Pith. sign in

REVIEW 3 major objections 3 minor 62 references

Neighbor-Sampling Based Momentum Stochastic Methods for Training Graph Neural Networks

T0 review · 3 major / 3 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 Neighbor-sampling Adam-type methods with control variates achieve the optimal convergence rate for nonconvex GCN training and outperform CV-SGD in node-classification experiments.

desk verdict Plausible Adam-plus-control-variate method for GNN training, but the supplied full text is corrupted, so the verdict is open until a readable version appears. read the letter →

arxiv 2508.00267 v1 pith:LXH6SZYS submitted 2025-08-01 math.OC cs.LG

Molly Noel , Gabriel Mancino-Ball , Yangyang Xu This is my paper
classification math.OCcs.LG MSC 90C2690C1568T07
keywords graphneuralnetworksneighborsamplingAdamoptimizationcontrolvariatesnonconvexconvergenceratenodeclassification
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

Training graph convolutional networks by sampling a subset of each node's neighbors keeps each update cheap, but it injects stochastic error that has made convergence guarantees hard to get. This paper introduces Adam-style (momentum plus adaptive step size) stochastic methods for that setting, and it claims that adding a control variate, an old full-gradient estimate subtracted from the sampled gradient to cancel its variance, restores the optimal convergence rate for nonconvex objectives under the standard assumptions used for Adam. The same methods are tested on node classification across several benchmark datasets, where they are reported to outperform the earlier neighbor-sampling SGD that also uses control variates, most clearly on large graphs. If the claim holds, practitioners get a principled Adam alternative to variance-reduced SGD for GNN training rather than an empirically motivated heuristic.

What carries the argument

The central mechanism is the control variate combined with an Adam-style update. The control variate reuses a stale full-neighbor aggregation as a baseline: the sampled gradient is adjusted by the difference between the current sampled estimate and the stored full estimate, which removes most of the sampling noise without biasing the estimator. On top of that, the Adam-type step applies per-coordinate scaling and momentum. The proof uses the variance reduction to control the stochastic error term that would otherwise slow convergence, while momentum and adaptivity are what the experiments show to be practically beneficial.

What would settle it

On a real large graph, for example one with a power-law degree distribution, measure the variance of the control-variate neighbor-sampled gradient throughout training; if the variance is not bounded by a data-independent constant as training progresses, the assumptions behind the optimal-rate theorem are violated. A second check is to run the proposed method on a graph with increasing numbers of layers and compare the measured gradient-norm decay to the predicted rate; a materially slower decay would show the theorem's hypotheses are not met in practice.

Watch

Extended reading notes

Core claim

The central claim is that neighbor-sampling based Adam-type methods with control variates solve the nonconvex GCN training problem at the optimal convergence rate. The authors present several variants of the Adam update in which the stochastic gradient from a sampled neighbor set is corrected by a control-variate estimator, keeping the update unbiased while reducing its variance; under smoothness and the usual bounded-variance assumptions for Adam-type methods, the squared gradient norm is driven to zero at the optimal rate. The empirical section reports that these methods outperform classic NS-based SGD equipped with the same control variate on node classification tasks, with the advantage most pronounced on large-scale graph datasets.

Load-bearing premise

The optimal-rate proof relies on the standard Adam-type assumptions, namely smoothness of the loss and bounded (or bounded-variance) stochastic gradients, holding for the nonconvex, neighbor-sampled GCN objective, and the paper does not verify those assumptions on the actual graph datasets used in the experiments.

Editorial extensions

If this is right

  • GNN practitioners can use Adam-style training with a formal guarantee that the nonconvex objective reaches a stationary point at the optimal rate under standard assumptions.
  • Large-scale node classification should benefit because control variates keep the sampled gradient close to the full gradient as the graph grows.
  • Existing neighbor-sampling SGD pipelines that already use control variates can be upgraded with momentum and adaptive step sizes without losing variance reduction.
  • The optimal rate implies the iteration complexity matches the best known lower bound for this class of nonconvex stochastic problems.

