Pith. sign in

REVIEW 4 major objections 6 minor 35 references

HGMP:Heterogeneous Graph Multi-Task Prompt Learning

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A single heterogeneous-graph pre-training and prompt framework, HGMP, reports the best 10-shot node, edge, and graph classification results on ACM and IMDB among nine baselines.

desk verdict Plausible new combination of known techniques for few-shot heterogeneous graph learning; the engineering is sound in outline, but the empirical 'significant outperformance' claim is not backed by the reported statistics and there are a couple of notation-level defects that need fixing before the results can be trusted. read the letter →

arxiv 2507.07405 v1 pith:CYY47YCB submitted 2025-07-10 cs.LG cs.AI

Pengfei Jiao , Jialong Ni , Di Jin , Xuan Guo , Huan Liu , Hongjiang Chen , Yanxian Bi This is my paper
classification cs.LGcs.AI
keywords heterogeneousgraphneuralnetworkspromptlearningcontrastivepre-trainingfew-shotmulti-taskclassificationnodeedge
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper sets out to show that prompt learning can span all main heterogeneous graph tasks rather than one task format. HGMP first rewrites node and edge classification as graph classification on $\tau$-hop induced subgraphs, then pre-trains a heterogeneous graph encoder with a graph-level contrastive objective, and finally adapts with per-type multiplicative feature prompts while the encoder stays frozen. The experiments report the best micro- and macro-F1 among nine baselines for 10-shot node, edge, and graph classification on ACM and IMDB. If that result stands, a single pre-trained heterogeneous encoder plus small prompt vectors could replace task-by-task fine-tuning in few-shot settings.

What carries the argument

The machinery of HGMP has three linked pieces. Task reformulation builds $\tau$-hop induced subgraphs—for a node, the target node plus its neighbors and typed edges; for an edge, both endpoints and their neighborhoods—and maps labels onto those subgraphs so node and edge classification become graph classification. The pre-training stage contrasts these subgraphs using heterogeneous node masking and heterogeneous edge permutation, with the number of masked nodes or permuted edges per type set by $r \cdot \mathrm{count}(i)^2 / \sum_j \mathrm{count}(j)^2$. The final piece is the Heterogeneous Graph Prompt Feature (HGPF): for each node type $T$, a learnable vector $p_T$ of the same dimension as that type's features, applied as $\hat{x}_i = x_i \times p_{T_i}$ so a single frozen encoder can be steered toward different downstream tasks.

What would settle it

Check the released code's contrastive loss: if it matches the Section 4.3 equation, training maximizes similarity between a positive and a negative augmented view, the opposite of the stated pre-training goal, and the reported accuracy would not be reproducible as described. If the code uses the standard positive-pair numerator, re-running the 10-shot protocol on ACM and IMDB would settle whether HGMP in fact achieves the best micro- and macro-F1 among the nine baselines.

Watch

Extended reading notes

Core claim

The paper's central claim is that the gap between pre-training and downstream tasks in heterogeneous graphs closes when both are expressed as graph-level problems. Node and edge labels are attached to induced subgraphs built from the target object's $\tau$-hop neighborhood, so every task becomes graph classification. A graph-level contrastive pre-training stage then operates on two type-aware augmentations—heterogeneous node masking and heterogeneous edge permutation—whose per-type ratios are adjusted by squared type counts. Downstream, the encoder is kept frozen and each node type receives a learnable Heterogeneous Graph Prompt Feature (HGPF), a vector multiplied elementwise into that type's raw features, which the paper argues reshapes the input toward the pre-training objective. Across ACM and IMDB, the paper reports that HGMP achieves the best micro- and macro-F1 on 10-shot node, edge, and graph classification among nine baselines.

Load-bearing premise

The load-bearing assumption is that the contrastive-loss formula printed in Section 4.3 is a typo: as written, it rewards making a positive augmented view and a negative augmented view similar, the opposite of the described goal, so the reported results can only follow if the intended loss pairs two views of the same graph.

Editorial extensions

If this is right

  • A frozen heterogeneous encoder can serve node, edge, and graph classification simultaneously through the induced-subgraph reformulation and small prompt vectors, so multi-task deployment no longer needs a separate fine-tuned model per task.
  • The ablation results indicate per-type multiplicative feature prompts contribute more than the heterogeneous augmentations alone, making prompt design the primary cheap lever for few-shot heterogeneous graph performance.
  • Because HGMP improves over supervised training on GCN, GAT, and HGT backbones in node and edge classification, the recipe is not tied to one encoder architecture and can be carried onto stronger backbones.
  • Graph-level pre-training over induced subgraphs creates a shared task space that makes heterogeneous graph prompt methods comparable across node, edge, and graph classification on the same protocol.

