Pith. sign in

REVIEW 5 major objections 6 minor 46 references

Variational Graph Convolutional Neural Networks

T0 review · 5 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 Variational graph convolutional networks give uncertainty estimates and slightly better accuracy

desk verdict A useful variational extension of GCNs to get output and attention uncertainty, but the empirical case for accuracy gains is not yet made. read the letter →

arxiv 2507.01699 v1 pith:HP7NQRYF submitted 2025-07-02 cs.LG

Illia Oleksiienko , Juho Kanniainen , Alexandros Iosifidis This is my paper
classification cs.LG
keywords VariationalneuralnetworksGraphconvolutionalUncertaintyestimationattentionSkeleton-basedactionrecognitionSocialtradinganalysisSpatio-temporalgraphs
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes turning four graph neural network architectures—GCN, GAT, ST-GCN, and AGCN—into variational neural networks, so that every layer outputs a Gaussian distribution rather than a fixed feature vector. On the Finnish board-membership social trading task and on skeleton-based action recognition benchmarks (NTU-60, NTU-120, Kinetics), these variational versions report a small consistent accuracy gain over their deterministic baselines. The same mechanism produces uncertainty estimates for the model's predictions and for its per-layer attention scores, which the paper offers as material for explainability and for deciding when a prediction needs expert review. The best reported result is a mean F1 of 0.637 for the Uncertainty-Aware Early Attention variant against 0.635 for the best variational baseline on the Finnish data.

What carries the argument

The carrying mechanism is the layer-wise reparameterized Gaussian sampling introduced by Variational Neural Networks and applied here to graph convolutions: a layer computes $\dot{\Pi}_\mu$ and $\dot{\Pi}_\sigma$ from the input, then draws $\tilde{S} \sim \mathcal{N}(\rho_\mu(\dot{\Pi}_\mu), \rho_\sigma(\dot{\Pi}_\sigma))$. In attention-based models the same template is applied to the attention matrix, $\tilde{\Lambda} \sim \mathcal{N}(\Lambda^\mu, \Lambda^\sigma)$, which is what turns attention weights into quantities with estimable uncertainty. Repeated sampling plus Monte Carlo averaging delivers the expectation and variance of outputs and attentions; initialization from a pretrained deterministic model (the IVGCN/IVGAT/IVST-GCN/IV-AGCN procedure) stabilizes training and improves the variational models' accuracy.

What would settle it

Run the same input through a pretrained VGAT with the original sampled attention and with the fully Monte Carlo integrated filtered attention, then compare the two output distributions statistically; if the distributions differ, the no-retraining construction is not valid.

Watch

Extended reading notes

Core claim

The central claim is that replacing the deterministic graph convolution with a reparameterized Gaussian layer improves classification accuracy while adding uncertainty estimates. For each variational layer, two graph-convolution sub-layers compute mean and variance parameters, and the output is a sample from that Gaussian; in VGAT and V-AGCN the attention matrix is sampled from its own Gaussian, giving attention uncertainty. Monte Carlo integration across repeated forward passes yields expected outputs and expected attentions together with their variances. Uncertainty-aware variants filter attentions according to their variance, and the Early Attention version, which must be trained from scratch, achieves the best accuracy among the uncertainty-aware models.

Load-bearing premise

The load-bearing assumption is that replacing each layer's sampled, input-dependent attention with a fixed filtered attention matrix leaves the output distribution unchanged, which is what would allow the Fully Monte Carlo Integrated models to be built from pretrained VGAT networks without retraining.