Reading between the lines

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

  • Beyond the paper: the same variance-reduction-plus-Adam recipe likely transfers to other message-passing GNN architectures, but the paper's proof is written for the GCN training objective and does not establish that transfer.
  • Beyond the paper: the assumptions used for the rate, such as bounded stochastic variance, are unlikely to be exactly true on real graphs with heavy-tailed degrees, so the guarantee may be a worst-case statement rather than a description of the observed speedups.
  • Beyond the paper: storing the control-variate baseline across the graph adds memory; on very large graphs this overhead could become a practical bottleneck that the abstract-level experiments do not address.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 3 minor

Summary. The paper proposes neighbor-sampling-based Adam-type stochastic methods with control variates for training graph convolutional networks (GCNs). The abstract claims that, under standard assumptions for Adam-type methods, the proposed methods achieve an optimal convergence rate for a nonconvex GCN training problem, and that numerical experiments on node classification benchmarks show superior performance over classic NS-based SGD with the control-variate technique, especially on large-scale graphs. The authors also provide a public code repository. The supplied full text, however, is an unreadable sequence of replacement characters, so I could only evaluate the abstract; no theorem statement, proof, assumption list, algorithm description, or experimental table could be inspected.

Significance. If correct, the paper would address a practical gap: it would give GNN practitioners Adam-style adaptive methods with neighbor sampling, control variates, and convergence guarantees, while improving over CV-SGD empirically. The topic is timely and the abstract is written in a measured way. The promised code release and the explicit attribution of the control-variate technique to prior work are positive features. That said, the significance cannot be fully assessed because the body of the manuscript is unreadable in the supplied form; the optimal-rate claim and the empirical superiority claim both require detailed verification that is currently impossible.

major comments (3)
  1. [Full text (all sections)] The supplied manuscript body is an unreadable sequence of replacement characters; no assumptions, theorem statements, proofs, algorithm definitions, or experimental tables can be inspected. This blocks verification of both central claims of the paper, namely the optimal convergence rate and the superior empirical performance. The authors should provide a readable PDF or otherwise fix the encoding before the manuscript can be meaningfully reviewed.
  2. [Abstract] The abstract states that the methods achieve the optimal convergence rate under standard assumptions for Adam-type methods, but it does not state those assumptions or define the oracle/complexity class with respect to which the rate is optimal. Please state the assumptions explicitly and identify the comparison class, for example stochastic first-order methods with bounded variance; otherwise the optimality claim is ambiguous and cannot be checked against the proof.
  3. [Abstract; experiments] The empirical claim of superior performance over classic NS-based SGD with control variates is not accompanied in the supplied text by any tables, dataset-specific hyperparameters, or standard deviations. In addition, a standard GCN with ReLU activations is not globally L-smooth and its gradients are not bounded, while Adam-type convergence proofs typically require such conditions. Please clarify whether the theorem applies to the exact objective trained in the experiments or to a smoothed/regularized surrogate, and explain how the experimental setup satisfies the theorem's assumptions.
minor comments (3)
  1. [Abstract] The numeric reference [1] for the control-variate technique does not have a visible bibliography entry in the supplied text; please ensure all citations are rendered in the final version.
  2. [Abstract] The proposed methods are not given distinct names or abbreviations in the abstract; naming the variants would make the paper easier to read and cite.
  3. [Abstract] The code repository URL is a welcome addition; please also specify the license and the exact software versions used for the reported experiments.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity identified; the abstract's claims are conditional on stated standard assumptions and do not reduce by construction to their inputs.

full rationale

The paper's central claims are (i) convergence of neighbor-sampling Adam-type methods under standard assumptions for Adam-type methods and (ii) empirical superiority over CV-SGD. Neither claim is shown, in any legible portion of the supplied text, to be equivalent to its own inputs. The only external technique invoked is the control-variate method 'proposed by [1]', which is used as a variance-reduction tool and is not presented as the source of the convergence theorem. The 'optimal convergence rate' statement is explicitly conditional on assumptions and therefore does not smuggle the conclusion into the premise. The full text is almost entirely corrupted by encoding replacement characters, so no equation, theorem, or experimental table can be inspected; however, unverifiability is not circularity. There is no quoted derivation in which a parameter is fitted and then renamed a prediction, no self-citation chain that forces the choice, and no ansatz imported via citation that defines the target result. A low score is therefore the honest finding, with the caveat that the corrupted full text prevents a complete audit of the proof and experiments.

