Pith. sign in

REVIEW 2 major objections 5 minor 1 cited by

Demystifying Distributed Training of Graph Neural Networks for Link Prediction

T0 review · 2 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read The paper argues that distributed GNN link prediction degrades because partitioning breaks neighbor information and negative sampling is local, then fixes both with sparsification.

desk verdict Useful system insight and a plausible diagnostic claim, but the training/eval asymmetry between positive and negative neighborhoods could inflate the headline accuracy; a symmetric-neighborhood ablation should gate the claim. read the letter →

arxiv 2506.20818 v1 pith:3U2DPN36 submitted 2025-06-25 cs.LG

classification cs.LG
keywords distributedGNNtraininglinkpredictionnegativesamplinggraphpartitioningsparsificationeffectiveresistancecommunicationcostHits@100
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 asks why link prediction accuracy drops when a graph neural network is trained in a distributed setting, where each worker sees only a partitioned subgraph. It argues that the drop has two root causes: graph partitioning destroys part of each node's neighbor information, and workers can only draw negative samples (node pairs that are not connected by an edge) from their own partition, so the model never learns from global negative pairs. Sharing the full graph with every worker fixes the accuracy but transfers thousands of gigabytes, which defeats the purpose of distributed training. SpLPG instead keeps the full neighbor list of every node in the worker's own partition and replaces the other partitions with sparsified versions, so global negative samples can still be drawn without transferring full remote neighborhoods. On nine datasets the paper reports that this cuts communication overhead by up to about 80% while keeping accuracy close to that of centralized training.

What carries the argument

The central mechanism is negative sampling over sparsified remote subgraphs. Each worker draws positive samples from its own unsimplified partition, while negative destinations are chosen from the entire graph, with remote partitions stored in shared memory in sparsified form. The sparsification is built on the effective-resistance identity: an edge is sampled with probability $p_{u,v}\propto 1/d_u+1/d_v$, a degree-only approximation of effective resistance that the paper justifies through a spectral sparsification theorem and a degree-based bound on resistance. Because a global negative sample's endpoint in a remote partition now has far fewer $k$-hop neighbors in the computational graph, the bytes transferred per negative sample drop sharply, while the sample space for choosing negative destinations remains the entire graph.

What would settle it

Train exactly the SpLPG procedure, then evaluate on a test set where every negative pair is represented with full $k$-hop neighborhoods for both endpoints (or, symmetrically, train with sparsified neighborhoods for positives too); if Hits@100 drops well below the centralized baseline, the accuracy gain is an artifact of the density difference between positive and negative samples rather than better link prediction.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that the performance degradation in distributed GNN link prediction is not mainly a data-distribution mismatch introduced by partitioning; it is the combination of partitioned neighbor loss and a shrunken negative-sample space. The authors show that giving every worker access to the whole graph restores centralized-level accuracy, then propose SpLPG to obtain the same effect more cheaply: each worker retains full neighborhoods for its own partition, while remote partitions are sparsified by sampling edges with probability proportional to a degree-based approximation of effective resistance and are used only when drawing negative samples. The experimental claim is that this design recovers centralized-level Hits@100 on most datasets and reduces communication traffic by up to about 80% compared to full data sharing, with the residual accuracy gap concentrated on small graphs such as Citeseer and Cora.

Load-bearing premise

The accuracy claim rests on the assumption that a model trained with full neighborhoods for positive samples but sparsified neighborhoods for negative samples does not learn to exploit that density difference as a shortcut; if it does, the reported Hits@100 would overstate true link-prediction ability.

Editorial extensions

If this is right

  • If the reported results hold, distributed GNN link prediction no longer needs full remote neighborhoods on every worker; sparsified copies suffice for accurate global negative sampling.
  • The ablation showing that both full local neighbors and global negative samples are needed gives a concrete diagnostic: a distributed system that fixes only one of the two will still see degraded accuracy.
  • The communication saving is tunable: at about 85% edge removal, the paper measures roughly 68% cost savings, and more aggressive sparsification trades some accuracy for larger savings.
  • Because the experiments cover GCN, GraphSAGE, GAT, and GATv2, the method is presented as architecture-agnostic rather than tied to one GNN variant.

