Pith. sign in

REVIEW 4 major objections 6 minor 47 references

DAGPrompT: Pushing the Limits of Graph Prompting with a Distribution-aware Graph Prompt Tuning Approach

T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper claims graph prompting can be pushed to heterophily graphs by low-rank encoder adaptation plus hop-specific class prompts.

desk verdict A plausible and well-ablated prompting method whose SOTA claim is undercut by an under-specified hyperparameter-selection protocol; deserves review, not blind trust. read the letter →

arxiv 2501.15142 v1 pith:XKKIUQQB submitted 2025-01-25 cs.LG cs.AI

Qin Chen , Liang Wang , Bo Zheng , Guojie Song This is my paper · ORCID
classification cs.LGcs.AI
keywords graphneuralnetworkspromptingfew-shotlearningheterophilygraphslow-rankadaptationprompttuningnodeclassification
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

Existing graph prompting methods freeze the pre-trained GNN encoder and therefore fail on heterophily graphs, where connected nodes often have different labels and different hops carry different class distributions. The paper proposes DAGPrompT, which tunes the frozen encoder through a low-rank adaptation of both the projection matrices and the message-passing adjacency, and couples this with hop-specific class prompts that are combined by learned coefficients. The claim, supported by experiments on ten datasets and fourteen baselines, is that this distribution-aware design improves few-shot node and graph classification accuracy by up to 4.79 percent while keeping tunable parameters in the thousands. If the claim holds, graph prompting can be extended to complex, heterophily-rich graphs without full fine-tuning.

What carries the argument

GLoRA (Graph Low-Rank Adaptation) is the module that carries the encoder-adaptation claim: it injects low-rank matrices into each GNN layer, $H^{(\ell)} = (A + P_A^{(\ell)} Q_A^{(\ell)\top}) H^{(\ell-1)} (W_0^{(\ell)} + P^{(\ell)}Q^{(\ell)\top})$, so that message passing can up- or down-weight edges and the projection can shift to the downstream distribution while the original weights $W_0$ stay frozen. The second machinery is Hop-specific Graph Prompting: the model keeps the intermediate embeddings $H^{(0)},\ldots,H^{(L)}$, creates layer-specific class tokens $P_c^{(\ell)}$ as the mean of training embeddings plus a learnable prompt, computes cosine similarity at each hop, and combines the hop scores with learnable coefficients $\gamma^{(\ell)}$ initialized as $\alpha(1-\alpha)^\ell$. Together these two mechanisms are what the paper claims makes prompting work on heterophily graphs: the first gives the encoder enough flexibility to separate different labels, and the second lets each hop contribute according to its own distribution.

What would settle it

Train DAGPrompT on Texas with 5-shot labels, then freeze the hop-prompt module and randomly permute the learned adjacency-adjustment weights while keeping all other parameters fixed; if accuracy does not drop materially, the claim that GLoRA's message-passing adaptation drives the heterophily gain is not supported. A complementary check is to test whether the learned edge weights assign higher scores to same-label edges than the unmodified adjacency does.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that a pre-trained GNN can be prompted effectively on heterophily graphs if the encoder is lightly adapted rather than left fully frozen. The GLoRA module writes the adapted layer as $H^{(\ell)} = (A + P_A^{(\ell)} Q_A^{(\ell)\top}) H^{(\ell-1)} (W_0^{(\ell)} + P^{(\ell)}Q^{(\ell)\top})$, with rank-1 factors on the adjacency and low-rank factors on the projection, so the model reweights which edges matter for message passing and adjusts feature projections while keeping the pre-trained weights intact. The hop-specific prompting module collects embeddings from every layer, builds a per-layer class prototype plus a learnable prompt for each class, scores each node against these prototypes at every hop by cosine similarity, and fuses the hop scores with learnable coefficients initially set to $\alpha(1-\alpha)^\ell$. The paper reports that this combination reaches state-of-the-art accuracy on all ten datasets, with the largest gains on strongly heterophily graphs such as Texas (up to 4.79 percent) and Chameleon, and that removing either module degrades performance.