Reading between the lines

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

  • An untested consequence of the squared-count adjustment is that augmentation concentrates on the most frequent node and edge types; comparing it with uniform or entropy-weighted ratios on skewed graphs, such as ACM's 56 subject nodes, would show whether the adjustment helps or merely tracks type frequency.
  • Because HGPF multiplies features by a prompt vector, zero-valued features cannot be shifted and negative features get rescaled rather than translated; an additive or residual prompt variant would separate the benefit of type-awareness from the choice of multiplicative combination.
  • The paper converts IMDB's multi-label graph classification into single-label classification and notes the information loss; a multi-label-capable prompt head is a direct extension that would reveal how much of the IMDB graph-classification gap is an artifact of that conversion.
  • The $\tau$-hop induced-subgraph construction grows with the neighborhood radius, so the reported protocol is only tested at small dataset scale; neighborhood sampling would be needed before claiming the same recipe transfers to large heterogeneous graphs.
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 HGMP, a heterogeneous graph multi-task prompt-learning framework. It reformulates node and edge classification into graph-level classification via induced subgraphs, pre-trains a GNN encoder with contrastive learning on heterogeneous graphs using type-weighted node masking and edge permutation augmentations, and then freezes the encoder and applies per-node-type multiplicative feature prompts to align with downstream tasks. Experiments on ACM and IMDB under 10-shot settings report micro-/macro-F1 for node, edge, and graph classification, together with ablations and backbone studies. The central claim is that HGMP adapts well and significantly outperforms baseline methods.

Significance. If the results hold, the contribution is a coherent recipe for unifying heterogeneous graph tasks at the graph level and for adapting frozen heterogeneous pre-trained encoders with cheap type-specific prompts; the graph-level contrastive pre-training with heterogeneity-aware augmentations is a sensible direction. The paper is clearly organized and includes ablation and backbone-flexibility checks. However, the strength of the empirical claim is currently unverified: reported margins are often within one standard deviation, no significance testing is provided, and the promised supplement/code is absent. The manuscript also contains a load-bearing inconsistency in the stated contrastive loss. With corrected notation and stronger statistical evidence, the framework could be a useful addition to the graph prompt-learning literature.

major comments (4)
  1. [Section 4.3, Eq. (Lcon)] The loss as written is Lcon = -\Sigma_i log[ exp(sim(z+_i, z-_i)/\tau) / \Sigma_j \mathbb{1}[i\neq j] exp(sim(z+_i, z_j)/\tau) ]. If z-_i denotes the negative representation, this maximizes the positive-negative similarity and inverts the intended objective; if z-_i is instead meant to be the positive view, the notation is seriously misleading. Please state the exact standard InfoNCE form and define z+ and z- unambiguously, since the pre-training objective is a central component of the method.
  2. [Tables 2-4 and Section 5.1] The claim that HGMP "significantly outperforms" baselines is not supported by the reported statistics. No significance tests, number of seeds, or per-seed results are given, and several best-vs-second margins are within overlapping standard deviations (e.g., IMDB node classification: HGMP 68.76±0.82 vs GCN 68.66±1.71; ACM graph classification: HGMP 98.30±0.45 vs GCN 98.16±0.58; ACM node classification: HGMP 84.60±3.18 vs HeCo 83.99±1.74). The supplement with hyperparameters promised in Section 5.1 is not available, so reproducibility cannot be checked. Please report a fixed number of seeds, add paired significance tests, and make code/supplement available.
  3. [Section 4.4] The "Why It Works" subsection invokes Fang et al.'s existence theorem for additive prompts (X + p*), but HGPF uses multiplicative prompts (\hat x_i = x_i \times p_{T_i}), and the text earlier says the prompt can be added or multiplied. The additive theorem does not automatically justify multiplicative prompting. Either provide a formal argument for the multiplicative case or present HGPF as a heuristic design choice supported by the ablation study rather than by that theorem.
  4. [Section 5.1 and Tables 2-4] HDMI is listed among the pre-train/fine-tune baselines, but it appears in none of the result tables or analyses. Either include its results or remove it from the baseline list, so that the claimed comparison against nine baselines is complete.
minor comments (6)
  1. [Section 4.1] The text says "the overall framework of HGMF" where the model is named HGMP; this should be corrected.
  2. [Section 4.3, Eq. (Lcon)] The notation "P2N" and "⊮" in the denominator is malformed; please use proper summation symbols and the indicator function \mathbb{1} with a correctly indexed sum over the 2N augmented graphs.
  3. [Section 5.5] The ablation section says all variants were trained using "the reconstruction task template based on induced subgraphs," but no reconstruction pre-training task is defined in Section 4; please clarify what this template is, or correct the description to contrastive pre-training.
  4. [Tables 5 and 6] There are formatting issues in the tables: "V ARIANT" has an unintended space and "Backbonemethod" is run together.
  5. [Section 5.4] The shot-curve figure is referenced as "Figs 3" and no information is given about how many runs are averaged or whether error bars are shown; please clarify the experimental protocol.
  6. [Section 4.4, Eq. (3)] The multiplication in \hat x_i = x_i \times p_{T_i} should be explicitly defined as elementwise multiplication, and the earlier sentence mentioning addition should be reconciled with the multiplicative definition.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the empirical claim is tested against external baselines and the author-overlapping citations are background only.

