Pith. sign in

REVIEW 3 major objections 7 minor 60 references

FedGrAINS: Personalized SubGraph Federated Learning with Adaptive Neighbor Sampling

T0 review · 3 major / 7 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read FedGrAINS shows that learning which neighbors matter per client and per layer, then pruning the message-passing graph to those neighbors, consistently improves personalized subgraph federated learning.

desk verdict A plausible new GFlowNet-based neighbor-sampling idea for subgraph FL with a broad evaluation, but the 'consistently improves all baselines' claim is contradicted by the paper's own tables and the missing random-sampling baseline plus code should be addressed before acceptance. read the letter →

arxiv 2501.12592 v2 pith:PWZVPMVH submitted 2025-01-22 cs.LG cs.AIcs.DCcs.IR

Emir Ceyani , Han Xie , Baturalp Buyukates , Carl Yang , Salman Avestimehr This is my paper · ORCID
classification cs.LGcs.AIcs.DCcs.IR
keywords federatedlearninggraphneuralnetworkspersonalizedsubgraphGFlowNetsadaptiveneighborsamplingnodeclassificationtrajectorybalancenode-degreeheterogeneity
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

Federated learning of graph neural networks becomes unreliable when each client holds only a subgraph of a larger network, because missing cross-client edges distort each node's neighborhood and standard federated averaging does not adapt to that distortion. FedGrAINS is a plug-in regularizer that learns, per client and per message-passing layer, which neighbors actually matter for the client's node-classification task, then prunes the aggregation to those neighbors. The importance model is a generative flow network (GFlowNet) trained with a trajectory balance objective against a reward given by the exponentiated negative classification loss, so the sampling policy is task-driven rather than fixed by graph topology. The paper's central claim, tested on six datasets under disjoint and overlapping subgraph splits, is that adding this sampler improves every subgraph FL baseline it was combined with by at least about one percentage point, with no additional communication cost and no sharing of raw graph data.

What carries the argument

The load-bearing object is a client-local GFlowNet whose state is a sequence of adjacency matrices $s = (A_0, \ldots, A_L)$ and whose forward policy assigns each candidate neighbor an independent Bernoulli inclusion logit. The Gumbel-max trick turns those logits, plus Gumbel noise, into a sample of exactly $k$ neighbors per layer, so the message-passing graph is pruned in a task-aware way. Trajectory balance, with the reward $R(s_L)=\exp(-\alpha \mathcal{L}_{\mathrm{GNN}})$, provides the training signal that aligns the sampling policy with classification performance, and the normalizer $Z(s_0)$ is treated as a constant in that loss. The GFlowNet and the GNN are trained jointly on each client, while only GNN parameters are communicated to the server.

What would settle it

Run FedGrAINS with the same sampling budget $k$ but replace the GFlowNet logits with uniform random logits; if random $k$-neighbor sampling matches the reported accuracy gains, then learned task-aware importance is not what causes the improvements.

Watch

Extended reading notes

Core claim

FedGrAINS claims that heterogeneous node degrees, not just missing links, are what break federated GNN training, and that the right remedy is to replace full-neighborhood aggregation with a learned, per-layer selection of the $k$ most task-relevant neighbors. Each client runs a local GFlowNet whose states are sequences of sampled adjacency matrices; at every layer the forward policy produces per-node Bernoulli inclusion logits, and the Gumbel-max trick converts these into an exact top-$k$ sample. The reward is $R(s_L)=\exp(-\alpha\,\mathcal{L}_{\mathrm{GNN}})$, the exponentiated negative node-classification loss, and the sampler is trained by a trajectory balance loss while the GNN is trained by classification loss. Only the GNN weights are exchanged with the server, so the adaptive sampler is a personalization layer that leaves the communication pattern of the base FL algorithm unchanged. The paper reports that this consistently improves FedAvg, FedPer, GCFL+ and FED-PUB across Cora, CiteSeer, Pubmed, Amazon-Computer, Amazon-Photo and ogbn-arxiv.

Load-bearing premise

The method assumes the total-flow normalizer $Z(s_0)$ is constant and omits it from the trajectory balance loss; if that normalizer actually varies across clients or sampled trajectories, the learned sampling policy is biased and the reported gains are not fully controlled.