Editorial extensions

If this is right

  • Standard GCN and GAT models can be converted into variational form and trained to output both predictions and uncertainty values, giving practitioners a drop-in way to flag low-confidence predictions.
  • Attention uncertainty provides a per-link measure of how much the model relies on each graph edge, so an expert can inspect which investor connections or body-joint links drive a decision.
  • In critical applications, the output uncertainty can be used as a gate: only predictions above a confidence threshold are executed automatically, and the rest are sent for human review.
  • The variational versions improve over deterministic baselines on the tested benchmarks, with the largest gain on the Finnish social trading task (mean F1 0.637 vs 0.635 for the best baseline).
  • The same construction transfers from spatial to spatio-temporal GCNs by making both the spatial graph-convolution block and the temporal convolution block variational.

Reading between the lines

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

  • The accuracy gains reported are small, so the more durable value is probably the calibration of uncertainties; a natural extension is to measure calibration directly (e.g., expected calibration error on the reported datasets) instead of accuracy alone.
  • If attention uncertainty is reliable, it could serve as an input-agnostic graph-rewiring criterion, letting users prune unreliable edges before deployment rather than only filtering attentions during forward passes.
  • The claim that fully Monte Carlo integrated models can be created from pretrained VGAT without retraining is directly testable by comparing output distributions with and without filtered attention; that check would settle whether the efficiency advantage holds.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper proposes variational versions of Graph Convolutional Networks based on the authors' earlier Variational Neural Network (VNN) framework. It introduces VGCN and VGAT for spatial graph tasks and VST-GCN and VAGCN for spatio-temporal tasks, where each layer outputs a Gaussian distribution parameterized by two sub-layers. The paper further proposes two uncertainty-aware variants: UA-EA-VGAT, which combines mean and variance attention branches early and requires retraining, and UA-FMCI-VGAT, which is claimed to preserve the output distribution and can be applied post hoc to a pretrained VGAT. Experiments are reported on the Finnish board membership dataset for social trading analysis and on NTU-60, NTU-120, and Kinetics for skeleton-based action recognition. The paper claims that the variational models improve classification performance while providing uncertainty in both outputs and attentions.

Significance. If the claims are substantiated, the paper would make a useful contribution by extending VNN-based uncertainty estimation to graph convolutional architectures, including attention uncertainty and layer-wise visualization. The provision of code links and the detailed mathematical formulations are strengths, as is the attempt to quantify attention uncertainty in a principled way. However, the central empirical claim of classification improvement is supported only by weak evidence: the largest reported gain is 0.002 in mean F1 with no error bars or significance tests, and the spatio-temporal results are shown only as plots. The FMCI no-retraining claim is theoretically questionable. The paper's value would be significantly enhanced by rigorous statistical evaluation and calibration assessment of the uncertainties.