full rationale

The central claim (HGMP outperforms nine baselines in 10-shot node, edge, and graph classification) is evaluated on public ACM/IMDB data against external methods; no reported F1 is a fitted parameter renamed as a prediction. The graph-level task reformulation is explicitly adapted from prior homogeneous-graph prompt work (ProG) and is applied uniformly to all compared methods, so it does not by construction determine HGMP's superiority. The only author-overlapping references (Tan et al. 2023 and Zhang et al. 2022, both involving co-author Huan Liu) appear in background lists and carry none of the derivation. Section 4.4's 'Why It Works' cites Fang et al.'s additive-prompt theorem while HGPF uses multiplicative prompts; that is an unsupported extrapolation and a theoretical-validity gap, but the reported results do not derive from that theorem, so it is not a circular reduction. Section 4.3's contrastive-loss formula appears inconsistent with its prose (numerator uses a positive-negative pair); this is a correctness/typo issue, not a self-referential derivation. With no equation reducing to its own input and no load-bearing self-citation, the appropriate finding is no significant circularity.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The method introduces no new physical or conceptual entities. The free parameters are standard hyperparameters whose values are not reported. The axioms include the core transfer assumption of pre-training, an ad hoc augmentation weighting, and an unproven extrapolation from additive to multiplicative prompts. The paper does not provide independent evidence for the multiplicative-prompt assumption beyond the end-task accuracy numbers.

free parameters (3)
  • augmentation ratio r = not reported in main text
    Controls the fraction of nodes and edges masked or permuted during pre-training; no sensitivity analysis or chosen value is given.
  • contrastive temperature tau = not reported
    Temperature in the contrastive loss L_con; value affects gradient behavior and is not specified.
  • hop radius tau for induced subgraphs = not reported
    Determines the size of the induced subgraphs that reformulate node and edge tasks into graph tasks; no ablation over this value.
assumptions (5)
  • standard math InfoNCE contrastive loss is an appropriate objective for graph-level representation learning.
    Adopted from contrastive learning literature, though the formula as written appears to contain a sign/sample error.
  • domain assumption Graph-level contrastive pre-training on augmented heterogeneous graphs transfers to downstream node, edge, and graph classification.
    Core premise of the method; no theoretical or empirical isolation of this transfer is provided beyond end-task results.
  • ad hoc to paper Squared-type-count weighting in augmentation is a good way to allocate masking and permutation across node and edge types.
    Introduced in Section 4.3 without motivation or ablation over alternative weightings.
  • domain assumption tau-hop induced subgraphs with target labels preserve the information needed for downstream classification.
    Adopted from homogeneous graph prompt methods (ProG, GraphPrompt) and extended to heterogeneous graphs in Section 4.2.
  • ad hoc to paper A multiplicative per-type feature prompt can align a frozen pre-trained encoder with downstream tasks, by analogy to the additive-prompt theorem of Fang et al.
    Used in Section 4.4; the cited theorem covers additive prompts, not multiplication.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HGMP:Heterogeneous Graph Multi-Task Prompt Learning." pith.science (2026). https://pith.science/paper/CYY47YCB

@misc{pith2026250707405,
  author       = {Pith},
  title        = {Pith review of: HGMP:Heterogeneous Graph Multi-Task Prompt Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CYY47YCB}},
  note         = {Machine review of arXiv:2507.07405}
}
read the original abstract