Editorial extensions

If this is right

  • On every dataset and split tested, adding FedGrAINS raises node-classification accuracy by at least one percentage point over each base algorithm it was combined with: FedAvg, FedPer, GCFL+, and FED-PUB.
  • The largest gains appear on higher-degree graphs such as Amazon-Computer, Amazon-Photo, and ogbn-arxiv, where neighborhood heterogeneity is most severe.
  • Because only GNN weights are aggregated, FedGrAINS can be layered onto existing subgraph FL methods without extra communication rounds or additional privacy risk.
  • Although accuracy still falls as the number of clients grows, FedGrAINS consistently shrinks that degradation relative to the unmodified baselines.

Reading between the lines

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

  • Beyond the paper: a natural ablation is to replace the GFlowNet logits with uniform random logits at the same budget $k$; if random pruning matches the gains, much of the improvement may be a regularization effect rather than task-aware importance.
  • Beyond the paper: the deferred estimation of $Z(s_0)$ could change the learned policy, so comparing constant-$Z$ training against estimated-$Z$ training would reveal how much of the reported behavior depends on that approximation.
  • Beyond the paper: the same reward-wrapper design should transfer to other federated graph tasks with a scalar loss, such as link prediction or graph classification, because the GFlowNet needs only a tractable reward signal.
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 / 7 minor

Summary. The paper introduces FedGrAINS, a plug-in regularization method for personalized subgraph federated learning. Each client trains a GFlowNet that assigns inclusion probabilities to neighbors, samples k neighbors per GNN layer via the Gumbel-top-k trick, and updates both the GNN classifier and the GFlowNet using a trajectory balance objective whose reward is the classification loss. Experiments on six datasets with disjoint and overlapping client partitions compare several subgraph FL baselines with and without FedGrAINS, reporting mean and standard deviation of accuracy over three runs. The paper claims that adding FedGrAINS consistently improves all baselines and that this comes without extra communication cost or privacy leakage.

Significance. If the central claim held, FedGrAINS would be a practically useful off-the-shelf personalization module for subgraph FL: it adds only local computation, avoids sharing raw graph data, and can be combined with existing methods. The paper also draws attention to node-degree heterogeneity as a distinct challenge in subgraph FL and connects adaptive neighbor sampling to GFlowNet-based structured generative modeling. The paper is honest in its conclusion about the Z(s0) assumption, and the experimental scope across six datasets is broad. However, the evidence presented does not currently establish the headline claim, and the lack of a random-sampling control leaves open the possibility that the gains come from the regularization effect of neighbor subsampling rather than from the learned importance weights.

major comments (3)
  1. [Section 4.2] The claim that 'the inclusion of FedGrAINS as a personalization method consistently improves the performance of all baselines' and that it 'improves the accuracy of all algorithms (except FedSage+) by at least 1%' is contradicted by the paper's own Tables 1 and 2. The following augmented cells report lower mean accuracy than their base counterpart: Table 1, Amazon-Computer 5-client FedAvg (87.59 vs 88.03); Table 1, PubMed 5-client GCFL+ (85.68 vs 85.77); Table 2, PubMed 10-client FedAvg (84.59 vs 84.66); Table 2, PubMed 10-client GCFL+ (84.51 vs 85.08); Table 2, Amazon-Photo 50-client FED-PUB (91.45 vs 91.71). A universal claim requires zero negative differences, and several of these differences are also far below the stated 1% threshold. The text should be revised to a more precise claim, with significance testing or a discussion of the negative cells, or the experiments should be rerun to support the universal statement.
  2. [Eq. (3.5) and Conclusion] The trajectory balance loss in Eq. (3.5) omits the normalizer Z(s0) by treating it as a constant, and the conclusion explicitly says this was assumed 'to reduce the workload at the edge' and defers estimation to future work. No justification is given that the omitted normalizer does not change the fixed point of the TB objective or the learned sampling distribution. Since the reward is the classification loss, a biased sampler could explain both the reported gains and the performance drops. The paper should either provide evidence that the learned sampling policy remains reward-proportional under this assumption or analyze the bias introduced by the omitted term.
  3. [Section 4.2, Tables 1 and 2] The experiments do not include a random neighbor-sampling baseline with the same sampling budget k and the same GNN architecture. Without this control, the observed improvements over full-neighborhood baselines could be attributed to the regularization effect of dropping neighbors rather than to the GFlowNet's learned importance weights. Adding an ablation that uses uniform or Gumbel-top-k sampling with the same k, same GNN, and same number of layers is necessary to isolate the contribution of the adaptive importance component.
minor comments (7)
  1. [Appendix A] The text refers to 'Algorithm 0' twice, but the appendices present Algorithms 1 and 2; the in-text references should be corrected.
  2. [Algorithm 1, line 19] The update for the GFlowNet parameters appears to contain a typo: it assigns to ϕ^i_GNN while using ϕ^i_GFN on the right-hand side, and the gradient is taken with respect to the GNN parameters rather than the GFlowNet parameters.
  3. [Algorithm 1, line 17] The loss is referenced as 'Eq. 3.2', but the FedGrAINS loss is defined in Eq. (3.5); the reference should be fixed.
  4. [Algorithm 1, line 13] The notation K^l is used without definition; the intended set is presumably V_l^k from Eq. (3.6).
  5. [Appendix B.1.1, Tables 5 and 6] Table 5 is captioned 'non-overlapping' while Table 6 is captioned 'disjoint'; the same setting is described with two different terms in the text and table captions, which should be made consistent.
  6. [Appendix B.1] The metric name 'Jenson-Shannon divergence' should be 'Jensen-Shannon divergence'.
  7. [Reproducibility] The paper does not state code availability or list the random seeds used across the three runs; given the complexity of jointly training a GFlowNet and a GNN in a federated loop, code or a detailed per-dataset hyperparameter table would materially improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the GFlowNet sampler and classifier are jointly trained on matching objectives, but no predicted result is equivalent to a fitted input by construction.