major comments (5)
  1. [V-A, Table I] The central accuracy claim rests on a difference of 0.002 in mean F1 between IVGAT (0.635) and UA-EA-VGAT (0.637). No standard deviations, confidence intervals, or statistical significance tests are reported, despite the text stating that each model is trained with 10 random seeds and that variational models average over multiple tests. With stochastic variational inference, this difference is well within typical noise. In addition, the text states that 'the 20 best models are shown for the variational networks' while the figure caption says 'For each model type, 20 best models are displayed'; it is not clear whether identical hyperparameter selection was applied to the deterministic baselines. This selection bias, combined with the missing uncertainty quantification, means the conclusion that variational models 'provide an improvement in classification performance' is not established.
  2. [V-B, Figure 5] The skeleton-based action recognition results are presented only as plots with no numeric accuracy values, error bars, or confidence intervals. The text describes a 'slight improvement' but does not quantify it. For a paper whose headline claim is improvement in accuracy, omitting the numerical results and their variance makes it impossible to assess the magnitude or reliability of the claimed gains across NTU-60, NTU-120, and Kinetics.
  3. [IV-A, Eq. (17)] The claim that UA-FMCI-VGAT 'can be created directly from VGAT models without the need to train the network again' because 'the distribution of the output remains identical to the VGAT models' is not supported by the equations. In Eq. (9), the attention matrices Λμ and Λσ are random variables because they are functions of the sampled input Ŝiν; in Eq. (17), these are replaced by the deterministic filtered matrices Λ̃μ_filtered and Λ̃σ_filtered. Replacing input-dependent stochastic attention with fixed matrices removes a source of stochasticity, so the output distribution changes. The method may still be a useful approximation, but the stated distribution-preservation premise is false, and the post-hoc applicability claim collapses.
  4. [IV-A, Eq. (13)] The filtering condition in the second sub-equation of Eq. (13) appears to contain a typo: it filters Λ̃σ[k,q] based on the condition Λμ[k,q] ≤ l Λσ[k,q], whereas the accompanying text says the attention mean and variance are 'filtered based on the attention variance values.' The first sub-equation uses the condition Λσ[k,q] ≤ l Λμ[k,q]. The inconsistency makes the definition of the filter ambiguous and prevents exact reproduction of the method; the authors should correct the condition or explain the intended asymmetry.
  5. [IV, V] The paper does not evaluate the quality of the estimated uncertainties at all. Since uncertainty estimation is a central claimed contribution, it is insufficient to only provide qualitative figures (Figures 2 and 6). The authors should report calibration metrics such as expected calibration error, reliability diagrams, or coverage of predictive intervals for the variational models, and ideally compare these against standard baselines like Monte Carlo dropout or ensembles. Without such evaluation, the usefulness of the output and attention uncertainties remains unverified.
minor comments (6)
  1. [IV, VST-GCN block description] The text says 'This part applies two temporal convolutions TC µ(·) and TC µ(·)', which appears to be a typo; the second should likely be TC σ(·).
  2. [Throughout] The notation 'V AGCN' appears with a space in several places; it should be written consistently as 'VAGCN'.
  3. [V-A, Figure 4] The figure caption says 'For each model type, 20 best models with different hyperparameters are displayed.' To ensure a fair comparison, the authors should clearly state whether the same number of hyperparameter configurations and the same selection rule (e.g., best validation performance) were used for the deterministic baselines GCN and GAT.
  4. [IV-A, Eq. (15)] The expression for the reparameterized VGAT output is ambiguous: the term ρσ(ρ(Λσ) Ŝiσ) appears with a superscript '1 2' that is not clearly typeset as an exponent. Please clarify whether the variance term is squared or square-rooted and whether the activation is applied to the variance before or after the exponent.
  5. [V] The paper does not report concrete hyperparameter values, such as the attention filter limit l, the number of Monte Carlo samples, activation function placements, and training schedules. Including these details (or referring to the provided code) would substantially improve reproducibility.
  6. [V-B] For the spatio-temporal experiments, the authors state that only the 'best activation options' were chosen based on the smaller VGCN/VGAT networks; this selection process should be described explicitly so the reader can assess whether a fair comparison was made for each architecture.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the variational architectures are defined from first principles in this paper, and the accuracy claims are evaluated against independent baselines on public benchmarks.

full rationale

I examined the derivation chain from the variational layer definitions (Eqs. 8, 9, 11, 12) through the Uncertainty-Aware variants (Eqs. 13, 14, 17) to the experiments. The core method is constructed by replacing deterministic GCN/GAT/ST-GCN/AGCN layer computations with Gaussian-parametrized variational layers, and the subsequent UA-EA and UA-FMCI variants are explicit modifications of those equations. No equation reduces to a fitted parameter renamed as a prediction, and no result is defined in terms of the quantity it is claimed to predict. The paper does cite prior work by the same authors for the underlying Variational Neural Network framework ([12], [24]) and for the initialization strategy ([42]), but those citations provide the building blocks rather than the claimed outcome; the tested claim is an empirical comparison against deterministic baselines on Finnish board membership, NTU-60, NTU-120, and Kinetics, which is externally grounded. The most suspicious passage is the FMCI claim in Section IV-A: 'Since the distribution of the output remains identical to the VGAT models, UA-FMCI-VGAT models can be created directly from VGAT models without the need to train the network again.' This is an asserted identity, not a derived one, and replacing input-dependent stochastic attentions with fixed filtered matrices likely changes the output distribution. However, that is a correctness or validity concern, not circularity: it does not make the central accuracy or uncertainty claim equivalent to its own inputs by construction. Likewise, the selection of 'the 20 best models' in Section V-A is a statistical reporting issue that could bias the comparison, but it is not a circular step. I therefore find no significant circularity and assign score 0.