Reading between the lines

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

  • A natural next experiment is to test SpLPG on graphs where remote $k$-hop neighborhoods are precomputed and cached; part of the reported communication saving may shift to storage rather than disappear.
  • The positive/negative asymmetry in neighborhood density is a confound worth checking directly: if positive samples always see full neighborhoods while negative samples see sparsified ones, Hits@100 could partly reflect a density cue. Evaluating test samples with both endpoints expanded to full neighborhoods would settle whether the score is genuine link prediction.
  • The fixed global sparsification level $\alpha=0.15$ leaves small graphs behind; an adaptive per-partition sparsity based on local density is a testable extension that might close the remaining gap on Cora and Citeseer.
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

2 major / 5 minor

Summary. The paper studies distributed training of GNNs for link prediction and identifies two root causes of accuracy degradation: information loss from graph partitioning and the restricted negative-sample space when workers only access local subgraphs. The authors propose SpLPG, which keeps full local neighborhoods for positive-sample computation, sparsifies remote subgraphs used only for negative sampling, and shares these sparsified subgraphs across workers to reduce communication cost. Experiments on nine datasets with GCN, GraphSAGE, GAT, and GATv2 report that SpLPG reduces communication overhead by up to about 80% while roughly preserving centralized link-prediction accuracy.

Significance. If the accuracy-preservation claim holds, SpLPG is a practical contribution to distributed GNN training for link prediction, where existing frameworks are mainly optimized for node classification. The paper provides code, extensive experiments across multiple models and datasets, and a useful ablation (Fig. 12) isolating the effects of full-neighbor information and global negative sampling. However, the central claim is weakened by a potential confound in the training/evaluation protocol: positive and negative samples have asymmetric neighborhood densities, and the test-time graph is unspecified. The theoretical part (Theorems 1–2) is taken from prior work, and the proposed sparsification uses only a degree-based approximation, so the contribution is primarily empirical. Because of the confound, the significance is conditional on the outcome of a symmetric-neighborhood test.

major comments (2)
  1. [Algorithm 1, Lines 20–21; §V-B; footnote 2] Same as above.
  2. [§V-B, Fig. 11] Same as above.
minor comments (5)
  1. [Figure 4] Same as above.
  2. [§V-A] Same as above.
  3. [Abstract; Table III] Same as above.
  4. [§V-A, Fig. 11] Same as above.
  5. [§IV-A, Fig. 6] Same as above.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found; the sparsification theory is imported from external mathematical results, the root-cause claims are empirical and benchmarked against centralized training, and no fitted quantity is renamed as a prediction.

full rationale

I walked the derivation chain rather than the narrative. SpLPG partitions the graph with METIS, keeps each worker's full local neighborhood, sparsifies remote subgraphs by effective-resistance sampling, and draws global negative samples from sparsified remote subgraphs. The spectral guarantee (Theorem 1) is quoted from Spielman-Srivastava [34] and the degree-based resistance bound (Theorem 2) from Lovasz [39]; both are external, machine-independent results that do not depend on SpLPG's own accuracy claims. The paper's main claims are empirical: the performance drop is diagnosed by comparing distributed variants with centralized training (Figs. 3, 4, 12), and the accuracy/communication trade-off is measured against complete-data-sharing baselines and centralized models (Figs. 8, 10, 11). No equation is defined in terms of the quantity it is said to predict, and no fitted parameter is relabeled as a prediction; the sparsification level alpha is a hyperparameter explored in Table III, not a parameter fitted to the reported Hits@100 values. The only self-citations ([8], [10]) appear as introductory examples of GNN-related work and are not load-bearing. The nearest passage to a caveat is in Sec. IV-A and footnote 2, where the paper concedes that each global negative sample from outside the assigned subgraph 'comes with a much fewer number of k-hop neighbors' after sparsification. That asymmetry between positive and negative neighborhoods could confound the accuracy comparison if test-time embeddings are computed in the same way, but it is an evaluation-validity concern, not a circularity: it does not make any derivation equivalent to its own inputs. I therefore find no significant circularity.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central claims depend on three unstated assumptions: the validity of the degree-based sparsification as a stand-in for effective-resistance sampling, the match between training and test evaluation protocols, and the fidelity of the shared-memory byte count as a communication metric. These are domain assumptions rather than invented physical entities.