full rationale

The paper's claimed derivation chain is an empirical optimization loop: the GFlowNet learns a neighbor-sampling policy via the trajectory balance loss (Eq. 3.5), the reward is exp(-alpha * classification loss) on the GFlowNet-sampled subgraph (Eq. 3.4), and the classifier and sampler are updated jointly (Eq. 3.7). This is coupled reinforcement-learning training, not a circular derivation: the sampler is not fitted to the test metric and then reported as a prediction of it. Evaluation is on held-out nodes under standard federated subgraph benchmarks against external baselines (FedAvg, FedPer, GCFL, FedSage+, FED-PUB). The normalizer Z(s0) is explicitly approximated as constant and the authors disclose this limitation in the Conclusion, which is an approximation/bias concern rather than a circularity. The claim that FedGrAINS 'consistently improves the performance of all baselines' is in tension with some cells in Tables 1 and 2 (e.g., PubMed-10 FedAvg 84.66 vs 84.59), but that is an internal consistency or correctness issue, not circularity. No load-bearing self-citation appears: the GFlowNet methods and subgraph FL baselines cited are external works, and the authors' own prior papers are used only as context or baseline references. The central method therefore stands on its own experimental implementation rather than reducing to its own inputs.

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

The central claim depends on three tuned hyperparameters (alpha, k, beta_GFN) that are not fully reported, plus four explicit modeling assumptions. The most fragile is the constant-Z assumption, which the authors themselves flag. No new physical or conceptual entities are invented.

free parameters (3)
  • reward scaling alpha = 1e5 (tuned on Cora validation; sensitivity range 1e4 to 1e6)
    Multiplies the classification loss in the log-reward. Table 3 shows strong sensitivity, with validation accuracy falling from 82.92% to 55.27% when alpha changes from 1e5 to 1e6.
  • sampling budget k = not reported
    Number of neighbors sampled at each message-passing layer. The authors state they tune it on validation accuracy but do not give per-dataset values.
  • GFlowNet learning rate beta_GFN = 0.001
    The sensitivity analysis shows that beta=1e-2 with alpha=1e5 gives 79.42% instead of 82.92%, so the fixed value is impactful.
assumptions (5)
  • standard math Trajectory balance (TB) drives the forward policy to be proportional to the reward (Eq. 2.3, from Malkin et al. 2022).
    The paper relies on the TB convergence result without proof.
  • domain assumption The state s=(A0,...,A_l) stores the full trajectory, making the backward transition PB(s_l|s_{l-1}) = 1 deterministic.
    Stated in Section 3.1; this is true by construction of the state representation.
  • ad hoc to paper The normalizer Z(s0) is constant and can be omitted from optimization (Eq. 3.5).
    The authors treat Z as a constant and defer learning it. If Z varies across trajectories, this omitted term biases the learned sampling policy. The Conclusion explicitly assumes Z(s0) is constant.
  • ad hoc to paper Gumbel-top-k on independent Bernoulli logits yields a valid behavior policy for off-policy TB training.
    The paper asserts off-policy training works without change, citing off-policy GFlowNets [35], but does not verify the distributional match between the Bernoulli forward model and the Gumbel-top-k sampler.
  • domain assumption The reward exp(-alpha * L_GNN) is well-defined and informative for all sampled subgraphs.
    The classification loss must be finite and sufficiently smooth for the TB regression to be stable. With alpha=1e5, the log-reward target can be extreme.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FedGrAINS: Personalized SubGraph Federated Learning with Adaptive Neighbor Sampling." pith.science (2026). https://pith.science/paper/PWZVPMVH