Assumptions & free parameters 2 free parameters · 2 assumptions · 0 invented entities

No fitted scientific constants are involved; the free parameters listed are experimental hyperparameters that affect the empirical claim. The theoretical claim is conditional on standard optimization assumptions, and no new physical or mathematical entities are introduced.

free parameters (2)
  • learning rate schedule = not disclosed in abstract
    The empirical comparison with CV-SGD depends on how learning rates are tuned per dataset; the abstract gives no protocol.
  • neighbor sample size and minibatch size = not disclosed in abstract
    These sampling choices set the variance of the gradient estimator and can determine whether the method beats the baseline; values are not visible in the abstract.
assumptions (2)
  • domain assumption The GCN training objective is smooth and satisfies the stochastic assumptions required by Adam-type analyses, such as bounded gradients or bounded variance.
    Invoked in the abstract through the phrase 'under standard assumptions for Adam-type methods'; these assumptions are not stated and may not hold for real graph data.
  • domain assumption The control-variate estimator from prior work is unbiased and has reduced variance under neighbor sampling.
    The paper says it 'utilizes the control variate technique proposed by [1]'; the correctness of the variance reduction and the resulting convergence rate rests on that prior estimator's properties.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Neighbor-Sampling Based Momentum Stochastic Methods for Training Graph Neural Networks." pith.science (2026). https://pith.science/paper/LXH6SZYS

@misc{pith2026250800267,
  author       = {Pith},
  title        = {Pith review of: Neighbor-Sampling Based Momentum Stochastic Methods for Training Graph Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LXH6SZYS}},
  note         = {Machine review of arXiv:2508.00267}
}
read the original abstract

Graph convolutional networks (GCNs) are a powerful tool for graph representation learning. Due to the recursive neighborhood aggregations employed by GCNs, efficient training methods suffer from a lack of theoretical guarantees or are missing important practical elements from modern deep learning algorithms, such as adaptivity and momentum. In this paper, we present several neighbor-sampling (NS) based Adam-type stochastic methods for solving a nonconvex GCN training problem. We utilize the control variate technique proposed by [1] to reduce the stochastic error caused by neighbor sampling. Under standard assumptions for Adam-type methods, we show that our methods enjoy the optimal convergence rate. In addition, we conduct extensive numerical experiments on node classification tasks with several benchmark datasets. The results demonstrate superior performance of our methods over classic NS-based SGD that also uses the control-variate technique, especially for large-scale graph datasets. Our code is available at https://github.com/RPI-OPT/CV-ADAM-GNN .

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

62 extracted references · 44 canonical work pages

  1. [1]

    , Zhu , J

    botherref Chen , J. , Zhu , J. , Song , L. : Stochastic training of graph convolutional networks with variance reduction. arXiv preprint arXiv:1710.10568 (2017) botherref

    arXiv 2017

  2. [2]

    , Wang , Y.-C

    barticle Chen , F. , Wang , Y.-C. , Wang , B. , Kuo , C.-C.J. : Graph representation learning: a survey . APSIPA Transactions on Signal and Information Processing 9 , 15 ( 2020 ) barticle

    work page 2020

  3. [3]

    , Ma , Y

    bchapter Fan , W. , Ma , Y. , Li , Q. , He , Y. , Zhao , E. , Tang , J. , Yin , D. : Graph neural networks for social recommendation . In: The World Wide Web Conference , pp. 417 -- 426 ( 2019 ) bchapter

    work page 2019

  4. [4]

    , Lee , Y.-C

    barticle Sharma , K. , Lee , Y.-C. , Nambi , S. , Salian , A. , Shah , S. , Kim , S.-W. , Kumar , S. : A survey of graph neural networks for social recommender systems . ACM Computing Surveys 56 ( 10 ), 1 -- 34 ( 2024 ) barticle

    work page 2024

  5. [5]

    , Sanchez-Gonzalez , A

    barticle Lam , R. , Sanchez-Gonzalez , A. , Willson , M. , Wirnsberger , P. , Fortunato , M. , Alet , F. , Ravuri , S. , Ewalds , T. , Eaton-Rosen , Z. , Hu , W. , Merose , A. , Hoyer , S. , Holland , G. , Vinyals , O. , Stott , J. , Pritzel , A. , Mohamed , S. , Battaglia , P. : Learning skillful medium-range global weather forecasting . Science 382 ( 66...

