Pith. sign in

REVIEW 5 major objections 7 minor 2 cited by

Multi-view Fuzzy Graph Attention Networks for Enhanced Graph Learning

T0 review · 5 major / 7 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read This paper claims that adding a multi-view Transformation Block and learnable weighted-sum pooling to a Fuzzy Graph Attention Network improves graph classification accuracy beyond GCN, GAT, GraphSAGE, and FGAT on PROTEINS, NCI1, and…

desk verdict Coherent, incremental architecture paper whose single-run accuracy claims need seeds, error bars, and a holdout view-count choice before the central claim can be taken seriously. read the letter →

arxiv 2412.17271 v1 pith:EQNEOJ3L submitted 2024-12-23 cs.LG

Jinming Xing , Dongwen Luo , Qisen Cheng , Chang Xue , Ruilin Xing This is my paper · ORCID
classification cs.LG
keywords multi-viewlearningfuzzygraphattentionnetworkclassificationroughsetsneuralnetworkslearnablepoolingweightedsum
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 tries to establish that a graph neural network gains accuracy when it models the data from several learned perspectives at once instead of one. Its proposed model, MFGAT, wraps the existing Fuzzy Graph Attention Network in a Transformation Block that creates multiple linear views of each node's features and merges them with learnable weights, then adds a learnable weighted-sum global pooling. On three standard graph-classification benchmarks, MFGAT with three views reports the highest accuracy against GCN, GAT, GraphSAGE, and the base FGAT. The intended upshot is that fuzzy graph attention is stronger when fed multi-view representations, and that this combination sets a new practical level for graph classification.

What carries the argument

The Transformation Block is the load-bearing object. For each of $m$ views, it applies a learnable affine map $W_j x_i + b_j$ to every node, gathers the $m$ representations, and merges them by element-wise multiplication with learned view weights $W$. A second multi-view weighted-sum pooling, $Pool_{ws}^{mv}(X) = W^{mv} \odot (\Vert_j W_j X)$, turns node features into a graph-level vector. Together these give the network a tunable number of perspectives and a differentiable way to weight them, which is what the paper credits for the accuracy gains.

What would settle it

Run MFGAT and FGAT on PROTEINS, NCI1, and Mutagenicity with ten or more random seeds, holding all hyperparameters fixed except the seed, and report mean and standard deviation. If the 95% confidence intervals for MFGAT and FGAT overlap on any of the three datasets, the paper's central claim of consistent superiority is not supported.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that MFGAT, with the number of views set to three, outperforms all four baselines on all three datasets: PROTEINS (0.7630 vs. FGAT's 0.7188), NCI1 (0.6865 vs. FGAT's 0.6712), and Mutagenicity (0.8097 vs. the best baseline GCN's 0.7604). The mechanism is a Transformation Block that computes view-specific embeddings $x_i^j = W_j x_i + b_j$, pools them element-wise with learned weights into a unified representation, feeds that into FGAT convolutions, and finishes with a multi-view weighted-sum pooling that concatenates per-view weighted sums. The paper interprets this as evidence that multi-view dependencies and fuzzy relations complement each other for graph-level learning.

Load-bearing premise

The accuracy differences between MFGAT and the baselines are treated as meaningful even though each number comes from a single run without error bars or significance tests, so they are assumed to reflect a real improvement and not random seed variation.

Editorial extensions

If this is right

  • On the three benchmarks, setting the view count to three is the best choice; one view (equivalent to a plain FGAT front-end) and ten views both score lower, so view count acts as a tunable capacity knob.
  • If the gain is real, the same Transformation Block should lift other GNN backbones, not just FGAT, because the block is architecture-agnostic.
  • Multi-view weighted-sum pooling is a drop-in replacement for mean or max pooling in graph-level tasks, and should generalize beyond the fuzzy setting.
  • The paper's proposed next step, applying MFGAT to node classification and link prediction, is a direct corollary since the convolution and pooling components operate at node and graph level respectively.