@misc{pith2026250112592,
  author       = {Pith},
  title        = {Pith review of: FedGrAINS: Personalized SubGraph Federated Learning with Adaptive Neighbor Sampling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PWZVPMVH}},
  note         = {Machine review of arXiv:2501.12592}
}
read the original abstract

Graphs are crucial for modeling relational and biological data. As datasets grow larger in real-world scenarios, the risk of exposing sensitive information increases, making privacy-preserving training methods like federated learning (FL) essential to ensure data security and compliance with privacy regulations. Recently proposed personalized subgraph FL methods have become the de-facto standard for training personalized Graph Neural Networks (GNNs) in a federated manner while dealing with the missing links across clients' subgraphs due to privacy restrictions. However, personalized subgraph FL faces significant challenges due to the heterogeneity in client subgraphs, such as degree distributions among the nodes, which complicate federated training of graph models. To address these challenges, we propose \textit{FedGrAINS}, a novel data-adaptive and sampling-based regularization method for subgraph FL. FedGrAINS leverages generative flow networks (GFlowNets) to evaluate node importance concerning clients' tasks, dynamically adjusting the message-passing step in clients' GNNs. This adaptation reflects task-optimized sampling aligned with a trajectory balance objective. Experimental results demonstrate that the inclusion of \textit{FedGrAINS} as a regularizer consistently improves the FL performance compared to baselines that do not leverage such regularization.

Figures

Figures reproduced from arXiv: 2501.12592 by the authors.

Figure 1
Figure 1. The illustration of the proposed joint training scheme of 2-layer GNN and GFlowNet for our personalized [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

60 extracted references · 55 canonical work pages

  1. [1]

    Ahmed, Z

    K. Ahmed, Z. Zeng, M. Niepert, and G. V. den Broeck, SIMPLE: A gradient estimator for k-subset sampling, in The Eleventh International Conference on Learning Representations, 2023

    work page 2023

  2. [2]

    S. Alam, T. Zhang, T. Feng, H. Shen, Z. Cao, D. Zhao, J. Ko, K. Somasundaram, S. S. Narayanan, S. A vestimehr, et al., Fedaiot: A fed- Copyright © 2025 by SIAM Unauthorized reproduction of this article is prohibited erated learning benchmark for artificial intelligence of things, arXiv preprint arXiv:2310.00109, (2023)

    work page Pith review arXiv 2023

  3. [3]

    M. G. Arivazhagan, V. Aggarwal, A. K. Singh, and S. Choudhary, Federated learning with personal- ization layers , 2019

    work page 2019

  4. [4]

    J. Baek, W. Jeong, J. Jin, J. Yoon, and S. J. Hwang, Personalized subgraph federated learning , in Proceedings of the 40th International Conference on Machine Learning, 2023, pp. 1396–1415

    work page 2023

  5. [5]

    J. Baek, M. Kang, and S. J. Hwang, Accurate learning of graph representations with graph multiset pooling, in 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021, 2021

    work page 2021

  6. [6]

    Bengio, M

    E. Bengio, M. Jain, M. Korablyov, D. Precup, and Y. Bengio, Flow network based generative mod- els for non-iterative diverse candidate generation , Ad- vances in Neural Information Processing Systems, 34 (2021), pp. 27381–27394

    work page 2021

  7. [7]

    Bengio, S

    Y. Bengio, S. Lahlou, T. Deleu, E. J. Hu, M. Ti- wari, and E. Bengio, Gflownet foundations , arXiv preprint arXiv:2111.09266, (2021)

    arXiv 2021

  8. [8]

    V. D. Blondel, J.-L. Guillaume, R. Lambiotte, and E. Lefebvre, Fast unfolding of communities in large networks, Journal of statistical mechanics: theory and experiment, 2008 (2008), p. P10008

    work page 2008

Show all 60 references
  1. [9]

