Conference on Learni...

    2023

  46. [54]

    Chain-of-thought prompting elicits reasoning in large language models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian ichter, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems, volume 35, pages 24824–24837. Curran A...

    2022

  47. [55]

    Discovering invariant rationales for graph neural networks

    Yingxin Wu, Xiang Wang, An Zhang, Xiangnan He, and Tat-Seng Chua. Discovering invariant rationales for graph neural networks. In International Conference on Learning Representations, 2022

    2022

  48. [56]

    Feinberg, Joseph Gomes, Caleb Geniesse, Aneesh S

    Zhenqin Wu, Bharath Ramsundar, Evan N. Feinberg, Joseph Gomes, Caleb Geniesse, Aneesh S. Pappu, Karl Leswing, and Vijay Pande. Moleculenet: a benchmark for molecular machine learning. 9:513–530, 2018

    2018

  49. [57]

    Jun Xia, Lirong Wu, Jintao Chen, Bozhen Hu, and Stan Z. Li. Simgrace: A simple framework for graph contrastive learning without data augmentation. In Proceedings of the ACM Web Conference 2022, page 1070–1079, 2022

    2022

  50. [58]

    Hovy, Minh-Thang Luong, and Quoc V

    Qizhe Xie, Eduard H. Hovy, Minh-Thang Luong, and Quoc V . Le. Self-training with noisy stu- dent improves imagenet classification. Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684–10695, 2020

    2020

  51. [59]

    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. In The Twelfth International Conference on Learning Representations, 2024

    2024

  52. [60]

    Generalized out-of-distribution detection: A survey

    Jingkang Yang, Kaiyang Zhou, Yixuan Li, and Ziwei Liu. Generalized out-of-distribution detection: A survey. Int. J. Comput. Vision, 132(12):5635–5662, 2024

    2024

  53. [61]

    Generalized source-free domain adaptation

    Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Generalized source-free domain adaptation. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 8958–8967, 2021

    2021

  54. [62]

    Cohen, and Ruslan Salakhutdinov

    Zhilin Yang, William W. Cohen, and Ruslan Salakhutdinov. Revisiting semi-supervised learning with graph embeddings. In Proceedings of the 33rd International Conference on International Conference on Machine Learning - Volume 48, page 40–48, 2016

    2016

  55. [63]

    Gnnex- plainer: Generating explanations for graph neural networks

    Zhitao Ying, Dylan Bourgeois, Jiaxuan You, Marinka Zitnik, and Jure Leskovec. Gnnex- plainer: Generating explanations for graph neural networks. In Advances in Neural Information Processing Systems, volume 32, 2019. 14

    2019

  56. [64]

    Kaichao You, Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I. Jordan. Universal domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019

    2019

  57. [65]

    Graph contrastive learning with augmentations

    Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. Graph contrastive learning with augmentations. In Advances in Neural Information Processing Systems, volume 33, pages 5812–5823, 2020

    2020

  58. [66]

    Generalized graph prompt: Toward a unification of pre-training and downstream tasks on graphs

    Xingtong Yu, Zhenghao Liu, Yuan Fang, Zemin Liu, Sihong Chen, and Xinming Zhang. Generalized graph prompt: Toward a unification of pre-training and downstream tasks on graphs. IEEE Transactions on Knowledge and Data Engineering, 36(11):6237–6250, 2024

    2024

  59. [67]

    Node-time conditional prompt learning in dynamic graphs

    Xingtong Yu, Zhenghao Liu, Xinming Zhang, and Yuan Fang. Node-time conditional prompt learning in dynamic graphs. In The Thirteenth International Conference on Learning Represen- tations, 2025

    2025

  60. [68]

    Multigprompt for multi-task pre-training and prompting on graphs

    Xingtong Yu, Chang Zhou, Yuan Fang, and Xinming Zhang. Multigprompt for multi-task pre-training and prompting on graphs. In Proceedings of the ACM Web Conference 2024, page 515–526. Association for Computing Machinery, 2024

    2024

  61. [69]

    H. Yuan, H. Yu, S. Gui, and S. Ji. Explainability in graph neural networks: A taxonomic survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(05):5782–5799, 2023

    2023

  62. [70]

    Graphsaint: Graph sampling based inductive learning method

    Hanqing Zeng, Hongkuan Zhou, Ajitesh Srivastava, Rajgopal Kannan, and Viktor Prasanna. Graphsaint: Graph sampling based inductive learning method. In International Conference on Learning Representations, 2020

    2020

  63. [71]

    Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling

    Bowen Zhang, Yidong Wang, Wenxin Hou, HAO WU, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. In Advances in Neural Information Processing Systems, volume 34, pages 18408–18419, 2021

    2021

  64. [72]

    Large language models are human-level prompt engineers

    Yongchao Zhou, Andrei Ioan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy Ba. Large language models are human-level prompt engineers. In The Eleventh International Conference on Learning Representations, 2023

    2023

  65. [73]

    Beyond homophily in graph neural networks: Current limitations and effective designs

    Jiong Zhu, Yujun Yan, Lingxiao Zhao, Mark Heimann, Leman Akoglu, and Danai Koutra. Beyond homophily in graph neural networks: Current limitations and effective designs. In Advances in Neural Information Processing Systems, volume 33, pages 7793–7804, 2020

    2020

  66. [74]

    Shift-robust gnns: Overcoming the limitations of localized graph training data

    Qi Zhu, Natalia Ponomareva, Jiawei Han, and Bryan Perozzi. Shift-robust gnns: Overcoming the limitations of localized graph training data. Advances in Neural Information Processing Systems, 34, 2021

    2021

  67. [75]

    A comprehensive survey on transfer learning

    Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A comprehensive survey on transfer learning. Proceedings of the IEEE, 109:43–76, 2021

    2021

  68. [76]

    Prog: A graph prompt learning benchmark

    Chenyi Zi, Haihong Zhao, Xiangguo Sun, Yiqing Lin, Hong Cheng, and Jia Li. Prog: A graph prompt learning benchmark. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024. 15 A Appendix A.1 Algorithm Algorithm 1 presents our...

    2024

  69. [77]

    All the results are reported for both GCN and GAT when we have distribution shifts based on edge homophily for graph classification and PR node classification. Graph classification results for both GCN and GAT base models in Tables 13 illustrate the superior performance of UGP...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.