Reading between the lines

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

  • Because Table 2 reports one accuracy number per model with no error bars, the headline margins (for example +1.5 points on NCI1) could plausibly be seed noise; a multi-seed rerun is the decisive check.
  • The extra parameters of the Transformation Block may be the true source of the gain, independent of fuzzy rough sets; ablating FGAT versus a plain GAT with the same multi-view front-end would separate the two.
  • The choice of three views was made after observing test performance on the same datasets, so the reported numbers likely overstate the gain a practitioner would see; selecting $m$ on validation alone is the honest protocol.
  • A natural extension would apply MFGAT to datasets with genuinely distinct feature channels so the learned views can be compared with externally defined views.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 7 minor

Summary. The paper proposes MFGAT, an extension of the Fuzzy Graph Attention Network (FGAT) that adds a multi-view Transformation Block, weighted-sum view aggregation, and a multi-view learnable global pooling mechanism. The method is evaluated on three TUDataset graph-classification benchmarks (PROTEINS, NCI1, Mutagenicity) against GCN, GAT, GraphSAGE, and FGAT. The central claim is that MFGAT with m=3 views outperforms all baselines on all three datasets, as reported in Table 2, and that the number of views has a systematic effect on accuracy, as shown in Figure 2.

Significance. If the empirical claims were reliable, the paper would demonstrate a simple and potentially useful recipe: adding trainable linear multi-view projections and weighted-sum pooling to FGAT improves graph-classification accuracy on standard benchmarks. The method is easy to understand and the problem of integrating fuzzy rough sets with multi-view graph learning is of some interest. However, the contribution is incremental and the paper's value rests almost entirely on the experimental evidence, which is currently not sufficient to support the stated conclusions: the headline comparison is based on single runs without variance, the number of views is selected on the test set, no ablations isolate the new components, and no code or full hyperparameter configuration is provided. The paper does not include machine-checked proofs or reproducible artifacts, so the significance assessment depends on whether the reported accuracy gains survive repeated runs and properly controlled comparisons.