    doi:10.1126/science.adi2336 2023

  6. [6]

    , Schoenholz , S.S

    bchapter Gilmer , J. , Schoenholz , S.S. , Riley , P.F. , Vinyals , O. , Dahl , G.E. : Neural message passing for quantum chemistry . In: International Conference on Machine Learning , pp. 1263 -- 1272 ( 2017 ). PMLR bchapter

    work page 2017

  7. [7]

    , Zeng , J

    bchapter Liu , J. , Zeng , J. , Wang , X. , Liang , Z. : Learning graph-based code representations for source-level functional similarity detection . In: 2023 IEEE/ACM 45th International Conference on Software Engineering (ICSE) , pp. 345 -- 357 ( 2023 ). 10.1109/ICSE48619.2023.00040 bchapter

    arXiv 2023

  8. [8]

    , Monfardini , G

    bchapter Gori , M. , Monfardini , G. , Scarselli , F. : A new model for learning in graph domains . In: Proceedings. 2005 IEEE International Joint Conference on Neural Networks, 2005. , vol. 2 , pp. 729 -- 734 ( 2005 ). IEEE bchapter

    work page 2005

Show all 62 references
  1. [9]

    , Gori , M

    barticle Scarselli , F. , Gori , M. , Tsoi , A.C. , Hagenbuchner , M. , Monfardini , G. : The graph neural network model . IEEE transactions on neural networks 20 ( 1 ), 61 -- 80 ( 2008 ) barticle

    2008

  2. [10]

    , Welling , M

    botherref Kipf , T.N. , Welling , M. : Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907 (2016) botherref

    2016 arXiv

  3. [11]

    , Xiong , C

    botherref Li , G. , Xiong , C. , Thabet , A. , Ghanem , B. : Deepergcn: All you need to train deeper gcns. arXiv preprint arXiv:2006.07739 (2020) botherref

    2020 arXiv

  4. [12]

    , Yang , J

    botherref Wolfe , C.R. , Yang , J. , Liao , F. , Chowdhury , A. , Dun , C. , Bayer , A. , Segarra , S. , Kyrillidis , A. : Gist: Distributed training for large-scale graph convolutional networks. Journal of Applied and Computational Topology, 1--53 (2023) botherref

    2023

  5. [13]

    , Ying , Z

    botherref Hamilton , W. , Ying , Z. , Leskovec , J. : Inductive representation learning on large graphs. Advances in neural information processing systems 30 (2017) botherref

    2017

  6. [14]

    , Ma , T

    botherref Chen , J. , Ma , T. , Xiao , C. : Fastgcn: fast learning with graph convolutional networks via importance sampling. arXiv preprint arXiv:1801.10247 (2018) botherref

    2018 arXiv

  7. [15]

    , Zhou , H

    botherref Zeng , H. , Zhou , H. , Srivastava , A. , Kannan , R. , Prasanna , V. : Graphsaint: Graph sampling based inductive learning method. arXiv preprint arXiv:1907.04931 (2019) botherref

    2019 arXiv

  8. [16]

    , M \"u ller , M

    bchapter Li , G. , M \"u ller , M. , Ghanem , B. , Koltun , V. : Training graph neural networks with 1000 layers . In: International Conference on Machine Learning , pp. 6437 -- 6449 ( 2021 ). PMLR bchapter

    2021

  9. [17]

    , Monro , S

    botherref Robbins , H. , Monro , S. : A stochastic approximation method. The annals of mathematical statistics, 400--407 (1951) botherref

    1951

  10. [18]

    : Introductory lectures on convex optimization - a basic course

    bchapter Nesterov , Y. : Introductory lectures on convex optimization - a basic course . In: Applied Optimization ( 2014 ). https://api.semanticscholar.org/CorpusID:62288331 bchapter

    2014

  11. [19]

    , Ba , J

    botherref Kingma , D.P. , Ba , J. : Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) botherref

    2014 arXiv

  12. [20]