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

The variational GCN relies on the standard reparametrization trick and on prior graph architectures. The new free parameter is the attention filter limit l, plus unspecified MC sample counts and architecture choices. No new physical entities are introduced. The main ad hoc assumption is the independent per-entry filtering of attention uncertainties.

free parameters (5)
  • attention filter limit l
    In Equation (13), l controls which attention values are replaced by the fallback p based on uncertainty. Its value is not specified in the paper and would be tuned per dataset.
  • number of Monte Carlo samples
    Estimating attention mean and variance via repeated forward passes requires a sample count; the paper does not state how many are used for the reported results.
  • global variance flag
    Mentioned in Section V-A as a hyperparameter for VGCNs, controlling whether variance is shared; the chosen value is not reported.
  • activation positions for mean and variance
    The paper experiments with identity versus ReLU for rho_mu and rho_sigma, and reports only the 20 best models, so the effective hyperparameters are fitted to validation performance.
  • network architecture sizes
    Layer counts, hidden units, and training epochs are not specified; these are free choices that affect the accuracy comparisons.
assumptions (4)
  • standard math Reparametrization trick allows differentiable sampling from a Gaussian distribution.
    Used in Equations (8), (15), and (16) to backpropagate through stochastic samples.
  • domain assumption GCN, GAT, ST-GCN, and AGCN formulations as given in prior work are valid building blocks.
    Equations (1) through (7) adopt existing architectures without re-derivation; the variational wrapper assumes these are correct.
  • domain assumption A VNN layer with two sub-layers for mean and variance is a valid variational approximation.
    Borrowed from the authors' earlier VNN papers; the paper does not analyze the quality of the posterior approximation.
  • ad hoc to paper Attention uncertainty can be summarized by mean and variance of Monte Carlo samples and filtered independently per entry.
    Equation (13) filters each attention entry based on its own mean and variance; this ignores correlations between entries and is a heuristic without theoretical justification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Variational Graph Convolutional Neural Networks." pith.science (2026). https://pith.science/paper/HP7NQRYF

@misc{pith2026250701699,
  author       = {Pith},
  title        = {Pith review of: Variational Graph Convolutional Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HP7NQRYF}},
  note         = {Machine review of arXiv:2507.01699}
}
read the original abstract

Estimation of model uncertainty can help improve the explainability of Graph Convolutional Networks and the accuracy of the models at the same time. Uncertainty can also be used in critical applications to verify the results of the model by an expert or additional models. In this paper, we propose Variational Neural Network versions of spatial and spatio-temporal Graph Convolutional Networks. We estimate uncertainty in both outputs and layer-wise attentions of the models, which has the potential for improving model explainability. We showcase the benefits of these models in the social trading analysis and the skeleton-based human action recognition tasks on the Finnish board membership, NTU-60, NTU-120 and Kinetics datasets, where we show improvement in model accuracy in addition to estimated model uncertainties.

Figures

Figures reproduced from arXiv: 2507.01699 by the authors.