major comments (5)
  1. [§4.3, Table 2] Each cell in Table 2 is a single accuracy value with no standard deviation, no number of runs, and no significance test. Because the reported gains (e.g., 0.7630 vs. 0.7188 on PROTEINS and 0.8097 vs. 0.7228 on Mutagenicity) are of a size that could plausibly arise from seed noise on datasets of 1,113 to 4,337 graphs, the central claim that MFGAT consistently outperformed all baselines is not verifiable as reported. Please report mean ± standard deviation over at least 10 random seeds, perform paired significance tests, and release the code and random seeds needed for reproduction.
  2. [§4.3, Figure 2] The number of views m=3 is selected after observing test-set accuracy in Figure 2, and the same test sets are then used for the headline results in Table 2. This selection-on-test procedure makes the reported comparison optimistically biased and is conceptually indistinguishable from tuning the model on the test set. The paper should select m using nested validation or a separate held-out test set, and should report the full view sweep with confidence intervals rather than a single best value.
  3. [§3.1, Eq. (2); §3.3, Eq. (7)] The core aggregation operations are not precisely defined. In Eq. (2), x_mv_i is introduced as the set of view vectors and W as a set of weights, so the expression W ⊙ x_mv_i is not a valid element-wise product unless the views are first concatenated or the notation is otherwise specified; if a weighted sum is intended, it should be written explicitly as Σ_j w_j x^j_i. In Eq. (7), the concatenation operator || and element-wise product ⊙ are combined without defining the dimensions, and the term W_j X is not connected to the view transformation in Eq. (1), which includes a bias b_j. Since the Transformation Block and the pooling mechanism are the paper's two main contributions, these formulas need to be corrected and aligned with the actual implementation.
  4. [§4.2] The experimental setup is under-specified. The paper reports only max epochs=200, learning rate=0.01, optimizer=Adam, and a 70/10/20 split, but does not give the hidden dimension, number of FGAT layers, dropout probability, weight decay, batch size, early-stopping patience, initialization seed, or the rule used to select the best validation model. Without this information, and without a description of the hyperparameter search used for each baseline, the comparison in Table 2 cannot be judged as fair or independently reproduced. Please provide the full configuration for MFGAT and each baseline.
  5. [§3.1–§3.4, Table 2] No ablation isolates the contributions of the Transformation Block and the multi-view global pooling. The paper attributes the accuracy improvement to both new components, but the only reported comparison is full MFGAT versus FGAT and other baselines, so the reader cannot determine whether the gain comes from the linear view projections, the weighted-sum view aggregation, the pooling change, or simply from additional parameters. Add ablations such as FGAT, FGAT + Transformation Block, FGAT + proposed pooling, and full MFGAT, all with matched hyperparameters and seeds.
minor comments (7)
  1. [§4.2, reference [28]] The GCN baseline is cited to Chen et al., 'Simple and deep graph convolutional networks,' which is not the original GCN paper; if the baseline is standard GCN, the citation should be to Kipf and Welling (2017), and if the baseline is actually GCNII, the text should say so.
  2. [§4.3, Figure 2] Figure 2 reports no numerical values and no error bars; please provide the exact accuracy values for m=1, 3, 5, 10 in a table alongside the figure.
  3. [§4.2] The text says both 'Cross-validation and early stopping were employed' and that the data were split into 70/10/20; please clarify what form of cross-validation was used and how the validation set interacts with early stopping.
  4. [§3.2, Eq. (3)] The attention equation omits the masking of non-neighbor nodes and does not specify the LeakyReLU slope; adding these details would make the GAT component unambiguous.
  5. [§4.3, text below Figure 2] The sentence 'too high may adversely affected performance' contains a grammatical error and should read 'too high may adversely affect performance.'
  6. [§2.3 and reference [33]] The paper relies on reference [33], a same-author arXiv preprint on pooling in LLMs, as the authority for 'common pooling techniques'; either justify this reliance or replace it with a more standard reference on pooling for graph neural networks.
  7. [Throughout] The term 'state-of-the-art baselines' overstates the comparison set: GCN, GAT, GraphSAGE, and FGAT are standard baselines but not current state-of-the-art on TUDataset. Please soften this wording or add stronger baselines such as GIN or PNA.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: MFGAT's empirical claims come from training, not from definitional or self-referential construction.

full rationale

The paper's central claim is that MFGAT with three views outperforms baselines on PROTEINS, NCI1, and Mutagenicity. These numbers are produced by training the model on standard TUDataset benchmarks; no equation in the paper defines a prediction in terms of the data being predicted. Equation (1) defines view-specific linear transformations, Equation (2) defines the weighted-sum fusion of views, and Equation (7) defines the multi-view pooling mechanism. These are architectural definitions, not derived quantities that reduce to their own outputs. The self-citations to FGAT [6], FGATT [10], and the pooling analysis [33] provide motivation and building blocks, but the MFGAT results are independently generated by running the model, and the baselines are external. The choice of m=3 after observing test performance in Figure 2 is a model-selection / overfitting concern, not a circularity concern, because the reported accuracies are still the result of training rather than being fixed by construction. No fitted parameter from the comparison datasets is renamed as a prediction, and no uniqueness theorem is imported from the authors' prior work. The paper is self-contained against external benchmarks, and no circular step can be exhibited from the text.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The central claim rests on the transferred effectiveness of FGAT, on the assumption that random linear projections are meaningful views, on weighted-sum pooling being the right aggregation, and on the stability of single-run test accuracies. None of these is validated by an ablation or significance test in the paper.