free parameters (1)
  • sparsification level alpha = 0.15
    Controls the number of sampled edges L_i = alpha * |E_i| in each sparsified subgraph, removing about 85% of edges. Chosen by hand based on a Cora ablation (Table III) and then used as the default across all datasets.
assumptions (3)
  • domain assumption Sparsifying remote subgraphs by sampling edges with probability proportional to 1/d_u + 1/d_v preserves enough graph structure for negative-sample neighbor aggregation to yield an unbiased training signal.
    The paper replaces exact effective resistance with the degree-based bound from Theorem 2 without proving the spectral guarantee still holds for this approximation, and assumes sparsified negative-sample neighborhoods do not create a shortcut. See Sec IV-A and footnote 2.
  • domain assumption Random 80/10/10 splits produce validation/test sets whose negative sampling matches the training distribution under the same distributed protocol.
    The test-time graph used for computing node embeddings is unspecified, so the reported Hits@100 depends on an unstated match between training and test evaluation. See Sec V-A.
  • domain assumption The communication cost metric (total bytes transferred from master shared memory to workers per epoch) is a faithful proxy for real distributed communication overhead.
    Experiments use shared memory on a single machine, so the claim of 'communication reduction up to 80%' rests on this proxy rather than measured network traffic in a multi-machine cluster. See Sec V-B.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Demystifying Distributed Training of Graph Neural Networks for Link Prediction." pith.science (2026). https://pith.science/paper/3U2DPN36

@misc{pith2026250620818,
  author       = {Pith},
  title        = {Pith review of: Demystifying Distributed Training of Graph Neural Networks for Link Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3U2DPN36}},
  note         = {Machine review of arXiv:2506.20818}
}
read the original abstract

Graph neural networks (GNNs) are powerful tools for solving graph-related problems. Distributed GNN frameworks and systems enhance the scalability of GNNs and accelerate model training, yet most are optimized for node classification. Their performance on link prediction remains underexplored. This paper demystifies distributed training of GNNs for link prediction by investigating the issue of performance degradation when each worker trains a GNN on its assigned partitioned subgraph without having access to the entire graph. We discover that the main sources of the issue come from not only the information loss caused by graph partitioning but also the ways of drawing negative samples during model training. While sharing the complete graph information with each worker resolves the issue and preserves link prediction accuracy, it incurs a high communication cost. We propose SpLPG, which effectively leverages graph sparsification to mitigate the issue of performance degradation at a reduced communication cost. Experiment results on several public real-world datasets demonstrate the effectiveness of SpLPG, which reduces the communication overhead by up to about 80% while mostly preserving link prediction accuracy.

Figures

Figures reproduced from arXiv: 2506.20818 by the authors.

Figure 1
Figure 1. Illustration of (a) link prediction and (b) the computational graph of [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Illustration of GNNs for link prediction. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Link prediction accuracy of GraphSAGE models trained by the state-of-the-art methods. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Accuracy and communication cost of the state-of-the-art methods with the complete data-sharing strategy. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Negative samples. edges.1 Nonetheless, as shown in [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 6
Figure 6. Figure 6: Accuracy of GNNs w/ and w/o graph sparsification. [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 7
Figure 7. Figure 7: Illustration of SpLPG. easily obtained, this approximation of r(u,v) is straightforward to compute. Thus, from Theorems 1 and 2, the graph spar￾sification algorithm based on the approximation of effective resistance can be used to sparsify a graph efficiently. It is wo…
Figure 8
Figure 8. Figure 8: Improvement of communication cost achieved by SpLPG for (a)–(c) GCN and (d)–(f) GraphSAGE. [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Improvement of communication cost achieved by SpLPG over [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 10
Figure 10. Figure 10: Improvement of accuracy achieved by SpLPG for (a)–(c) GCN and (d)–(f) GraphSAGE. [PITH_FULL_IMAGE:figures/full_fig_p009_10.png]
Figure 11
Figure 11. Figure 11: Accuracy of GNNs trained by SpLPG. of each node in each partitioned subgraph managed by each worker to eliminate the need for transferring them during model training, which further reduces the communication cost. SpLPG introduces an additional cost due to graph sparsi…
Figure 13
Figure 13. Figure 13: Impact of batch size. TABLE III IMPACT OF SPARSIFICATION LEVEL Communication cost saving Accuracy p= 4 p= 8 p= 16 p= 4 p= 8 p= 16 α = 0.05 82.3% 82.5% 82.9% 0.716 0.701 0.688 α = 0.10 75.5% 75.3% 75.0% 0.722 0.713 0.711 α = 0.15 68.3% 67.6% 68.1% 0.725 0.719 0.718 α =…
Figure 14
Figure 14. Figure 14: Accuracy of different GNN models on (a)–(d) Cora and (e)–(h) [PITH_FULL_IMAGE:figures/full_fig_p010_14.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Graph Neural Network for Product Recommendation on the Amazon Co-purchase Graph