Load-bearing premise

The load-bearing premise is that a one-dimensional low-rank adjustment to the graph's adjacency, reduced in Appendix A to a single learned edge-weight vector applied only to edges connected to training-set nodes, is expressive enough to reshape message passing so that nodes of different classes become separable.

Editorial extensions

If this is right

  • Frozen-encoder prompting is not the only viable graph-prompting design; a lightly tuned encoder with low-rank updates extends prompting to heterophily graphs.
  • Hop-wise class tokens and learned hop weights let a prompting method exploit the fact that first-hop and later-hop neighborhoods can have opposite label patterns.
  • The method keeps the parameter-efficiency advantage of prompting: on Chameleon it uses roughly 6.4K tunable parameters at the downstream stage, compared with 331K for full fine-tuning.
  • Cross-domain pre-training transfers: pre-training on Texas and then prompting on Cornell, Wisconsin, or Chameleon outperforms training from scratch on those targets.
  • The full-shot evaluation shows the same qualitative pattern, with DAGPrompT improving accuracy by 3.79 percent on Chameleon even when labels are plentiful.

Reading between the lines

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

  • An editorial reading of the ablation table is that the hop-specific prompting module is the more consistently load-bearing component: removing GLoRA costs 1.9 to 2.5 points on the reported datasets, while removing the layer-specific prompts or the learned hop coefficients also costs between 0.8 and 3.5 points, so the exact division of credit is dataset-dependent.
  • A natural testable extension is to attach GLoRA to a heterophily-specific GNN backbone; if the gains there are smaller, the benefit of low-rank adjacency adaptation may be specific to GCN-style encoders whose message passing assumes homophily.
  • Because Appendix A reduces the adjacency update to a single edge-weight vector applied only to training-set edges, an implied conjecture is that most of GLoRA's message-passing benefit comes from reweighting edges incident to labeled nodes; checking performance when the edge weights are applied to all edges would test this directly.
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

4 major / 6 minor

Summary. The paper proposes DAGPrompT, a graph prompt tuning method for few-shot node and graph classification, with two components: (i) GLoRA, a low-rank adaptation of the GNN projection matrices and the message-passing adjacency matrix (Eq. 4), and (ii) hop-specific layer-wise class prompts (Eqs. 5-8) that combine per-hop similarity scores with learned coefficients. The method is evaluated on 10 datasets against 14 baselines, reporting state-of-the-art accuracy with improvements up to 4.79% on node classification (Table 3) and up to 5.89% on graph classification (Table 4), along with ablations (Table 7), transfer experiments (Table 5), efficiency analysis (Table 6), and additional experiments in the appendix.

Significance. If the empirical results are taken at face value, the paper makes a useful contribution: it extends graph prompting to heterophily graphs, a setting where prior prompting methods freeze the encoder and underperform, and it does so with a small number of tunable parameters (about 6.4K on Chameleon) while maintaining competitive efficiency. The ablations give some evidence that both the low-rank adaptation and the hop-specific prompts contribute, and the transfer experiment (Table 5) shows a clear gain from cross-domain pre-training. The code is announced as available at GitHub, and the appendix provides extended experiments with a GAT backbone, full-shot settings, and parameter analyses. However, the central empirical claim currently rests on an incompletely described hyperparameter-selection procedure, and several reported gains are within one standard deviation of the best baselines, which makes the headline result not yet fully supported.