free parameters (4)
  • number of views m = 3
    Hand-selected after evaluating m in {1,3,5,10} on the same test sets; best-on-test selection can inflate claimed gains.
  • learning rate = 0.01
    Chosen for all models, no sensitivity analysis; affects absolute accuracies and comparisons.
  • max epochs = 200
    Fixed cap with early stopping; stopping criterion is not specified, so comparisons may not be apples-to-apples.
  • FGAT layer count, hidden dimension, dropout probability = not reported
    These architectural hyperparameters are unspecified, so the exact model is underdetermined and not fully reproducible.
assumptions (4)
  • domain assumption FGAT from [6], developed for link prediction, transfers to graph classification without modification beyond pooling.
    Section 3.2 uses FGAT layers as described in [6]; no adaptation or validation for graph-level tasks is given.
  • ad hoc to paper Each learned linear projection W_j defines a semantically meaningful view.
    Eq. (1) defines views as arbitrary linear transforms of the same node features; there is no evidence these correspond to distinct perspectives.
  • domain assumption Weighted sum pooling is the appropriate aggregation across views.
    Section 3.1 cites [33] for pooling options but gives no criterion for choosing weighted sum over mean or max for multi-view fusion.
  • domain assumption TUDataset splits and labels are reliable and the 70/10/20 split replaces the standard benchmark protocol.
    Section 4.2 uses a custom split; TUDataset typically uses fixed 10-fold cross-validation, so results may not be comparable to published numbers.
invented entities (1)
  • Multi-view linear projections ('views')
    purpose: To capture dependencies from multiple perspectives.
    The views are just learned linear maps of the same input; no independent signal or data modality supports them, and no experiment isolates what each view captures.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-view Fuzzy Graph Attention Networks for Enhanced Graph Learning." pith.science (2026). https://pith.science/paper/EQNEOJ3L

@misc{pith2026241217271,
  author       = {Pith},
  title        = {Pith review of: Multi-view Fuzzy Graph Attention Networks for Enhanced Graph Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EQNEOJ3L}},
  note         = {Machine review of arXiv:2412.17271}
}
read the original abstract

Fuzzy Graph Attention Network (FGAT), which combines Fuzzy Rough Sets and Graph Attention Networks, has shown promise in tasks requiring robust graph-based learning. However, existing models struggle to effectively capture dependencies from multiple perspectives, limiting their ability to model complex data. To address this gap, we propose the Multi-view Fuzzy Graph Attention Network (MFGAT), a novel framework that constructs and aggregates multi-view information using a specially designed Transformation Block. This block dynamically transforms data from multiple aspects and aggregates the resulting representations via a weighted sum mechanism, enabling comprehensive multi-view modeling. The aggregated information is fed into FGAT to enhance fuzzy graph convolutions. Additionally, we introduce a simple yet effective learnable global pooling mechanism for improved graph-level understanding. Extensive experiments on graph classification tasks demonstrate that MFGAT outperforms state-of-the-art baselines, underscoring its effectiveness and versatility.

Figures

Figures reproduced from arXiv: 2412.17271 by the authors.

Figure 1
Figure 1. Multi-view Fuzzy Graph Attention Network [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Impact of Number of Views on MFGAT A consistent trend can be observed across all three datasets. MF￾GAT with three views achieved the best performance, followed by five views. Notably, setting the number of views too low or too high may adversely affected performance. Specifically, too few views limit the model’s ability to capture diverse perspectives while too many views may introduce more noise than information, … view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. RecMind: LLM-Enhanced Graph Neural Networks for Personalized Consumer Recommendations

    cs.LG 2025-09 conditional novelty 5.0 of 10

    RecMind aligns LLM text embeddings with LightGCN collaborative embeddings via contrastive learning and a learned gate, achieving the best reported scores on all 8 ranking metrics across two datasets.

  2. Enhanced Convolutional Neural Networks for Improved Image Classification

    cs.CV 2025-02 reject novelty 2.0 of 10

    An enhanced CNN with standard techniques claims 84.95% on CIFAR-10, but weak baselines and missing evidence undermine the contribution.