    , Xu , Y

    barticle Xu , Y. , Xu , Y. , Yan , Y. , Sutcher-Shepard , C. , Grinberg , L. , Chen , J. : Parallel and distributed asynchronous adaptive stochastic gradient methods . Mathematical Programming Computation 15 ( 3 ), 471 -- 508 ( 2023 ) barticle

    2023

  13. [21]

    : Some methods of speeding up the convergence of iteration methods

    barticle Polyak , B.T. : Some methods of speeding up the convergence of iteration methods . Ussr computational mathematics and mathematical physics 4 ( 5 ), 1 -- 17 ( 1964 ) barticle

    1964

  14. [22]

    , Kale , S

    botherref Reddi , S.J. , Kale , S. , Kumar , S. : On the convergence of adam and beyond. arXiv preprint arXiv:1904.09237 (2019) botherref

    2019 arXiv

  15. [23]

    , Hazan , E

    botherref Duchi , J. , Hazan , E. , Singer , Y. : Adaptive subgradient methods for online learning and stochastic optimization. Journal of machine learning research 12(7) (2011) botherref

    2011

  16. [24]

    , Liu , S

    botherref Chen , X. , Liu , S. , Sun , R. , Hong , M. : On the convergence of a class of adam-type algorithms for non-convex optimization. arXiv preprint arXiv:1808.02941 (2018) botherref

    2018 arXiv

  17. [25]

    , Tarlow , D

    botherref Li , Y. , Tarlow , D. , Brockschmidt , M. , Zemel , R. : Gated graph sequence neural networks. arXiv preprint arXiv:1511.05493 (2015) botherref

    2015 arXiv

  18. [26]

    , Micheli , A

    bchapter Gallicchio , C. , Micheli , A. : Graph echo state networks . In: The 2010 International Joint Conference on Neural Networks (IJCNN) , pp. 1 -- 8 ( 2010 ). IEEE bchapter

    2010

  19. [27]

    , Kozareva , Z

    bchapter Dai , H. , Kozareva , Z. , Dai , B. , Smola , A. , Song , L. : Learning steady-states of iterative algorithms over graphs . In: International Conference on Machine Learning , pp. 1106 -- 1114 ( 2018 ). PMLR bchapter

    2018

  20. [28]

    , Pan , S

    barticle Wu , Z. , Pan , S. , Chen , F. , Long , G. , Zhang , C. , Philip , S.Y. : A comprehensive survey on graph neural networks . IEEE transactions on neural networks and learning systems 32 ( 1 ), 4 -- 24 ( 2020 ) barticle

    2020

  21. [29]

    , Zaremba , W

    botherref Bruna , J. , Zaremba , W. , Szlam , A. , LeCun , Y. : Spectral networks and locally connected networks on graphs. arXiv preprint arXiv:1312.6203 (2013) botherref

    2013 arXiv

  22. [30]

    , Bruna , J

    botherref Henaff , M. , Bruna , J. , LeCun , Y. : Deep convolutional networks on graph-structured data. arXiv preprint arXiv:1506.05163 (2015) botherref

    2015 arXiv

  23. [31]

    , Bresson , X

    botherref Defferrard , M. , Bresson , X. , Vandergheynst , P. : Convolutional neural networks on graphs with fast localized spectral filtering. Advances in neural information processing systems 29 (2016) botherref

    2016

  24. [32]

    , Monti , F

    barticle Levie , R. , Monti , F. , Bresson , X. , Bronstein , M.M. : Cayleynets: Graph convolutional neural networks with complex rational spectral filters . IEEE Transactions on Signal Processing 67 ( 1 ), 97 -- 109 ( 2018 ) barticle

    2018

  25. [33]

    , Wang , S

    bchapter Li , R. , Wang , S. , Zhu , F. , Huang , J. : Adaptive graph convolutional neural networks . In: Proceedings of the AAAI Conference on Artificial Intelligence , vol. 32 ( 2018 ) bchapter

    2018

  26. [34]

    , Ma , Q

    bchapter Zhuang , C. , Ma , Q. : Dual graph convolutional networks for graph-based semi-supervised classification . In: Proceedings of the 2018 World Wide Web Conference , pp. 499 -- 508 ( 2018 ) bchapter

    2018

  27. [35]

    , Wang , Z

    bchapter Gao , H. , Wang , Z. , Ji , S. : Large-scale learnable graph convolutional networks . In: Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , pp. 1416 -- 1424 ( 2018 ) bchapter