Figure 1
Figure 1. An example of spatio-temporal human body poses based on human [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. An example of layer-wise attention graphs in a single investor ego [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Structure of a Variational ST-GCN block. The input features [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Results of ST-GCN, AGCN and variational versions of ST-GCN and [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: An example of layer-wise attention graphs for a single NTU-60 [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 42 canonical work pages

  1. [1]

    Semi-supervised classification with graph convolutional networks,

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

    work page 2017

  2. [2]

    Graph attention networks,

    Petar Veli ˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Li `o, and Yoshua Bengio, “Graph attention networks,” in International Conference on Learning Representations , 2018

    work page 2018

  3. [3]

    Social me- dia sentiment analysis based on dependency graph and co-occurrence graph,

    Zhigang Jin, Manyue Tao, Xiaofang Zhao, and Yi Hu, “Social me- dia sentiment analysis based on dependency graph and co-occurrence graph,” Cognitive Computation, vol. 14, no. 3, pp. 1039–1054, 2022

    work page 2022

  4. [4]

    Rumor detection on social media with bi-directional graph convolutional networks,

    Tian Bian, Xi Xiao, Tingyang Xu, Peilin Zhao, Wenbing Huang, Yu Rong, and Junzhou Huang, “Rumor detection on social media with bi-directional graph convolutional networks,” AAAI Conference on Artificial Intelligence , vol. 34, no. 01, pp. 549–556, 2020

    work page 2020

  5. [5]

    Predicting the trading behavior of socially connected investors: Graph neural network approach with implications to market surveillance,

    Ke ¸stutis Baltakys, Margarita Baltakien˙e, Negar Heidari, Alexandros Iosi- fidis, and Juho Kanniainen, “Predicting the trading behavior of socially connected investors: Graph neural network approach with implications to market surveillance,” Expert Systems with Applications , vol. 228, pp. 120285, 2023

    work page 2023

  6. [6]

    Mdgnn: Multi-relational dynamic graph neural network for comprehensive and dynamic stock investment prediction,

    Hao Qian, Hongting Zhou, Qian Zhao, Hao Chen, Hongxiang Yao, Jing- wei Wang, Ziqi Liu, Fei Yu, Zhiqiang Zhang, and Jun Zhou, “Mdgnn: Multi-relational dynamic graph neural network for comprehensive and dynamic stock investment prediction,” AAAI Conference on Artificial Intelligence, vol. 38, no. 13, pp. 14642–14650, Mar. 2024

    work page 2024

  7. [7]

    Spatial temporal graph convolutional networks for skeleton-based action recognition,

    Sijie Yan, Yuanjun Xiong, and Dahua Lin, “Spatial temporal graph convolutional networks for skeleton-based action recognition,” arxiv:1801.07455, 2018. 1https://gitlab.au.dk/maleci/skeleton/skeleton-based-action-recognition 2https://github.com/iliiliiliili/insider-influence ILLIA OLEKSIIENKO et al.: V ARIATIONAL GRAPH CONVOLUTIONAL NEURAL NETWORKS 9

    arXiv 2018

  8. [8]

    Two-stream adap- tive graph convolutional networks for skeleton-based action recognition,

    Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu, “Two-stream adap- tive graph convolutional networks for skeleton-based action recognition,” in IEEE Conference on Computer Vision and Pattern Recognition, 2019

    work page 2019

Show all 46 references
  1. [9]

    Temporal attention-augmented graph convolutional network for efficient skeleton-based human action recognition,

    Negar Heidari and Alexandros Iosifidis, “Temporal attention-augmented graph convolutional network for efficient skeleton-based human action recognition,” in International Conference on Pattern Recognition, 2021, pp. 7907–7914

    2021

  2. [10]

    Gnn-ddas: Drug discovery for identifying anti- schistosome small molecules based on graph neural network,

    Xin Zeng, Peng-Kun Feng, Shu-Juan Li, Shuang-Qing Lv, Meng-Liang Wen, and Yi Li, “Gnn-ddas: Drug discovery for identifying anti- schistosome small molecules based on graph neural network,” Journal of Computational Chemistry , vol. 45, no. 32, pp. 2825–2834, 2024

    2024

  3. [11]

    Gcrnn: graph convolutional recurrent neural network for compound–protein interaction prediction,