The pre-training and fine-tuning methods have gained widespread attention in the field of heterogeneous graph neural networks due to their ability to leverage large amounts of unlabeled data during the pre-training phase, allowing the model to learn rich structural features. However, these methods face the issue of a mismatch between the pre-trained model and downstream tasks, leading to suboptimal performance in certain application scenarios. Prompt learning methods have emerged as a new direction in heterogeneous graph tasks, as they allow flexible adaptation of task representations to address target inconsistency. Building on this idea, this paper proposes a novel multi-task prompt framework for the heterogeneous graph domain, named HGMP. First, to bridge the gap between the pre-trained model and downstream tasks, we reformulate all downstream tasks into a unified graph-level task format. Next, we address the limitations of existing graph prompt learning methods, which struggle to integrate contrastive pre-training strategies in the heterogeneous graph domain. We design a graph-level contrastive pre-training strategy to better leverage heterogeneous information and enhance performance in multi-task scenarios. Finally, we introduce heterogeneous feature prompts, which enhance model performance by refining the representation of input graph features. Experimental results on public datasets show that our proposed method adapts well to various tasks and significantly outperforms baseline methods.

Figures

Figures reproduced from arXiv: 2507.07405 by the authors.

Figure 1
Figure 1. Induced subgraphs for nodes and edges This subgraph captures both the structural context of the tar￾get node and the type-specific semantics of its neighboring nodes and edges. For edge-level tasks, the induced subgraph includes the two endpoints of the target edge and their respec￾tive τ -hop neighbors, retaining all nodes, edges, and their types within this subgraph, as depicted in Fig 1b. This en￾sures that the s… view at source ↗
Figure 2
Figure 2. Overview of the HGMP architecture: The input graph is first subjected to heterogeneous graph augmentation, resulting in an [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Impact of shots on ACM. However, HeCo and DMGI, as they only model the features of target nodes, remain unable to capture complete graph￾level information and consequently fell out of the top three in this task. 5.3 Flexibility on backbones To validate the adaptability and performance of the proposed HGMP model across different backbones, we conducted ex￾periments using GCN, GAT, and HGT on node classification, edge… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 31 canonical work pages

  1. [1]

    Knowledge-preserving incremental social event detection via heterogeneous gnns

    [Cao et al., 2021] Yuwei Cao, Hao Peng, Jia Wu, Yingtong Dou, Jianxin Li, and Philip S Yu. Knowledge-preserving incremental social event detection via heterogeneous gnns. In Proceedings of the Web Conference 2021, pages 3383– 3395,

    work page 2021

  2. [7]

    Strategies for pre-training graph neural net- works

    [Hu et al., 2019] Weihua Hu, Bowen Liu, Joseph Gomes, Marinka Zitnik, Percy Liang, Vijay Pande, and Jure Leskovec. Strategies for pre-training graph neural net- works. arXiv preprint arXiv:1905.12265,

    arXiv 2019

  3. [8]

    Heterogeneous graph transformer

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

    work page 2020

  4. [9]

    Prodigy: Enabling in-context learning over graphs

    [Huang et al., 2024] Qian Huang, Hongyu Ren, Peng Chen, Gregor Kr ˇzmanc, Daniel Zeng, Percy S Liang, and Jure Leskovec. Prodigy: Enabling in-context learning over graphs. Advances in Neural Information Processing Sys- tems, 36,

    work page 2024

  5. [10]

    A survey on knowl- edge graphs: Representation, acquisition, and applica- tions

    [Ji et al., 2021] Shaoxiong Ji, Shirui Pan, Erik Cambria, Pekka Marttinen, and S Yu Philip. A survey on knowl- edge graphs: Representation, acquisition, and applica- tions. IEEE transactions on neural networks and learning systems, 33(2):494–514,

    work page 2021

  6. [11]

    Hdmi: High-order deep multiplex infomax

    [Jing et al., 2021] Baoyu Jing, Chanyoung Park, and Hang- hang Tong. Hdmi: High-order deep multiplex infomax. In Proceedings of the Web Conference 2021, pages 2414– 2424,

    work page 2021

  7. [12]

    Semi-supervised classification with graph convolutional networks

    [Kipf and Welling, 2016] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. International Conference on Learning Repre- sentations,

    work page 2016

  8. [14]

    The power of scale for parameter-efficient prompt tuning

    [Lester et al., 2021] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. Empirical Methods in Natural Language Processing,

    work page 2021

Show all 35 references
  1. [15]