    2018

  28. [36]

    , Hu , Z

    botherref Zou , D. , Hu , Z. , Wang , Y. , Jiang , S. , Sun , Y. , Gu , Q. : Layer-dependent importance sampling for training deep and large graph convolutional networks. Advances in neural information processing systems 32 (2019) botherref

    2019

  29. [37]

    , Zhang , Y

    botherref Huang , T. , Zhang , Y. , Wu , J. , Fang , J. , Zheng , Z. : Mg-gcn: Fast and effective learning with mix-grained aggregators for training large graph convolutional networks. arXiv preprint arXiv:2011.09900 (2020) botherref

    2020 arXiv

  30. [38]

    , Zhang , T

    botherref Huang , W. , Zhang , T. , Rong , Y. , Huang , J. : Adaptive sampling towards fast graph representation learning. Advances in neural information processing systems 31 (2018) botherref

    2018

  31. [39]

    , Liu , X

    bchapter Chiang , W.-L. , Liu , X. , Si , S. , Li , Y. , Bengio , S. , Hsieh , C.-J. : Cluster-gcn: An efficient algorithm for training deep and large graph convolutional networks . In: Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data M...

    2019

  32. [40]

    , Tian , Y

    botherref Ji , S. , Tian , Y. , Liu , F. , Li , X. , Wu , L. : Promptgcn: Bridging subgraph gaps in lightweight gcns. arXiv preprint arXiv:2410.10089 (2024) botherref

    2024 arXiv

  33. [41]

    , Yang , X

    barticle Wang , X. , Yang , X. , Wang , P. , Yu , H. , Xu , T. : Ssgcn: a sampling sequential guided graph convolutional network . International Journal of Machine Learning and Cybernetics 15 ( 5 ), 2023 -- 2038 ( 2024 ) barticle

    2023

  34. [42]

    , Wang , Y

    bchapter Wang , J. , Wang , Y. , Yang , Z. , Yang , L. , Guo , Y. : Bi-gcn: Binary graph convolutional network . In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pp. 1561 -- 1570 ( 2021 ) bchapter

    2021

  35. [43]

    , Alattas , H

    botherref Bellei , C. , Alattas , H. , Kaaniche , N. : Label-gcn: An effective method for adding label propagation to graph convolutional networks. arXiv preprint arXiv:2104.02153 (2021) botherref

    2021 arXiv

  36. [44]

    , Liu , G

    botherref Jiang , M. , Liu , G. , Su , Y. , Wu , X. : Gcn-sl: Graph convolutional networks with structure learning for graphs under heterophily. arXiv preprint arXiv:2105.13795 (2021) botherref

    2021 arXiv

  37. [45]

    , Liu , G

    barticle Jiang , M. , Liu , G. , Su , Y. , Wu , X. : Self-attention empowered graph convolutional network for structure learning and node embedding . Pattern Recognition 153 , 110537 ( 2024 ) barticle

    2024

  38. [46]

    , Lu , J

    barticle Huang , A. , Lu , J. , Wu , Z. , Chen , Z. , Chen , Y. , Wang , S. , Zhang , H. : Geometric localized graph convolutional network for multi-view semi-supervised classification . Information Sciences 677 , 120769 ( 2024 ) barticle

    2024

  39. [47]

    , Ramezani , M

    botherref Cong , W. , Ramezani , M. , Mahdavi , M. : On the importance of sampling in training gcns: Tighter analysis and variance reduction. 2021. URL http://arxiv. org/abs/2103 2696 (2021) botherref

    2021

  40. [48]

    , Lou , H

    barticle Ma , Y. , Lou , H. , Yan , M. , Sun , F. , Li , G. : Spatio-temporal fusion graph convolutional network for traffic flow forecasting . Information Fusion 104 , 102196 ( 2024 ) barticle

    2024

  41. [49]

    , Feng , T

    barticle Liu , Y. , Feng , T. , Rasouli , S. , Wong , M. : St-dagcn: A spatiotemporal dual adaptive graph convolutional network model for traffic prediction . Neurocomputing 601 , 128175 ( 2024 ) barticle

    2024

  42. [50]