    Ermal Elbasani, Soualihou Ngnamsie Njimbouom, Tae-Jin Oh, Eung- Hee Kim, Hyun Lee, and Jeong-Dong Kim, “Gcrnn: graph convolutional recurrent neural network for compound–protein interaction prediction,” BMC Bioinformatics, vol. 22, no. 5, pp. 616, 2022

    2022

  4. [12]

    Variational neural networks,

    Illia Oleksiienko, Dat Thanh Tran, and Alexandros Iosifidis, “Variational neural networks,” Procedia Computer Science, vol. 222C, pp. 104–113, 2023

    2023

  5. [13]

    A survey of uncertainty in deep neural networks,

    Jakob Gawlikowski, Cedrique Rovile Njieutcheu Tassi, Mohsin Ali, Jongseok Lee, Matthias Humt, Jianxiang Feng, Anna M. Kruspe, Rudolph Triebel, Peter Jung, Ribana Roscher, Muhammad Shahzad, Wen Yang, Richard Bamler, and Xiao Xiang Zhu, “A survey of uncertainty in deep neural ne...

    2023

  6. [14]

    Evidential deep learning to quantify classification uncertainty,

    Murat Sensoy, Lance Kaplan, and Melih Kandemir, “Evidential deep learning to quantify classification uncertainty,” in Advances on Neural Information Processing Systems , 2018, p. 3183–3193

    2018

  7. [15]

    Uncertainty-aware voxel based 3d object detection and tracking with von-mises loss,

    Yuanxin Zhong, Minghan Zhu, and Huei Peng, “Uncertainty-aware voxel based 3d object detection and tracking with von-mises loss,” arXiv:2011.02553, 2020

    2011 arXiv

  8. [16]

    Weight Uncertainty in Neural Networks,

    Charles Blundell, Julien Cornebise, Koray Kavukcuoglu, and Daan Wierstra, “Weight Uncertainty in Neural Networks,” in International Confernece on Machine Learning , 2015

    2015

  9. [17]

    Bayesian learning for neural networks: an algorithmic survey,

    Martin Magris and Alexandros Iosifidis, “Bayesian learning for neural networks: an algorithmic survey,” Artificial Intelligence Review , 2023

    2023

  10. [18]

    Randomized prior functions for deep reinforcement learning,

    Ian Osband, John Aslanides, and Albin Cassirer, “Randomized prior functions for deep reinforcement learning,” in Advances on Neural Information Processing Systems , 2018, vol. 31, pp. 8626–8638

    2018

  11. [19]

    Deep sub-ensembles for fast uncertainty estimation in image classification,

    Matias Valdenegro-Toro, “Deep sub-ensembles for fast uncertainty estimation in image classification,” NeurIPS Workshop on Bayesian Deep Learning, 2019

    2019

  12. [20]

    Layer ensembles,

    Illia Oleksiienko and Alexandros Iosifidis, “Layer ensembles,” IEEE International Workshop on Machine Learning for Signal Processing , 2023

    2023

  13. [21]

    Au- tomatic brain tumor segmentation using convolutional neural networks with test-time augmentation,

    Guotai Wang, Wenqi Li, S ´ebastien Ourselin, and Tom Vercauteren, “Au- tomatic brain tumor segmentation using convolutional neural networks with test-time augmentation,” in BrainLes. 2018, vol. 11384, pp. 61–72, Springer

    2018

  14. [22]

    Aleatoric uncertainty estimation with test-time augmentation for medical image segmentation with convolu- tional neural networks,

    Guotai Wang, Wenqi Li, Michael Aertsen, Jan Deprest, S ´ebastien Ourselin, and Tom Vercauteren, “Aleatoric uncertainty estimation with test-time augmentation for medical image segmentation with convolu- tional neural networks,” Neurocomputing, vol. 338, pp. 34–45, 2019

    2019

  15. [23]

    Improving convolutional neural networks performance for image classification using test time augmen- tation: a case study using MURA dataset,