    cs.IR 2025-08 unverdicted novelty 2.0 of 10

    A benchmark of four GNN architectures (LightGCN, GraphSAGE, GAT, PinSAGE) for link prediction on the Amazon co-purchase graph, reporting trade-offs between accuracy, training cost, and scalability.

Reference graph

Works this paper leans on

44 extracted references · 40 canonical work pages · cited by 1 Pith paper

  1. [1]

    The link prediction problem for social networks,

    D. Liben-Nowell and J. Kleinberg, “The link prediction problem for social networks,” inACM CIKM, 2003, pp. 556–559

  2. [2]

    Learning entity and relation embeddings for knowledge graph completion,

    Y . Lin, Z. Liu, M. Sun, Y . Liu, and X. Zhu, “Learning entity and relation embeddings for knowledge graph completion,” inAAAI, 2015

  3. [3]

    Link prediction in complex networks: A survey,

    L. L ¨u and T. Zhou, “Link prediction in complex networks: A survey,” Physica A: Statistical Mechanics and its Applications, pp. 1150–1170, 2011

  4. [4]

    Link prediction tech- niques, applications, and performance: A survey,

    A. Kumar, S. S. Singh, K. Singh, and B. Biswas, “Link prediction tech- niques, applications, and performance: A survey,”Physica A: Statistical Mechanics and its Applications, p. 124289, 2020

  5. [5]

    Graph neural networks: foundation, frontiers and applications,

    L. Wu, P. Cui, J. Pei, L. Zhao, and X. Guo, “Graph neural networks: foundation, frontiers and applications,” inACM KDD, 2022, pp. 4840– 4841

  6. [6]

    Ma and J

    Y . Ma and J. Tang,Deep learning on graphs. Cambridge University Press, 2021

  7. [7]

    Graph representation learning,

    W. L. Hamilton, “Graph representation learning,”Synthesis Lectures on Artificial Intelligence and Machine Learning, pp. 1–159, 2020

  8. [8]

    Characterizing the efficiency of graph neural network frameworks with a magnifying glass,

    X. Huang, J. Kim, B. Rees, and C.-H. Lee, “Characterizing the efficiency of graph neural network frameworks with a magnifying glass,” inIEEE IISWC, 2022, pp. 160–170

Show all 44 references
  1. [9]

    Heterogeneous spatio- temporal graph convolution network for traffic forecasting with missing values,

    W. Zhong, Q. Suo, X. Jia, A. Zhang, and L. Su, “Heterogeneous spatio- temporal graph convolution network for traffic forecasting with missing values,” inIEEE ICDCS, 2021, pp. 707–717

  2. [10]

    GRAFICS: Graph embedding-based floor identification using crowdsourced RF signals,

    W. Zhuo, Z. Zhao, K. H. Chiu, S. Li, S. Ha, C.-H. Lee, and S.-H. G. Chan, “GRAFICS: Graph embedding-based floor identification using crowdsourced RF signals,” inIEEE ICDCS, 2022, pp. 1051–1061

  3. [11]

    Embedding communication for federated graph neural networks with privacy guarantees,

    X. Wu, Z. Ji, and C.-L. Wang, “Embedding communication for federated graph neural networks with privacy guarantees,” inIEEE ICDCS, 2023, pp. 305–315

  4. [12]

    Mega: More efficient graph attention for GNNs,

    W. Deng and J. Rao, “Mega: More efficient graph attention for GNNs,” inIEEE ICDCS, 2024, pp. 71–81

  5. [13]

    Semi-supervised classification with graph convolutional networks,

    T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” inICLR, 2017

  6. [14]

    Graph attention networks,

    P. Veli ˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P. Li `o, and Y . Bengio, “Graph attention networks,” inICLR, 2018

  7. [15]

    Inductive representation learning on large graphs,

    W. Hamilton, Z. Ying, and J. Leskovec, “Inductive representation learning on large graphs,” inNeurIPS, 2017, pp. 1025–1035

  8. [16]

    How powerful are graph neural networks?