    , Zhang , Y

    botherref Liu , A. , Zhang , Y. : Spatial--temporal dynamic graph convolutional network with interactive learning for traffic forecasting. IEEE Transactions on Intelligent Transportation Systems (2024) botherref

    2024

  43. [51]

    , Khalid , M

    barticle Ali , F. , Khalid , M. , Almuhaimeed , A. , Masmoudi , A. , Alghamdi , W. , Yafoz , A. : Ip-gcn: A deep learning model for prediction of insulin using graph convolutional network for diabetes drug design . Journal of Computational Science 81 , 102388 ( 2024 ) barticle

    2024

  44. [52]

    , Zheng , Z

    barticle Wang , F. , Zheng , Z. , Zhang , Y. , Li , Y. , Yang , K. , Zhu , C. : To see further: Knowledge graph-aware deep graph convolutional network for recommender systems . Information Sciences 647 , 119465 ( 2023 ) barticle

    2023

  45. [53]

    , Stich , S.U

    botherref Ajalloeian , A. , Stich , S.U. : On the convergence of sgd with biased gradients. arXiv preprint arXiv:2008.00051 (2020) botherref

    2020 arXiv

  46. [54]

    , Yin , W

    barticle Xu , Y. , Yin , W. : Block stochastic gradient iteration for convex and nonconvex optimization . SIAM Journal on Optimization 25 ( 3 ), 1686 -- 1716 ( 2015 ) barticle

    2015

  47. [55]

    , Carmon , Y

    barticle Arjevani , Y. , Carmon , Y. , Duchi , J.C. , Foster , D.J. , Srebro , N. , Woodworth , B. : Lower bounds for non-convex stochastic optimization . Mathematical Programming 199 ( 1 ), 165 -- 214 ( 2023 ) barticle

    2023

  48. [56]

    , Hou , Z

    botherref Cen , Y. , Hou , Z. , Wang , Y. , Chen , Q. , Luo , Y. , Yu , Z. , Zhang , H. , Yao , X. , Zeng , A. , Guo , S. , Dong , Y. , Yang , Y. , Zhang , P. , Dai , G. , Wang , Y. , Zhou , C. , Yang , H. , Tang , J. : Cogdl: A toolkit for deep learning on graphs. arXiv prepr...

    2021 arXiv

  49. [57]

    , Gross , S

    botherref Paszke , A. , Gross , S. , Massa , F. , Lerer , A. , Bradbury , J. , Chanan , G. , Killeen , T. , Lin , Z. , Gimelshein , N. , Antiga , L. , et al.: Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing system...

    2019

  50. [58]

    , Namata , G

    barticle Sen , P. , Namata , G. , Bilgic , M. , Getoor , L. , Galligher , B. , Eliassi-Rad , T. : Collective classification in network data . AI magazine 29 ( 3 ), 93 -- 93 ( 2008 ) barticle

    2008

  51. [59]

    , Fey , M

    barticle Hu , W. , Fey , M. , Zitnik , M. , Dong , Y. , Ren , H. , Liu , B. , Catasta , M. , Leskovec , J. : Open graph benchmark: Datasets for machine learning on graphs . Advances in neural information processing systems 33 , 22118 -- 22133 ( 2020 ) barticle

    2020

  52. [60]

    sn-basic.bst

    FUNCTION identify.basic.version "sn-basic.bst" " [2024/07/19 v1.1 bibliography style]" * top ENTRY address archive author booktitle chapter doi edition editor eid eprint howpublished institution journal key keywords month note number organization pages publisher school series ...

    2024

  53. [61]

    write newline

    " write newline "" before.all 'output.state := FUNCTION add.period duplicate empty 'skip "." * add.blank if FUNCTION if.digit duplicate "0" = swap duplicate "1" = swap duplicate "2" = swap duplicate "3" = swap duplicate "4" = swap duplicate "5" = swap duplicate "6" = swap dupl...

  54. [62]

    write newline

    " write newline "" before.all 'output.state := FUNCTION string.to.integer 't := t text.length 'k := #1 'char.num := t char.num #1 substring 's := s is.num s "." = or char.num k = not and char.num #1 + 'char.num := while char.num #1 - 'char.num := t #1 char.num substring FUNCTI...

Pith tools

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