major comments (4)
  1. [Section 5.2] The paper states that 'Hyperparameters are selected based on performance' using an 'end-to-end' grid search, but it does not describe any held-out validation split or nested cross-validation. If the hyperparameters (alpha, rank r, learning rate, weight decay, and possibly temperature) were selected on the downstream test labels or transductive labels, every entry in Tables 3 and 4 would be optimistically biased. Since the paper's central claim is state-of-the-art accuracy, please specify the exact selection protocol, report the chosen hyperparameter values per dataset, and provide the code or a commit hash so that the protocol can be verified.
  2. [Table 3] Several headline improvements are within or close to one standard deviation of the best baseline, e.g., Texas 5-shot (80.64±3.75 vs GCOPE 75.85±2.36), Cornell 10-shot (85.13±2.08 vs GPR-GNN 82.87±4.02), and Wisconsin 5-shot (73.86±2.45 vs GCOPE 71.45±2.86). The paper does not report the number of seeds, significance tests, or confidence intervals. The claim of 'consistently outperforming' would be much more convincing with paired significance tests across the same splits and seeds, or at least a statement of how many runs were averaged.
  3. [Appendix A] Appendix A states that for extremely large graphs P_A Q_A^T is reduced to a single edge-weight vector applied only to edges connected to nodes in the training set. This is a substantial restriction of the GLoRA message-passing adaptation, yet the paper does not analyze how such a local, scalar reweighting can reshape message passing for heterophily on large graphs such as Arxiv-year. Please discuss the expressiveness of this reduction, and clarify whether test nodes that are not adjacent to any training node receive any message-passing adaptation; if the reduction is used in the reported Arxiv-year experiments, its effect on the results should be evaluated explicitly.
  4. [Section 4] Theorem 1 is a standard Rademacher generalization bound for a generic hypothesis class; it does not itself compare GLoRA with full-parameter fine-tuning. The assertion that 'low-rank adaptations like GLoRA reduce model complexity' is not instantiated: no bound on R_D(H) is derived for the specific hypothesis class of low-rank adapted GNNs, and the bound holds for any hypothesis class of bounded complexity. The claim that 'in few-shot settings, low-rank adaptation proves to be more effective' is therefore not proven by Theorem 1. Please either provide a specific complexity measure for the GLoRA parameterization or present the theoretical statement as a heuristic interpretation of the standard bound, with the experimental comparison in Section E.6 as the primary evidence.
minor comments (6)
  1. [Abstract and Introduction] The abstract reports an improvement 'up to 4.79%' while Section 1 claims 'up to 7.55%' and Table 4 shows a 5.89% improvement on Texas* for graph classification; please reconcile these numbers and state the setting for the maximum improvement.
  2. [Eq. (8) and Eq. (9)] In Eq. (8), the summation for tilde-S is over l=1 to L, while the layer collection in Eq. (5) and the loss in Eq. (9) are over l=0 to L. Please align the indexing or explain why the l=0 term is omitted in the final scoring.
  3. [Algorithm 1] In Algorithm 1, after computing H <- f(X, A; Theta), the line 'H <- matmul(H, A)' appears redundant or possibly erroneous, since message passing is already performed inside the GNN forward; please clarify or remove it.
  4. [References] Reference [6] lists only 'Diederik' as the author of the Adam paper; please use the full standard citation.
  5. [Appendix E] The appendix uses the name 'DAGPrompt' in Tables 8-10 and in Section E.2, while the main text uses 'DAGPrompT'; please standardize the naming throughout.
  6. [General] The paper states that 'Codes are available at GitHub' but does not provide a URL or commit hash throughout the manuscript; please include a working link.

Circularity Check

2 steps flagged · score 6.0 of 10

The method equations are internally consistent, but the headline SOTA accuracy is selected on downstream performance with no held-out split, and the hop-preference conclusion in Appendix E.3 restates the tuned alpha rather than an independent finding.

  1. fitted input called prediction [Section 5.2, 'Involved Baselines & Settings' (also Section 5.1, 'Datasets and Settings')]
    "We choose 𝑟 in{8, 16, 32}, and 𝛼 in{0.1, 0.3, 0.5, 0.7, 0.9}. Hyperparameters are selected based on performance. We use an "end-to-end" approach by grid-searching various settings, applying them during pre-training, and evaluating performance on downstream tasks with the same setting."