    Prefix- tuning: Optimizing continuous prompts for generation

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

    2021

  2. [16]

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

    [Liu et al., 2023] Zemin Liu, Xingtong Yu, Yuan Fang, and Xinming Zhang. Graphprompt: Unifying pre-training and downstream tasks for graph neural networks. In Proceed- ings of the ACM Web Conference 2023 , pages 417–428,

    2023

  3. [17]

    Are we really mak- ing much progress? revisiting, benchmarking and refining heterogeneous graph neural networks

    [Lv et al., 2021] Qingsong Lv, Ming Ding, Qiang Liu, Yuxi- ang Chen, Wenzheng Feng, Siming He, Chang Zhou, Jian- guo Jiang, Yuxiao Dong, and Jie Tang. Are we really mak- ing much progress? revisiting, benchmarking and refining heterogeneous graph neural networks. In Proceeding...

    2021

  4. [18]

    Hetgpt: Harnessing the power of prompt tuning in pre-trained heterogeneous graph neural networks

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

    2024

  5. [19]

    Hinormer: Representation learning on heterogeneous information networks with graph trans- former

    [Mao et al., 2023] Qiheng Mao, Zemin Liu, Chenghao Liu, and Jianling Sun. Hinormer: Representation learning on heterogeneous information networks with graph trans- former. In Proceedings of the ACM Web Conference 2023, pages 599–610,

    2023

  6. [20]

    Weisfeiler and leman go neural: Higher-order graph neural networks

    [Morris et al., 2019] Christopher Morris, Martin Ritzert, Matthias Fey, William L Hamilton, Jan Eric Lenssen, Gau- rav Rattan, and Martin Grohe. Weisfeiler and leman go neural: Higher-order graph neural networks. In Proceed- ings of the AAAI conference on artificial intelligen...

    2019

  7. [21]

    Unsupervised attributed multiplex network embedding

    [Park et al., 2020] Chanyoung Park, Donghyun Kim, Jiawei Han, and Hwanjo Yu. Unsupervised attributed multiplex network embedding. In Proceedings of the AAAI confer- ence on artificial intelligence , volume 34, pages 5371– 5378,

    2020

  8. [22]

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

    [Sun et al., 2022] Mingchen Sun, Kaixiong Zhou, Xin He, Ying Wang, and Xin Wang. Gppt: Graph pre-training and prompt tuning to generalize graph neural networks. InPro- ceedings of the 28th ACM SIGKDD Conference on Knowl- edge Discovery and Data Mining, pages 1717–1727,

    2022

  9. [23]

    Graph prompt learn- ing: A comprehensive survey and beyond

    [Sun et al., 2023b] Xiangguo Sun, Jiawen Zhang, Xixi Wu, Hong Cheng, Yun Xiong, and Jia Li. Graph prompt learn- ing: A comprehensive survey and beyond. arXiv preprint arXiv:2311.16534,

    arXiv

  10. [24]

    Virtual node tuning for few-shot node clas- sification

    [Tan et al., 2023] Zhen Tan, Ruocheng Guo, Kaize Ding, and Huan Liu. Virtual node tuning for few-shot node clas- sification. In Proceedings of the 29th ACM SIGKDD Con- ference on Knowledge Discovery and Data Mining, pages 2177–2188,

    2023

  11. [25]

    Graph attention networks

    [Veliˇckovi´c et al., 2017] Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. arXiv preprint arXiv:1710.10903,

    2017 arXiv

  12. [27]

    Self-supervised heterogeneous graph neural network with co-contrastive learning

    [Wang et al., 2021] Xiao Wang, Nian Liu, Hui Han, and Chuan Shi. Self-supervised heterogeneous graph neural network with co-contrastive learning. In Proceedings of the 27th ACM SIGKDD conference on knowledge discov- ery & data mining, pages 1726–1736,

    2021

  13. [28]

    Simgrace: A simple framework for graph contrastive learning without data augmentation

    [Xia et al., 2022] 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 , pages 1070–1079,

    2022

  14. [30]

    From local structures to size generalization in graph neural networks

    [Yehudai et al., 2021] Gilad Yehudai, Ethan Fetaya, Eli Meirom, Gal Chechik, and Haggai Maron. From local structures to size generalization in graph neural networks. In International Conference on Machine Learning , pages 11975–11986. PMLR,

    2021

  15. [31]

    Graph con- trastive learning with augmentations

    [You et al., 2020] Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. Graph con- trastive learning with augmentations. Advances in neural information processing systems, 33:5812–5823,