Reference graph

Works this paper leans on

35 extracted references · 28 canonical work pages · cited by 2 Pith papers

  1. [6]

    Enhancing Link Prediction with Fuzzy Graph Attention Networks and Dynamic Negative Sampling

    Xing, Jinming. "Enhancing Link Prediction with Fuzzy Graph Attention Networks and Dynamic Negative Sampling. " arXiv preprint arXiv:2411.07482 (2024)

    arXiv 2024

  2. [33]

    Comparative Analysis of Pool- ing Mechanisms in LLMs: A Sentiment Analysis Perspective

    Xing, Jinming, Ruilin Xing, and Yan Sun. "Comparative Analysis of Pool- ing Mechanisms in LLMs: A Sentiment Analysis Perspective. " arXiv preprint arXiv:2411.14654 (2024)

    arXiv 2024

  3. [1]

    Rough fuzzy sets and fuzzy rough sets

    Dubois, Didier, and Henri Prade. "Rough fuzzy sets and fuzzy rough sets. " Inter- national Journal of General System 17, no. 2-3 (1990): 191-209

    work page 1990

  4. [2]

    Parameterized maximum-entropy-based three-way approximate attribute reduction

    Gao, Can, et al. "Parameterized maximum-entropy-based three-way approximate attribute reduction. " International Journal of Approximate Reasoning 151 (2022): 85-100

    work page 2022

  5. [3]

    Novel variable precision fuzzy rough sets and three-way decision model with three strategies

    Zou, Dandan, Yaoliang Xu, Lingqiang Li, and Zhenming Ma. "Novel variable precision fuzzy rough sets and three-way decision model with three strategies. " Information Sciences 629 (2023): 222-248

    work page 2023

  6. [4]

    Weighted fuzzy rough sets-based tri- training and its application to medical diagnosis

    Xing, Jinming, Can Gao, and Jie Zhou. "Weighted fuzzy rough sets-based tri- training and its application to medical diagnosis. " Applied Soft Computing 124 (2022): 109025. Multi-view Fuzzy Graph Attention Networks for Enhanced Graph Learning ISMSI ’25, Tokyo, Japan,

    work page 2022

  7. [5]

    A survey on graph neural networks for time series: Forecasting, classification, imputation, and anomaly detection

    Jin, Ming, Huan Yee Koh, Qingsong Wen, Daniele Zambon, Cesare Alippi, Geof- frey I. Webb, Irwin King, and Shirui Pan. "A survey on graph neural networks for time series: Forecasting, classification, imputation, and anomaly detection. " IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

    work page 2024

  8. [7]

    Towards semi- supervised universal graph classification

    Luo, Xiao, Yusheng Zhao, Yifang Qin, Wei Ju, and Ming Zhang. "Towards semi- supervised universal graph classification. " IEEE Transactions on Knowledge and Data Engineering 36, no. 1 (2023): 416-428

    work page 2023

Show all 35 references
  1. [8]

    Attention is all you need

    Vaswani, A. "Attention is all you need. " Advances in Neural Information Process- ing Systems (2017)

    2017

  2. [9]

    Graph attention networks

    Veličković, Petar, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. "Graph attention networks. " arXiv preprint arXiv:1710.10903 (2017)

    2017 arXiv

  3. [10]

    FGATT: A Robust Framework for Wire- less Data Imputation Using Fuzzy Graph Attention Networks and Transformer Encoders

    Xing, Jinming, Ruilin Xing, and Yan Sun. "FGATT: A Robust Framework for Wire- less Data Imputation Using Fuzzy Graph Attention Networks and Transformer Encoders. " arXiv preprint arXiv:2412.01979 (2024)

    2024 arXiv

  4. [11]