    Ibrahem Kandel and Mauro Castelli, “Improving convolutional neural networks performance for image classification using test time augmen- tation: a case study using MURA dataset,” Health Inf. Sci. Syst. , vol. 9, no. 1, pp. 33, 2021

    2021

  16. [24]

    Variational neural networks implementation in pytorch and jax,

    Illia Oleksiienko, Dat Thanh Tran, and Alexandros Iosifidis, “Variational neural networks implementation in pytorch and jax,” Software Impacts, vol. 14, pp. 100431, 2022

    2022

  17. [25]

    Uncertainty in Graph Neural Networks: A Survey,

    Fangxin Wang, Yuqing Liu, Kay Liu, Yibo Wang, Sourav Medya, and Philip S. Yu, “Uncertainty in Graph Neural Networks: A Survey,” Transactions on Machine Learning Research , 2024

    2024

  18. [26]

    Graphpatcher: Mitigating degree bias for graph neural networks via test-time augmentation,

    Mingxuan Ju, Tong Zhao, Wenhao Yu, Neil Shah, and Yanfang Ye, “Graphpatcher: Mitigating degree bias for graph neural networks via test-time augmentation,” in Advances in Neural Information Processing Systems, 2023, vol. 36, pp. 55785–55801

    2023

  19. [27]

    Social influence prediction with train and test time augmentation for graph neural networks,

    Hongbo Bo, Ryan McConville, Jun Hong, and Weiru Liu, “Social influence prediction with train and test time augmentation for graph neural networks,” in International Joint Conference on Neural Networks, 2021, pp. 1–8

    2021

  20. [28]

    Epistemic Neural Networks,

    Ian Osband, Zheng Wen, Mohammad Asghari, Morteza Ibrahimi, Xiyuan Lu, and Benjamin Van Roy, “Epistemic Neural Networks,” Advances on Neural Information Processing Systems , 2023

    2023

  21. [29]

    Dropout as a bayesian approx- imation: Representing model uncertainty in deep learning,

    Yarin Gal and Zoubin Ghahramani, “Dropout as a bayesian approx- imation: Representing model uncertainty in deep learning,” in JMLR Workshop and Conference Proceedings , 2016, vol. 48, pp. 1050–1059

    2016

  22. [30]

    Bayesian graph convolutional neural networks for semi-supervised classification,

    Yingxue Zhang, Soumyasundar Pal, Mark Coates, and Deniz Ustebay, “Bayesian graph convolutional neural networks for semi-supervised classification,” AAAI Conference on Artificial Intelligence , vol. 33, no. 01, pp. 5829–5836, 2019

    2019

  23. [31]

    Bayesian graph neural networks with adaptive connection sampling,

    Arman Hasanzadeh, Ehsan Hajiramezanali, Shahin Boluki, Mingyuan Zhou, Nick Duffield, Krishna Narayanan, and Xiaoning Qian, “Bayesian graph neural networks with adaptive connection sampling,” in Interna- tional Conference on Machine Learning, 2020, vol. 119, pp. 4094–4104

    2020

  24. [32]

    Drope- dge: Towards deep graph convolutional networks on node classification,

    Yu Rong, Wenbing Huang, Tingyang Xu, and Junzhou Huang, “Drope- dge: Towards deep graph convolutional networks on node classification,” in International Conference on Learning Representations , 2020

    2020

  25. [33]

    The kinetics human action video dataset,

    Will Kay, Jo ˜ao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman, “The kinetics human action video dataset,” arxiv:1705.06950, 2017

    2017 arXiv

  26. [34]

    A closer look at spatiotemporal convolutions for action recognition,

    Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri, “A closer look at spatiotemporal convolutions for action recognition,” in IEEE Conference on Computer Vision and Pattern Recognition, 2018

    2018

  27. [35]

    Auto-Encoding Variational Bayes,

    Diederik P Kingma and Max Welling, “Auto-Encoding Variational Bayes,” in Interntional Conference on Learning Representations , 2014