    The few-shot experiments are described without any held-out validation split (Section 5.1 only states that accuracy is measured on five-shot and ten-shot settings). The grid search for learning rate, weight decay, rank r, and alpha is therefore conducted against the same downstream accuracies that are then reported as the model's performance. The headline improvement of up to 4.79% is the maximum over configurations selected by those very labels, not an independent prediction of a pre-specified configuration. The reported SOTA numbers are statistically forced by the selection criterion rather than by the method alone.

  2. fitted input called prediction [Appendix E.3, 'Parameter Analysis' (referencing Equations 7-8 and Section 5.2)]
    "The results show that Texas and Cornell generally perform better with larger 𝛼 values, up to 0.9, while Chameleon favors smaller values, down to 0.1. This indicates that Texas and Cornell benefit more from distant hop information (as larger 𝛼 assigns greater weight to them), whereas Chameleon relies more on local hop information with a smaller 𝛼."

    The hop-weighting conclusion is a direct restatement of the tuned hyperparameter alpha. Equations 7-8 define the predictor as a weighted sum of hop-wise similarities with the initialization gamma^(l)=alpha(1-alpha)^l, and alpha is chosen per dataset by downstream performance (Section 5.2). Saying that Texas 'benefits more from distant hops' is therefore just saying that the alpha value which maximized test accuracy was large; no independent hop-distribution measurement or out-of-sample validation supports the claim. The narrative is fitted to the same objective it is used to explain.

full rationale

DAGPrompT's methodological derivation is largely self-contained and not circular: GLoRA is a direct low-rank perturbation of the projection matrices and adjacency (Eq. 4), the hop-specific prompts are class prototypes plus learnable offsets (Eqs. 5-8), and the training loss (Eq. 9) matches the inference-time similarity scoring. These are ordinary supervised components, not predictions derived from their own inputs. The theoretical bound (Theorem 1) is a standard Rademacher bound and is not used as a self-citation. The central comparison against external baselines (Tables 3-5) gives the paper independent content. However, two evaluation/narrative steps are circular. First, Section 5.2 states that hyperparameters are selected 'based on performance' through an 'end-to-end' grid search, and no validation split is described for the few-shot experiments; the reported accuracies are thus the selection criterion itself, making the headline SOTA gain a selected maximum rather than an independent prediction. Second, Appendix E.3 presents the tuned alpha as evidence of hop-specific preferences, which is merely the chosen hyperparameter restated as an empirical discovery. These issues do not make the method vacuous, but they mean the strongest empirical claim, the up-to-4.79% improvement, is not independently grounded as written.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The method rests on standard pre-training (link prediction), a standard generalization bound, and three domain assumptions about heterophily and hop structure. The main ad hoc element is the GLoRA adjacency perturbation, whose shape is justified by analogy to LoRA rather than by analysis, and whose practical form for large graphs is an undocumented approximation. The tuned hyperparameters (alpha, rank, tau) are the fitted quantities that the narrative partly interprets as findings. No new physical or formal entities are postulated; pseudo-nodes and pseudo-graphs in Section 3.2.2 are explanatory devices inherited from prior prompting work and carry no independent falsifiable handle.