    2020

  16. [32]

    Hgprompt: Bridging homogeneous and heterogeneous graphs for few-shot prompt learning

    [Yu et al., 2024] Xingtong Yu, Yuan Fang, Zemin Liu, and Xinming Zhang. Hgprompt: Bridging homogeneous and heterogeneous graphs for few-shot prompt learning. In Proceedings of the AAAI Conference on Artificial Intelli- gence, pages 16578–16586,

    2024

  17. [33]

    Graph trans- former networks

    [Yun et al., 2019] Seongjun Yun, Minbyul Jeong, Raehyun Kim, Jaewoo Kang, and Hyunwoo J Kim. Graph trans- former networks. Advances in neural information process- ing systems, 32,

    2019

  18. [34]

    Few-shot learning on graphs

    [Zhang et al., 2022] Chuxu Zhang, Kaize Ding, Jundong Li, Xiangliang Zhang, Yanfang Ye, Nitesh V Chawla, and Huan Liu. Few-shot learning on graphs. International Joint Conference on Artificial Intelligence,

    2022

  19. [35]

    Prog: A graph prompt learning benchmark

    [Zi et al., 2024] Chenyi Zi, Haihong Zhao, Xiangguo Sun, Yiqing Lin, Hong Cheng, and Jia Li. Prog: A graph prompt learning benchmark. arXiv preprint arXiv:2406.05346, 2024

    2024 arXiv

  20. [2016]

    Understanding attention and gen- eralization in graph neural networks

    [Knyazev et al., 2019] Boris Knyazev, Graham W Taylor, and Mohamed Amer. Understanding attention and gen- eralization in graph neural networks. Advances in neural information processing systems, 32,

    2019

  21. [2017]

    Hetero- geneous graph attention network

    [Wang et al., 2019] Xiao Wang, Houye Ji, Chuan Shi, Bai Wang, Yanfang Ye, Peng Cui, and Philip S Yu. Hetero- geneous graph attention network. In The world wide web conference, pages 2022–2032,

    2019

  22. [2019]

    Universal prompt tuning for graph neural networks

    [Fang et al., 2024] Taoran Fang, Yunchao Zhang, Yang Yang, Chunping Wang, and Lei Chen. Universal prompt tuning for graph neural networks. Advances in Neural In- formation Processing Systems, 36,

    2024

  23. [2020]

    A survey of graph neural networks for recommender systems: Challenges, methods, and directions

    [Gao et al., 2023] Chen Gao, Yu Zheng, Nian Li, Yinfeng Li, Yingrong Qin, Jinghua Piao, Yuhan Quan, Jianxin Chang, Depeng Jin, Xiangnan He, et al. A survey of graph neural networks for recommender systems: Challenges, methods, and directions. ACM Transactions on Recommender Sy...

    2023

  24. [2021]

    Unified language model pre- training for natural language understanding and genera- tion

    [Dong et al., 2019] Li Dong, Nan Yang, Wenhui Wang, Furu Wei, Xiaodong Liu, Yu Wang, Jianfeng Gao, Ming Zhou, and Hsiao-Wuen Hon. Unified language model pre- training for natural language understanding and genera- tion. Advances in neural information processing systems, 32,

    2019

  25. [2022]

    Heterogeneous network rep- resentation learning: A unified framework with survey and benchmark

    [Yang et al., 2020] Carl Yang, Yuxin Xiao, Yu Zhang, Yizhou Sun, and Jiawei Han. Heterogeneous network rep- resentation learning: A unified framework with survey and benchmark. IEEE Transactions on Knowledge and Data Engineering, 34(10):4854–4873,

    2020

  26. [2023]

    Knowledge graphs

    [Hogan et al., 2021] Aidan Hogan, Eva Blomqvist, Michael Cochez, Claudia d’Amato, Gerard De Melo, Claudio Gutierrez, Sabrina Kirrane, Jos ´e Emilio Labra Gayo, Roberto Navigli, Sebastian Neumaier, et al. Knowledge graphs. ACM Computing Surveys (Csur) , 54(4):1–37,

    2021

  27. [2024]

    Magnn: Metapath aggregated graph neural network for heterogeneous graph embedding

    [Fu et al., 2020] Xinyu Fu, Jiani Zhang, Ziqiao Meng, and Irwin King. Magnn: Metapath aggregated graph neural network for heterogeneous graph embedding. In Proceed- ings of the web conference 2020, pages 2331–2341,

    2020

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.