    F. Chen, P. Li, T. Miyazaki, and C. Wu, Fed- graph: Federated graph learning with intelligent sam- pling, IEEE Transactions on Parallel & Distributed Systems, 33 (2022), pp. 1775–1786

    2022

  2. [10]

    F allah, A

    A. F allah, A. Mokhtari, and A. Ozdaglar, Per- sonalized federated learning with theoretical guarantees: A model-agnostic meta-learning approach , Advances in neural information processing systems, 33 (2020), pp. 3557–3568

    2020

  3. [11]

    F ang, Z

    T. F ang, Z. Xiao, C. W ang, J. Xu, X. Yang, and Y. Yang, Dropmessage: Unifying random dropping for graph neural networks , in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 37, 2023

    2023

  4. [12]

    Gilmer, S

    J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, and G. E. Dahl, Neural message pass- ing for quantum chemistry , in Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 2017, vol. 70 of Proceedings of Machin...

    2017

  5. [13]

    Girvan and M

    M. Girvan and M. E. J. Newman, Community struc- ture in social and biological networks , Proceedings of the National Academy of Sciences, 99 (2002), pp. 7821– 7826

    2002

  6. [14]

    Hamilton, Z

    W. Hamilton, Z. Ying, and J. Leskovec, Inductive representation learning on large graphs , in Advances in Neural Information Processing Systems, I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, eds., vol. 30, Curran Associates, Inc., 2017

    2017

  7. [15]

    W. L. Hamilton, Graph representation learning, Syn- thesis Lectures on Artificial Intelligence and Machine Learning, 14 (2020), pp. 1–159

    2020

  8. [16]

    D. K. Hammond, P. V andergheynst, and R. Gri- bonval, Wavelets on graphs via spectral graph the- ory, Applied and Computational Harmonic Analysis, 30 (2011), pp. 129–150

    2011

  9. [17]

    Hasanzadeh, E

    A. Hasanzadeh, E. Hajiramezanali, S. Boluki, M. Zhou, N. Duffield, K. Narayanan, and X. Qian, Bayesian graph neural networks with adap- tive connection sampling , in International conference on machine learning, PMLR, 2020, pp. 4094–4104

    2020

  10. [18]

    C. He, M. Annavaram, and S. A vestimehr, Group knowledge transfer: Federated learning of large cnns at the edge , Advances in Neural Information Processing Systems, 33 (2020), pp. 14068–14080

    2020

  11. [19]

    C. He, K. Balasubramanian, E. Ceyani, C. Yang, H. Xie, L. Sun, L. He, L. Yang, P. S. Yu, Y. Rong, et al., Fedgraphnn: A federated learning system and benchmark for graph neural networks , arXiv preprint arXiv:2104.07145, (2021)

    2021 arXiv

  12. [20]

    C. He, E. Ceyani, K. Balasubramanian, M. An- navaram, and S. A vestimehr, Spreadgnn: Serverless multi-task federated learning for graph neural networks, AAAI, (2022)

    2022

  13. [21]

    Horvath, S

    S. Horvath, S. Laskaridis, M. Almeida, I. Leon- tiadis, S. Venieris, and N. Lane, Fjord: Fair and accurate federated learning under heterogeneous targets with ordered dropout, Advances in Neural Information Processing Systems, 34 (2021), pp. 12876–12889

    2021

  14. [22]

    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 , Advances in neural information processing systems, 33 (2020), pp. 22118–22133

    2020

  15. [23]

    I. A. Huijben, W. Kool, M. B. Paulus, and R. J. V an Sloun, A review of the gumbel-max trick and its extensions for discrete stochasticity in machine learning, IEEE Transactions on Pattern Analysis and Machine Intelligence, 45 (2022), pp. 1353–1371

    2022

  16. [24]

    J. Jo, J. Baek, S. Lee, D. Kim, M. Kang, and S. J. Hwang, Edge representation learning with hy- pergraphs, in Advances in Neural Information Process- ing Systems, M. Ranzato, A. Beygelzimer, Y. Dauphin, P. Liang, and J. W. Vaughan, eds., vol. 34, Curran As- sociates, Inc., 202...