    Fuzzy rough sets and fuzzy rough neural networks for feature selection: A review

    Ji, Wanting, Yan Pang, Xiaoyun Jia, Zhongwei Wang, Feng Hou, Baoyan Song, Mingzhe Liu, and Ruili Wang. "Fuzzy rough sets and fuzzy rough neural networks for feature selection: A review. " Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery 11, no. 3 (2021): e1402

    2021

  5. [12]

    A new multi-objective decision-making method with diversified weights and Pythagorean fuzzy rough sets

    Ye, Jin, Bingzhen Sun, Qiang Bao, Chun Che, Qingchun Huang, and Xiaoli Chu. "A new multi-objective decision-making method with diversified weights and Pythagorean fuzzy rough sets. " Computers & Industrial Engineering 182 (2023): 109406

    2023

  6. [13]

    Attribute reduction methods in fuzzy rough set theory: An overview, com- parative experiments, and new directions

    Yuan, Zhong, Hongmei Chen, Peng Xie, Pengfei Zhang, Jia Liu, and Tianrui Li. "Attribute reduction methods in fuzzy rough set theory: An overview, com- parative experiments, and new directions. " Applied Soft Computing 107 (2021): 107353

    2021

  7. [14]

    Context helps determine spatial knowledge from tweets

    Xiao, Zhaomin, Yan Huang, and Eduardo Blanco. "Context helps determine spatial knowledge from tweets. " In Findings of the Association for Computational Linguistics: IJCNLP-AACL 2023 (Findings), pp. 149-160. 2023

    2023

  8. [15]

    Corporate Event Predictions Using Large Language Models

    Xiao, Zhaomin, Zhelu Mai, Zhuoer Xu, Yachen Cui, and Jiancheng Li. "Corporate Event Predictions Using Large Language Models. " In 2023 10th International Conference on Soft Computing & Machine Intelligence (ISCMI), pp. 193-197. IEEE, 2023

    2023

  9. [16]

    Financial Sentiment Analysis Meets LLaMA 3: A Comprehensive Analysis

    Mai, Zhelu, Jinran Zhang, Zhuoer Xu, and Zhaomin Xiao. "Financial Sentiment Analysis Meets LLaMA 3: A Comprehensive Analysis. " In Proceedings of the 2024 7th International Conference on Machine Learning and Machine Intelligence (MLMI), pp. 171-175. 2024

    2024

  10. [17]

    Botnet detection based on machine learning

    Yang, Xiaoran, Zhen Guo, and Zetian Mai. "Botnet detection based on machine learning. " In 2022 International Conference on Blockchain Technology and Infor- mation Security (ICBCTIS), pp. 213-217. IEEE, 2022

    2022

  11. [18]

    Research and application of knowledge graph technology for intelligent question answering

    Lin, Chengrong, Shaofan Chen, Xiaoran Yang, Caimao Li, Cong Qu, and Qiuhong Chen. "Research and application of knowledge graph technology for intelligent question answering. " In 2021 12th International Symposium on Parallel Architec- tures, Algorithms and Programming (PAAP), ...

    2021

  12. [19]

    Research on automatic composition based on multiple machine learning models

    Yang, Xiaoran. "Research on automatic composition based on multiple machine learning models. " In 2021 3rd International Conference on Artificial Intelligence and Advanced Manufacture, pp. 1206-1209. 2021

    2021

  13. [20]

    Image Retrieval Based on Blockchain

    Zhao, Sihao, Jiahui Xie, Chu Lin, Xiaolan Nie, Jun Ye, Xiaoran Yang, and Pengzhi Xu. "Image Retrieval Based on Blockchain. " In 2022 International Conference on Blockchain Technology and Information Security (ICBCTIS), pp. 210-212. IEEE, 2022

    2022

  14. [21]

    Analysis of Lifting Posture by Two Inertial Measurement Units and a Classification Model Based on a Convolutional Neural Network

    Liu, Juncheng, Xiaoran Yang, Jing-Chen Hong, and Hiroyasu Iwata. "Analysis of Lifting Posture by Two Inertial Measurement Units and a Classification Model Based on a Convolutional Neural Network. " In 2024 10th IEEE RAS/EMBS Inter- national Conference for Biomedical Robotics a...