    2014

  28. [36]

    How do investment ideas spread through social interaction? evidence from a ponzi scheme,

    VILLE RANTALA, “How do investment ideas spread through social interaction? evidence from a ponzi scheme,” The Journal of Finance , vol. 74, no. 5, pp. 2349–2389, 2019

    2019

  29. [37]

    A neural network with a case based dynamic window for stock trading prediction,

    Pei-Chann Chang, Chen-Hao Liu, Jun-Lin Lin, Chin-Yuan Fan, and Celeste S.P. Ng, “A neural network with a case based dynamic window for stock trading prediction,” Expert Systems with Applications , vol. 36, no. 3, Part 2, pp. 6889–6898, 2009

    2009

  30. [38]

    An innovative neural network approach for stock market prediction,

    Xiongwen Pang, Yanqiang Zhou, Pan Wang, Weiwei Lin, and Victor Chang, “An innovative neural network approach for stock market prediction,” The Journal of Supercomputing , vol. 76, no. 3, pp. 2098– 2118, Mar 2020

    2020

  31. [39]

    Deepinf: Social influence prediction with deep learning,

    Jiezhong Qiu, Jian Tang, Hao Ma, Yuxiao Dong, Kuansan Wang, and Jie Tang, “Deepinf: Social influence prediction with deep learning,” in International Conference on Knowledge Discovery & Data Mining , New York, NY , USA, 2018, p. 2110–2119, Association for Computing Machinery

    2018

  32. [40]

    An edge feature aware heterogeneous graph neural network model to support tax evasion detection,

    Bin Shi, Bo Dong, Yiming Xu, Jiaxiang Wang, Yunfan Wang, and Qinghua Zheng, “An edge feature aware heterogeneous graph neural network model to support tax evasion detection,” Expert Systems with Applications, vol. 213, pp. 118903, 2023

    2023

  33. [41]

    On the combination of graph data for assessing thin-file borrowers’ creditworthiness,

    Ricardo Mu ˜noz-Cancino, Cristi´an Bravo, Sebasti´an A. R´ıos, and Manuel Gra˜na, “On the combination of graph data for assessing thin-file borrowers’ creditworthiness,” Expert Systems with Applications , vol. 213, pp. 118809, 2023

    2023

  34. [42]

    Uncertainty-aware ab3dmot by variational 3d object detection,

    Illia Oleksiienko and Alexandros Iosifidis, “Uncertainty-aware ab3dmot by variational 3d object detection,” in IEEE International Conference on Image Processing , 2024, pp. 3389–3395

    2024

  35. [43]

    Ntu rgb+d: A large scale dataset for 3d human activity analysis,

    Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang, “Ntu rgb+d: A large scale dataset for 3d human activity analysis,” in Conference on Computer Vision and Pattern Recognition , 2016

    2016

  36. [44]

    Ntu rgb+d 120: A large-scale benchmark for 3d human activity understanding,

    Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C. Kot, “Ntu rgb+d 120: A large-scale benchmark for 3d human activity understanding,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 42, no. 10, pp. 2684–2701, 2020

    2020

  37. [45]

    Progressive spatio-temporal graph convolutional network for skeleton-based human action recog- nition,

    Negar Heidari and Alexandros Iosifidis, “Progressive spatio-temporal graph convolutional network for skeleton-based human action recog- nition,” in International Conference on Acoustics, Speech and Signal Processing, 2021, pp. 3220–3224

    2021

  38. [46]

    Revealing key details to see differences: A novel prototypical perspective for skeleton-based action recognition,

    Hongda Liu, Yunfan Liu, Min Ren, Hao Wang, Yunlong Wang, and Zhenan Sun, “Revealing key details to see differences: A novel prototypical perspective for skeleton-based action recognition,” arXiv:2411.18941, 2024

    2024 arXiv

Pith tools

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