free parameters (4)
  • alpha (hop-weight prior) = grid {0.1, 0.3, 0.5, 0.7, 0.9}, chosen per dataset
    Initializes the hop coefficients as gamma(l) = alpha(1-alpha)^l in Eq. 8; selected based on performance (Section 5.2), and Section E.3 interprets the chosen value as evidence of hop-level importance, making it a fitted quantity inside the paper's narrative.
  • GLoRA rank r = {8, 16, 32}, chosen per dataset
    Controls the capacity of the low-rank adaptation in Eq. 4; selected by grid search (Sections 5.2 and E.3) and directly trades off expressiveness against the generalization argument in Section 4.
  • temperature tau = not reported
    Used in the pre-training loss (Eq. 3) and the downstream loss (Eq. 9); its value is not stated, and it scales all similarity logits, so it is a hidden fitted constant.
  • learning rate and weight decay = eta in {0.1, ..., 10} x 10^-4; decay in {0, 2.5, 5} x 10^-6
    Grid-searched per dataset (Section 5.2); standard hyperparameters, but they are selected on downstream performance without a described validation split, which can inflate reported accuracy.
assumptions (5)
  • standard math Standard Rademacher generalization bound (Eq. 13)
    Theorem 1 restates the bound from Shalev-Shwartz and Ben-David [23]; used as the theoretical basis for preferring low-rank adaptation in Section 4.
  • domain assumption Link-prediction pre-training maps connected nodes to similar embeddings, which harms downstream class separation on heterophily graphs
    Motivating premise of the whole paper (Section 1, Figure 1); plausible and partially supported by Figure 2, but asserted rather than measured, and it justifies why the encoder must be adapted at all.
  • domain assumption Hop-wise distributions differ and per-hop decoupling with layer-wise prototypes is the right inductive bias
    Supported by Figure 6 (hop-level homophily histograms), but the jump from 'distributions differ' to 'layer-wise cosine prototypes with learned weights help' is assumed; the ablation in Table 7 gives only indirect support.
  • ad hoc to paper A rank-1 adjacency perturbation P_A Q_A^T, collapsed for large graphs to a single edge-weight vector applied only to training-adjacent edges, suffices to adapt message passing
    Eq. 4 and Appendix A; no expressiveness analysis is given for the adjacency adaptation, and the reduction to training-set edges is an undocumented approximation that could weaken the adaptation on large graphs.
  • domain assumption Baseline comparability under mixed backbones
    GCN backbone for all methods except GCOPE (FAGCN) and the heterophily-specific methods (H2GCN, GPR-GNN, ALT-GNN), as described in Section 5.2; the comparison is standard in the field but does not isolate the prompting contribution.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DAGPrompT: Pushing the Limits of Graph Prompting with a Distribution-aware Graph Prompt Tuning Approach." pith.science (2026). https://pith.science/paper/XKKIUQQB

@misc{pith2026250115142,
  author       = {Pith},
  title        = {Pith review of: DAGPrompT: Pushing the Limits of Graph Prompting with a Distribution-aware Graph Prompt Tuning Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XKKIUQQB}},
  note         = {Machine review of arXiv:2501.15142}
}
read the original abstract

The pre-train then fine-tune approach has advanced GNNs by enabling general knowledge capture without task-specific labels. However, an objective gap between pre-training and downstream tasks limits its effectiveness. Recent graph prompting methods aim to close this gap through task reformulations and learnable prompts. Despite this, they struggle with complex graphs like heterophily graphs. Freezing the GNN encoder can reduce the impact of prompting, while simple prompts fail to handle diverse hop-level distributions. This paper identifies two key challenges in adapting graph prompting methods for complex graphs: (1) adapting the model to new distributions in downstream tasks to mitigate pre-training and fine-tuning discrepancies from heterophily and (2) customizing prompts for hop-specific node requirements. To overcome these challenges, we propose Distribution-aware Graph Prompt Tuning (DAGPrompT), which integrates a GLoRA module for optimizing the GNN encoder's projection matrix and message-passing schema through low-rank adaptation. DAGPrompT also incorporates hop-specific prompts accounting for varying graph structures and distributions among hops. Evaluations on 10 datasets and 14 baselines demonstrate that DAGPrompT improves accuracy by up to 4.79 in node and graph classification tasks, setting a new state-of-the-art while preserving efficiency. Codes are available at GitHub.

Figures

Figures reproduced from arXiv: 2501.15142 by the authors.