    K. Xu, W. Hu, J. Leskovec, and S. Jegelka, “How powerful are graph neural networks?” inICLR, 2018

  9. [17]

    Position-aware graph neural net- works,

    J. You, R. Ying, and J. Leskovec, “Position-aware graph neural net- works,” inICML, 2019, pp. 7134–7143

  10. [18]

    Graph convolutional neural networks for web-scale rec- ommender systems,

    R. Ying, R. He, K. Chen, P. Eksombatchai, W. L. Hamilton, and J. Leskovec, “Graph convolutional neural networks for web-scale rec- ommender systems,” inACM KDD, 2018, pp. 974–983

  11. [19]

    Modeling relational data with graph convolutional networks,

    M. Schlichtkrull, T. N. Kipf, P. Bloem, R. Van Den Berg, I. Titov, and M. Welling, “Modeling relational data with graph convolutional networks,” inExtended Semantic Web Conference, 2018, pp. 593–607

  12. [20]

    Fast graph representation learning with PyTorch Geometric,

    M. Fey and J. E. Lenssen, “Fast graph representation learning with PyTorch Geometric,” inICLR Workshop on Representation Learning on Graphs and Manifolds, 2019

  13. [21]

    DistDGL: Distributed graph neural network training for billion-scale graphs,

    D. Zheng, C. Ma, M. Wang, J. Zhou, Q. Su, X. Song, Q. Gan, Z. Zhang, and G. Karypis, “DistDGL: Distributed graph neural network training for billion-scale graphs,” inIEEE/ACM IA3, 2020, pp. 36–44

  14. [22]

    Distributed hybrid CPU and GPU training for graph neural networks on billion-scale heterogeneous graphs,

    D. Zheng, X. Song, C. Yang, D. LaSalle, and G. Karypis, “Distributed hybrid CPU and GPU training for graph neural networks on billion-scale heterogeneous graphs,” inACM KDD, 2022, pp. 4582–4591

  15. [23]

    DistGNN: Scalable distributed training for large-scale graph neural networks,

    V . Md, S. Misra, G. Ma, R. Mohanty, E. Georganas, A. Heinecke, D. Kalamkar, N. K. Ahmed, and S. Avancha, “DistGNN: Scalable distributed training for large-scale graph neural networks,” inACM/IEEE SC, 2021, pp. 1–14

  16. [24]

    BGL: GPU-Efficient GNN training by optimizing graph data I/O and preprocessing,

    T. Liu, Y . Chen, D. Li, C. Wu, Y . Zhu, J. He, Y . Peng, H. Chen, H. Chen, and C. Guo, “BGL: GPU-Efficient GNN training by optimizing graph data I/O and preprocessing,” inUSENIX NSDI, 2023, pp. 103–118

  17. [25]

    Adaptive message quantization and parallelization for distributed full-graph GNN training,

    B. Wan, J. Zhao, and C. Wu, “Adaptive message quantization and parallelization for distributed full-graph GNN training,”MLSys, 2023

  18. [26]

    Simplifying distributed neural network training on massive graphs: Randomized partitions improve model aggregation,

    J. Zhu, A. A. Reganti, E. Huang, C. Dickens, N. Rao, K. Subbian, and D. Koutra, “Simplifying distributed neural network training on massive graphs: Randomized partitions improve model aggregation,” inICML Workshop on Localized Learning, 2023

  19. [27]

    Multilevel k-way partitioning scheme for irregular graphs,

    G. Karypis and V . Kumar, “Multilevel k-way partitioning scheme for irregular graphs,” inJ. Parallel Distrib. Comput., 1998, pp. 96–129

  20. [28]

    DeepWalk: Online learning of social representations,

    B. Perozzi, R. Al-Rfou, and S. Skiena, “DeepWalk: Online learning of social representations,” inACM KDD, 2014, pp. 701–710

  21. [29]

    node2vec: Scalable feature learning for networks,

    A. Grover and J. Leskovec, “node2vec: Scalable feature learning for networks,” inACM KDD, 2016, pp. 855–864

  22. [30]

    Does negative sampling matter? a review with insights into its theory and applications,

    Z. Yang, M. Ding, T. Huang, Y . Cen, J. Song, B. Xu, Y . Dong, and J. Tang, “Does negative sampling matter? a review with insights into its theory and applications,”IEEE Trans. Pattern Anal. Mach. Intell., 2024