    2021

  17. [25]

    Karypis, Metis: Unstructured graph partitioning and sparse matrix ordering system , Technical report, (1997)

    G. Karypis, Metis: Unstructured graph partitioning and sparse matrix ordering system , Technical report, (1997)

    1997

  18. [26]

    Karypis and V

    G. Karypis and V. Kumar, A fast and high qual- ity multilevel scheme for partitioning irregular graphs , SIAM Journal on scientific Computing, 20 (1998), pp. 359–392

    1998

  19. [27]

    D. P. Kingma and J. Ba, Adam: A method for stochastic optimization , in 3rd International Confer- ence on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Pro- ceedings, 2015

    2015

  20. [28]

    T. N. Kipf and M. Welling, Semi-supervised clas- sification with graph convolutional networks , in 5th In- Copyright © 2025 by SIAM Unauthorized reproduction of this article is prohibited ternational Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-2...

    2025

  21. [29]

    W. Kool, H. V an Hoof, and M. Welling, Stochas- tic beams and where to find them: The gumbel-top-k trick for sampling sequences without replacement, in In- ternational Conference on Machine Learning, PMLR, 2019, pp. 3499–3508

    2019

  22. [30]

    J. Li, S. Tian, R. Wu, L. Zhu, W. Zhao, C. Meng, L. Chen, Z. Zheng, and H. Yin, Less can be more: Unsupervised graph pruning for large-scale dynamic graphs, arXiv preprint arXiv:2305.10673, (2023)

    2023 arXiv

  23. [31]

    Q. Li, Z. Han, and X.-M. Wu , Deeper insights into graph convolutional networks for semi-supervised learning, in Proceedings of the AAAI conference on artificial intelligence, vol. 32, 2018

    2018

  24. [32]

    Q. Li, Z. Wen, Z. Wu, S. Hu, N. W ang, Y. Li, X. Liu, and B. He, A survey on federated learning systems: vision, hype and reality for data privacy and protection, IEEE Transactions on Knowledge and Data Engineering, (2021)

    2021

  25. [33]

    T. Li, A. K. Sahu, M. Zaheer, M. Sanjabi, A. Tal- walkar, and V. Smith, Federated optimization in heterogeneous networks , in Proceedings of Machine Learning and Systems 2020, MLSys 2020, Austin, TX, USA, March 2-4, 2020, mlsys.org, 2020

    2020

  26. [34]

    Malkin, M

    N. Malkin, M. Jain, E. Bengio, C. Sun, and Y. Bengio, Trajectory balance: Improved credit as- signment in gflownets, Advances in Neural Information Processing Systems, 35 (2022), pp. 5955–5967

    2022

  27. [35]

    Malkin, S

    N. Malkin, S. Lahlou, T. Deleu, X. Ji, E. Hu, K. Everett, D. Zhang, and Y. Bengio, Gflownets and variational inference , arXiv preprint arXiv:2210.00580, (2022)

    2022 arXiv

  28. [36]

    McAuley, C

    J. McAuley, C. Targett, Q. Shi, and A. V an Den Hengel, Image-based recommendations on styles and substitutes , in Proceedings of the 38th interna- tional ACM SIGIR conference on research and devel- opment in information retrieval, 2015, pp. 43–52

    2015

  29. [37]

    H. B. McMahan, E. Moore, D. Ramage, S. Hamp- son, and B. A. y Arcas, Communication-efficient learning of deep networks from decentralized data , in AISTATS, 2017

    2017

  30. [38]

    L. Pan, D. Zhang, A. Courville, L. Huang, and Y. Bengio, Generative augmented flow networks , in The Eleventh International Conference on Learning Representations, 2023

    2023

  31. [39]

    Radicchi, C

    F. Radicchi, C. Castellano, F. Cecconi, V. Loreto, and D. Parisi, Defining and identifying communities in networks , Proceedings of the national academy of sciences, 101 (2004), pp. 2658–2663

    2004

  32. [40]

    Y. Rong, W. Huang, T. Xu, and J. Huang , Dropedge: Towards deep graph convolutional networks on node classification , in International Conference on Learning Representations, 2020

    2020

  33. [41]

    Sattler, K.-R

    F. Sattler, K.-R. M¨uller, and W. Samek, Clus- tered federated learning: Model-agnostic distributed multitask optimization under privacy constraints, IEEE transactions on neural networks and learning systems, 32 (2020), pp. 3710–3722

    2020

  34. [42]

    P. Sen, G. Namata, M. Bilgic, L. Getoor, B. Gal- ligher, and T. Eliassi-Rad, Collective classification in network data , AI magazine, 29 (2008), pp. 93–93

    2008

  35. [43]