Figure 2
Figure 2. Conventional graph prompting techniques are less [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 1
Figure 1. Heterophily diminishes the effectiveness of prompt [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. The framework of Distribution-aware Graph Prompt Tuning. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Impact of data heterophily on Syn-Chameleon. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Impact of shots on Texas and Chameleon. effectiveness of prompt-based methods in learning embeddings. (iii) Fewer labels for training significantly hinder models trained from scratch, especially in heterophily settings. On graphs with smaller label ratios, Chameleon, S…
Figure 6
Figure 6. Figure 6: Hop-wise Local Heterophily Distributions. [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Parameter analysis. E Additional Experiments E.1 DAGPrompt with GAT as Backbone To evaluate the generalization capability of DAGPrompt, we con￾duct 5-shot and 10-shot node classification experiments using GAT as the backbone encoder2 . The results in [PITH_FULL_IMAGE:…
Figure 9
Figure 9. Figure 9: Loss curve of DAGPrompT on Chameleon. 𝛼 values, up to 0.9, while Chameleon favors smaller values, down to 0.1. This indicates that Texas and Cornell benefit more from distant hop information (as larger 𝛼 assigns greater weight to them), whereas Chameleon relies more on…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 16 canonical work pages

  1. [1]

    Armen Aghajanyan, Sonal Gupta, and Luke Zettlemoyer. 2021. Intrinsic Di- mensionality Explains the Effectiveness of Language Model Fine-Tuning. In Proceedings of the 59th Annual Meeting of the Association for Computational Lin- guistics and the 11th International Joint Conference on Natural Language Processing, ACL/IJCNLP 2021, (Volume 1: Long Papers), Vi...

    work page 2021

  2. [2]

    Deyu Bo, Xiao Wang, Chuan Shi, and Huawei Shen. 2021. Beyond low-frequency information in graph convolutional networks. In Proceedings of the AAAI confer- ence on artificial intelligence , Vol. 35. 3950–3957

    work page 2021

  3. [3]

    Ming Chen, Zhewei Wei, Zengfeng Huang, Bolin Ding, and Yaliang Li. 2020. Simple and deep graph convolutional networks. In International conference on machine learning. PMLR, 1725–1735

    2020

  4. [4]

    Eli Chien, Jianhao Peng, Pan Li, and Olgica Milenkovic. 2020. Adaptive universal generalized pagerank graph neural network. arXiv preprint arXiv:2006.07988 (2020)

    arXiv 2020

  5. [5]

    Siddhartha Shankar Das, SM Ferdous, Mahantesh M Halappanavar, Edoardo Serra, and Alex Pothen. 2024. AGS-GNN: Attribute-guided Sampling for Graph Neural Networks. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 538–549

    work page 2024

  6. [6]

    P Kingma Diederik. 2014. Adam: A method for stochastic optimization. (No Title) (2014)

    2014

  7. [7]

    Taoran Fang, Yunchao Zhang, Yang Yang, Chunping Wang, and Lei Chen. 2024. Universal prompt tuning for graph neural networks. Advances in Neural Infor- mation Processing Systems 36 (2024)

    2024

  8. [8]

    Zhenyu Hou, Xiao Liu, Yukuo Cen, Yuxiao Dong, Hongxia Yang, Chunjie Wang, and Jie Tang. 2022. Graphmae: Self-supervised masked graph autoencoders. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 594–604

    2022

Show all 47 references
  1. [9]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)

    2021 arXiv

  2. [10]

    Ziniu Hu, Yuxiao Dong, Kuansan Wang, Kai-Wei Chang, and Yizhou Sun. 2020. Gpt-gnn: Generative pre-training of graph neural networks. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining. 1857–1867

    2020

  3. [11]

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

    2016 arXiv

  4. [12]

    Thomas N Kipf and Max Welling. 2016. Variational graph auto-encoders. arXiv preprint arXiv:1611.07308 (2016)

    2016 arXiv

  5. [13]

    Nils Kriege and Petra Mutzel. 2012. Subgraph matching kernels for attributed graphs. Proceedings of the 29th International Conference on Machine Learning, ICML 2012, Edinburgh, Scotland, UK, June 26 - July 1, 2012 (2012)

    2012

  6. [14]

    Xiang Li, Renyu Zhu, Yao Cheng, Caihua Shan, Siqiang Luo, Dongsheng Li, and Weining Qian. 2022. Finding global homophily in graph neural networks when meeting heterophily. In International Conference on Machine Learning . PMLR, 13242–13256