    2024

  15. [22]

    Balancing Real-world Interaction and VR Immersion with AI Vision Robotic Arm

    Yang, Xiaoran, Yang Zhan, Yukiko Iwasaki, Miaohui Shi, Shijie Tang, and Hi- royasu Iwata. "Balancing Real-world Interaction and VR Immersion with AI Vision Robotic Arm. " In 2023 IEEE International Conference on Mechatronics and Automation (ICMA), pp. 2051-2057. IEEE, 2023

    2023

  16. [23]

    Short-Term Traffic Condition Prediction Based on Multi-source Data Fusion

    Deng, Xiaoru, Hui Zhou, Xiaoran Yang, and Chunyang Ye. "Short-Term Traffic Condition Prediction Based on Multi-source Data Fusion. " In International Con- ference on Data Mining and Big Data, pp. 327-335. Singapore: Springer Singapore, 2021

    2021

  17. [24]

    Improvement and Implementation of a Speech Emotion Recognition Model Based on Dual-Layer LSTM

    Yang, Xiaoran, Shuhan Yu, and Wenxi Xu. "Improvement and Implementation of a Speech Emotion Recognition Model Based on Dual-Layer LSTM. " arXiv preprint arXiv:2411.09189 (2024)

    2024 arXiv

  18. [25]

    Everything is connected: Graph neural networks

    Veličković, Petar. "Everything is connected: Graph neural networks. " Current Opinion in Structural Biology 79 (2023): 102538

    2023

  19. [26]

    A comprehensive survey on graph neural networks

    Wu, Zonghan, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and S. Yu Philip. "A comprehensive survey on graph neural networks. " IEEE transactions on neural networks and learning systems 32, no. 1 (2020): 4-24

    2020

  20. [27]

    A comprehensive review of graph convolutional networks: approaches and applications

    Xu, Xinzheng, Xiaoyang Zhao, Meng Wei, and Zhongnian Li. "A comprehensive review of graph convolutional networks: approaches and applications. " Electronic Research Archive 31, no. 7 (2023): 4185-4215

    2023

  21. [28]

    Simple and deep graph convolutional networks

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

    2020

  22. [29]

    How powerful are graph neural networks?

    Xu, Keyulu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. "How powerful are graph neural networks?. " arXiv preprint arXiv:1810.00826 (2018)

    2018 arXiv

  23. [30]

    How attentive are graph attention networks?

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

    2021 arXiv

  24. [31]

    Deep multi- view learning methods: A review

    Yan, Xiaoqiang, Shizhe Hu, Yiqiao Mao, Yangdong Ye, and Hui Yu. "Deep multi- view learning methods: A review. " Neurocomputing 448 (2021): 106-129

    2021

  25. [32]

    GMC: Graph-based multi-view clustering

    Wang, Hao, Yan Yang, and Bing Liu. "GMC: Graph-based multi-view clustering. " IEEE Transactions on Knowledge and Data Engineering 32, no. 6 (2019): 1116- 1129

    2019

  26. [34]

    Tudataset: A collection of benchmark datasets for learn- ing with graphs

    Morris, Christopher, Nils M. Kriege, Franka Bause, Kristian Kersting, Petra Mutzel, and Marion Neumann. "Tudataset: A collection of benchmark datasets for learn- ing with graphs. " arXiv preprint arXiv:2007.08663 (2020)

    2020 arXiv

  27. [35]

    Inductive representation learn- ing on large graphs

    Hamilton, Will, Zhitao Ying, and Jure Leskovec. "Inductive representation learn- ing on large graphs. " Advances in neural information processing systems 30 (2017)

    2017

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.