  23. [31]

    Under- standing negative sampling in graph representation learning,

    Z. Yang, M. Ding, C. Zhou, H. Yang, J. Zhou, and J. Tang, “Under- standing negative sampling in graph representation learning,” inACM KDD, 2020, pp. 1666–1676

  24. [32]

    Learn locally, correct globally: A distributed algorithm for training graph neural networks,

    M. Ramezani, W. Cong, M. Mahdavi, M. Kandemir, and A. Sivasub- ramaniam, “Learn locally, correct globally: A distributed algorithm for training graph neural networks,” inICLR, 2022

  25. [33]

    Demystifying graph sparsification algorithms in graph properties preservation,

    Y . Chen, H. Ye, S. Vedula, A. Bronstein, R. Dreslinski, T. Mudge, and N. Talati, “Demystifying graph sparsification algorithms in graph properties preservation,”Proceedings of the VLDB Endowment, pp. 427– 440, 2023

  26. [34]

    Graph sparsification by effective resistances,

    D. A. Spielman and N. Srivastava, “Graph sparsification by effective resistances,” inACM STOC, 2008, pp. 563–568

  27. [35]

    LightNE: A lightweight graph processing system for network embedding,

    J. Qiu, L. Dhulipala, J. Tang, R. Peng, and C. Wang, “LightNE: A lightweight graph processing system for network embedding,” inACM SIGMOD, 2021, pp. 2281–2289

  28. [36]

    DSpar: An embarrassingly simple strategy for efficient GNN training and inference via degree-based sparsification,

    Z. Liu, K. Zhou, Z. Jiang, L. Li, R. Chen, S.-H. Choi, and X. Hu, “DSpar: An embarrassingly simple strategy for efficient GNN training and inference via degree-based sparsification,”Transactions on Machine Learning Research, 2023

  29. [37]

    Sur- vey on graph neural network acceleration: An algorithmic perspective,

    X. Liu, M. Yan, L. Deng, G. Li, X. Ye, D. Fan, S. Pan, and Y . Xie, “Sur- vey on graph neural network acceleration: An algorithmic perspective,” inIJCAI, 2022, pp. 5521–5529

  30. [38]

    Open Graph Benchmark: Datasets for machine learning on graphs,

    W. Hu, M. Fey, M. Zitnik, Y . Dong, H. Ren, B. Liu, M. Catasta, and J. Leskovec, “Open Graph Benchmark: Datasets for machine learning on graphs,” inNeurIPS, 2020, pp. 22 118–22 133

  31. [39]

    Random walks on graphs: A survey,

    L. Lov ´asz, “Random walks on graphs: A survey,”Combinatorics, Paul Erd˝os Is Eighty, pp. 1–46, 1993

  32. [40]

    On the convergence of FedAvg on Non-IID data,

    X. Li, K. Huang, W. Yang, S. Wang, and Z. Zhang, “On the convergence of FedAvg on Non-IID data,” inICLR, 2019

  33. [41]

    PyTorch: An imperative style, high-performance deep learning library,

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antigaet al., “PyTorch: An imperative style, high-performance deep learning library,” inNeurIPS, 2019, pp. 8024–8035

  34. [42]

    Deep Graph Library: A graph-centric, highly-performant package for graph neural networks,

    M. Wang, D. Zheng, Z. Ye, Q. Gan, M. Li, X. Song, J. Zhou, C. Ma, L. Yu, Y . Gai, T. Xiao, T. He, G. Karypis, J. Li, and Z. Zhang, “Deep Graph Library: A graph-centric, highly-performant package for graph neural networks,”arXiv preprint arXiv:1909.01315, 2019

  35. [43]

    PyTorch distributed: Experi- ences on accelerating data parallel draining,

    S. Li, Y . Zhao, R. Varma, O. Salpekar, P. Noordhuis, T. Li, A. Paszke, J. Smith, B. Vaughan, P. Damaniaet al., “PyTorch distributed: Experi- ences on accelerating data parallel draining,”Proceedings of the VLDB Endowment, pp. 3005–3018, 2020

  36. [44]

    How attentive are graph attention networks?

    S. Brody, U. Alon, and E. Yahav, “How attentive are graph attention networks?” inICLR, 2022

Pith tools

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