    2022

  7. [15]

    Derek Lim, Felix Hohne, Xiuyu Li, Sijia Linda Huang, Vaishnavi Gupta, Omkar Bhalerao, and Ser Nam Lim. 2021. Large Scale Learning on Non-Homophilous Graphs: New Benchmarks and Strong Simple Methods. Advances in Neural Information Processing Systems 34 (2021)

    2021

  8. [16]

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

    2023

  9. [17]

    Yuanfu Lu, Xunqiang Jiang, Yuan Fang, and Chuan Shi. 2021. Learning to pre- train graph neural networks. In Proceedings of the AAAI conference on artificial intelligence, Vol. 35. 4276–4284

    2021

  10. [18]

    Xiaojun Ma, Qin Chen, Yuanyi Ren, Guojie Song, and Liang Wang. 2022. Meta- Weight Graph Neural Network: Push the Limits Beyond Global Homophily. In Proceedings of the ACM Web Conference 2022 . 1270–1280

    2022

  11. [19]

    Yihong Ma, Ning Yan, Jiayu Li, Masood Mortazavi, and Nitesh V Chawla. 2024. Hetgpt: Harnessing the power of prompt tuning in pre-trained heterogeneous graph neural networks. In Proceedings of the ACM on Web Conference 2024 . 1015– 1023

    2024

  12. [20]

    Miller McPherson, Lynn Smith-Lovin, and James M Cook. 2001. Birds of a feather: Homophily in social networks. Annual review of sociology 27, 1 (2001), 415–444

    2001

  13. [21]

    Hongbin Pei, Bingzhe Wei, Kevin Chen-Chuan Chang, Yu Lei, and Bo Yang

  14. [22]

    Benedek Rozemberczki, Carl Allen, and Rik Sarkar. 2021. Multi-scale attributed node embedding. Journal of Complex Networks 9, 2 (2021), cnab014

    2021

  15. [23]

    Shai Shalev-Shwartz and Shai Ben-David. 2014. Understanding machine learning: From theory to algorithms . Cambridge university press

    2014

  16. [24]

    Fan-Yun Sun, Jordan Hoffmann, Vikas Verma, and Jian Tang. 2019. Infograph: Un- supervised and semi-supervised graph-level representation learning via mutual information maximization. arXiv preprint arXiv:1908.01000 (2019)

    2019 arXiv

  17. [25]

    Mingchen Sun, Kaixiong Zhou, Xin He, Ying Wang, and Xin Wang. 2022. 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. 1717–1727

    2022

  18. [26]

    Xiangguo Sun, Hong Cheng, Jia Li, Bo Liu, and Jihong Guan. 2023. All in one: Multi-task prompting for graph neural networks. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 2120–2131

    2023

  19. [27]

    Petar Velickovic, William Fedus, William L Hamilton, Pietro Liò, Yoshua Bengio, and R Devon Hjelm. 2019. Deep graph infomax. ICLR (Poster) 2, 3 (2019), 4

    2019

  20. [28]

    Shaowen Wang, Linxi Yu, and Jian Li. 2024. LoRA-GA: Low-Rank Adaptation with Gradient Approximation. arXiv preprint arXiv:2407.05000 (2024)

    2024 arXiv

  21. [29]

    Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and S Yu Philip. 2020. A comprehensive survey on graph neural networks. IEEE transactions on neural networks and learning systems 32, 1 (2020), 4–24

    2020

  22. [30]

    Zhe Xu, Yuzhong Chen, Qinghai Zhou, Yuhang Wu, Menghai Pan, Hao Yang, and Hanghang Tong. 2023. Node classification beyond homophily: Towards a general solution. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 2862–2873

    2023

  23. [31]

    Yuchen Yan, Peiyan Zhang, Zheng Fang, and Qingqing Long. 2024. Inductive Graph Alignment Prompt: Bridging the Gap between Graph Pre-training and Inductive Fine-tuning From Spectral Perspective. In Proceedings of the ACM on Web Conference 2024. 4328–4339