    Shchur, M

    O. Shchur, M. Mumme, A. Bojchevski, and S. G¨unnemann, Pitfalls of graph neural network eval- uation, arXiv preprint arXiv:1811.05868, (2018)

    2018 arXiv

  36. [44]

    Srivastava, G

    N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov, Dropout: A simple way to prevent neural networks from overfit- ting, Journal of Machine Learning Research, 15 (2014), pp. 1929–1958

    2014

  37. [45]

    Y. Tan, Y. Liu, G. Long, J. Jiang, Q. Lu, and C. Zhang, Federated learning on non-iid graphs via structural knowledge sharing , arXiv preprint arXiv:2211.13009, (2022)

    2022 arXiv

  38. [46]

    W ang, W

    Z. W ang, W. Kuang, Y. Xie, L. Yao, Y. Li, B. Ding, and J. Zhou , Federatedscope-gnn: To- wards a unified, comprehensive and efficient package for federated graph learning , in KDD ’22: The 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Washington, DC, USA,...

    2022

  39. [47]

    C. Wu, F. Wu, Y. Cao, Y. Huang, and X. Xie, Fedgnn: Federated graph neural network for privacy- preserving recommendation, KDD, (2021)

    2021

  40. [48]

    Z. Wu, S. Pan, F. Chen, G. Long, C. Zhang, and P. S. Yu, A comprehensive survey on graph neural networks, IEEE Trans. Neural Networks Learn. Syst., 32 (2021), pp. 4–24

    2021

  41. [49]

    H. Xie, J. Ma, L. Xiong, and C. Yang, Federated graph classification over non-iid graphs , in Advances in Neural Information Processing Systems, vol. 34, Curran Associates, Inc., 2021, pp. 18839–18852

    2021

  42. [50]

    S. M. Xie and S. Ermon, Reparameterizable subset sampling via continuous relaxations , arXiv preprint arXiv:1901.10517, (2019)

    2019 arXiv

  43. [51]

    Yao and C

    Y. Yao and C. Joe-Wong , Fedgcn: Conver- gence and communication tradeoffs in federated train- ing of graph convolutional networks , arXiv preprint arXiv:2201.12433, (2022)

    2022 arXiv

  44. [52]

    Zhang, C

    K. Zhang, C. Yang, X. Li, L. Sun, and S. M. Yiu, Subgraph federated learning with missing neighbor gen- eration, in Advances in Neural Information Process- ing Systems, vol. 34, Curran Associates, Inc., 2021, pp. 6671–6682

    2021

  45. [53]

    J. Zhou, G. Cui, S. Hu, Z. Zhang, C. Yang, Z. Liu, L. W ang, C. Li, and M. Sun, Graph neural networks: A review of methods and applications , AI Open, 1 (2020), pp. 57–81. Copyright © 2025 by SIAM Unauthorized reproduction of this article is prohibited A Algorithm Outline In t...

    2020

  46. [54]

    Local: This method is the non-FL baseline, which only locally trains the model for each client without weight sharing

  47. [55]

    F edAvg: This method [37] is the FL baseline, where each client locally updates a model and sends it to a server, while the server aggregates the locally updated models with respect to their numbers of training samples and transmits the aggregated model back to the clients

  48. [56]

    F edPer: This method [3] is the personalized FL baseline, which shares only the base layers, while keeping the personalized classification layers in the local side

  49. [57]

    F edSage+: This method [52] is the subgraph FL baseline, which expands local subgraphs by generating additional nodes with the local graph generator. To train the graph generator, each client first receives node representations from other clients, and then calculates the gradi...

  50. [58]

    In particular, it uses the bi-partitioning scheme, which divides a set of clients into two disjoint client groups based on their gradient similarities

    GCFL: This method [49] is the graph FL baseline, which targets completely disjoint graphs (e.g., molecular graphs) as in image tasks. In particular, it uses the bi-partitioning scheme, which divides a set of clients into two disjoint client groups based on their gradient simil...

  51. [59]

    FED-PUB: This method [4] performs personalized aggregation based on subgraph similarities and models’ functional embeddings for discovering community structures, but also adaptively masks the received weights from the server to filter irrelevant weights from heterogeneous communities

  52. [60]

    FedGrAINS: Our plug-and-play personalized subgraph federated learning model, which employs a personalized GFlowNet for each client. For each node in the client’s graph, the GFlowNet estimates the important nodes over the k−hop neighborhood and samples using Gumbel-max trick [2...

    2025

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.