    2024

  24. [32]

    Zhilin Yang, William Cohen, and Ruslan Salakhudinov. 2016. Revisiting semi- supervised learning with graph embeddings. In International conference on ma- chine learning. PMLR, 40–48

    2016

  25. [33]

    Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. 2020. Graph contrastive learning with augmentations. Advances in neural information processing systems 33 (2020), 5812–5823

    2020

  26. [34]

    Xingtong Yu, Yuan Fang, Zemin Liu, and Xinming Zhang. 2024. Hgprompt: Bridging homogeneous and heterogeneous graphs for few-shot prompt learning. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 38. 16578– 16586

    2024

  27. [35]

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

    2024

  28. [36]

    Xingtong Yu, Jie Zhang, Yuan Fang, and Renhe Jiang. 2024. Non-homophilic graph pre-training and prompt learning. arXiv preprint arXiv:2408.12594 (2024)

    2024 arXiv

  29. [37]

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

    2024

  30. [38]

    Xingtong Yu, Chang Zhou, Yuan Fang, and Xinming Zhang. 2024. Text-Free Multi-domain Graph Pre-training: Toward Graph Foundation Models. arXiv preprint arXiv:2405.13934 (2024)

    2024 arXiv

  31. [39]

    Jiaqi Zeng and Pengtao Xie. 2021. Contrastive self-supervised learning for graph classification. In Proceedings of the AAAI conference on Artificial Intelligence , Vol. 35. 10824–10832

    2021

  32. [40]

    Haihong Zhao, Aochuan Chen, Xiangguo Sun, Hong Cheng, and Jia Li. 2024. All in one and one for all: A simple yet effective method towards cross-domain graph pretraining. arXiv preprint arXiv:2402.09834 (2024)

    2024 arXiv

  33. [41]

    Zhengyang Zhou, Qihe Huang, Gengyu Lin, Kuo Yang, Lei Bai, and Yang Wang

  34. [42]

    Jiong Zhu, Yujun Yan, Lingxiao Zhao, Mark Heimann, Leman Akoglu, and Danai Koutra. 2020. Beyond homophily in graph neural networks: Current limitations and effective designs. Advances in neural information processing systems 33 (2020), 7793–7804

    2020

  35. [43]

    Yun Zhu, Yaoke Wang, Haizhou Shi, Zhenshuo Zhang, Dian Jiao, and Siliang Tang. 2024. GraphControl: Adding Conditional Control to Universal Graph Pre- trained Models for Graph Domain Transfer Learning. In Proceedings of the ACM on Web Conference 2024. 539–550

    2024

  36. [44]

    Yanqiao Zhu, Yichen Xu, Feng Yu, Qiang Liu, Shu Wu, and Liang Wang. 2020. Deep graph contrastive representation learning. arXiv preprint arXiv:2006.04131 (2020). A Algorithm of DAGPrompT We detail the DAGPrompT algorithm from pre-training to prompt- ing in Algorithms algorithm...

    2020 arXiv

  37. [47]

    Labels for Texas, Cornell, and Wisconsin represent web page cate- gories, while Chameleon and Squirrel labels capture average monthly web traffic, grouped into five ranges

    consist of web pages, where nodes represent individual pages, node features are word embeddings, and edges reflect hyperlinks. Labels for Texas, Cornell, and Wisconsin represent web page cate- gories, while Chameleon and Squirrel labels capture average monthly web traffic, gro...

    2025

  38. [2020]

    arXiv preprint arXiv:2002.05287 (2020)

    Geom-gcn: Geometric graph convolutional networks. arXiv preprint arXiv:2002.05287 (2020)

    2020 arXiv

  39. [2023]

    In The eleventh international conference on learning representations

    Greto: Remedying dynamic graph topology-task discordance via target homophily. In The eleventh international conference on learning representations

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.