Pith. sign in

REVIEW 4 major objections 6 minor 69 references

DUPLEX claims that jointly optimizing network topology and per-worker graph sampling ratios—as one coordinated configuration rather than two separate techniques—cuts completion time by 20.1–48.8% and communication cost by 16.7–37.6% while i

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

DUPLEX jointly optimizes P2P network topology and per-device graph sampling ratios via deep reinforcement learning, cutting completion time by 20-49% and communication cost by 17-38% while improving accuracy.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection DUPLEX has a genuinely new joint-optimization idea for decentralized federated graph learning, but the consensus-distance estimator in Eq. (15) is structurally flawed and the evaluation has a baseline-selection problem. the 4 major comments →

arxiv 2509.08409 v1 pith:J5JRG5CU submitted 2025-09-10 cs.DC

Towards Communication-Efficient Decentralized Federated Graph Learning over Non-IID Data

classification cs.DC
keywords decentralized federated graph learningnetwork topology constructiongraph neighbor samplingnon-IID dataconsensus distancedeep reinforcement learningcommunication efficiencypeer-to-peer networks
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

Decentralized federated graph learning trains graph neural networks across devices without a central parameter server, but the price is heavy exchange of node embeddings between workers. Existing systems attack that cost either by pruning the peer-to-peer topology or by sampling fewer graph neighbors, and the paper claims these two levers interact: changing one changes what the other should do. Directly combining the two existing techniques, the paper shows, hurts accuracy more than using either alone. DUPLEX therefore treats the pair as a single configuration, re-optimized every round by a deep-reinforcement-learning agent that watches bandwidths, round times, model divergences, and losses. On physical edge-device testbeds over non-IID graphs, it reaches target accuracy 20.1–48.8% faster and with 16.7–37.6% less communication, and achieves 3.3–7.9% higher accuracy under equal resource budgets.

Core claim

The central claim is that network topology and graph sampling ratio are not separable levers in decentralized federated graph learning. In a peer-to-peer system, the amount of node-embedding traffic a worker should generate depends on whom it can talk to; conversely, the value of a communication link depends on how much graph information sampling makes available to send. The paper demonstrates this coupling empirically: a naive composition of topology construction and DRL-based sampling degrades accuracy and leaves communication reduction on the table, whereas DUPLEX, which outputs the adjacency matrix and the sampling-ratio vector together each round, cuts traffic by 22.1% compared with tha

What carries the argument

The coordinated configuration ⟨A,R⟩ is the object that carries the argument: A is a symmetric adjacency matrix selecting which workers exchange models and embeddings, and R assigns each worker a graph sampling ratio. DUPLEX generates both from the same DDPG policy, making the coupling explicit. Consensus distance C(k) — the Euclidean distance between a worker's local model and the global average (estimated from pairwise distances) — is the signal that links non-IID data to the configuration: workers whose models diverge are preferentially connected, and sampling ratios are scaled by consensus distance. The reward function converts three objectives (short round time, consensus distance below

Load-bearing premise

The coordinator's per-round collection of global state and broadcast of new configurations is assumed to cost so little (100–300 KB) that it can be ignored; if it becomes a bottleneck, the system's advantage over a parameter server disappears.

What would settle it

Run DUPLEX at the largest reported scale (500 workers) and log the coordinator's inbound and outbound traffic and the latency between collecting states and broadcasting actions. If this control traffic is not negligible compared with model and embedding exchanges, or if killing the coordinator stops adaptation, the paper's key assumption is falsified. A cleaner test: compare DUPLEX against a version whose coordinator only sees local two-hop information, to see whether the global view earns its cost.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • System builders can treat topology rewiring and sampling as a single control problem rather than two; the empirical comparison shows the naive combination loses to the joint configuration.
  • Adaptive reconfiguration guided by consensus distance reduces the accuracy drop as non-IID degree increases: DUPLEX degrades 11.4% on ogbn-products while baselines degrade 15.2–21.2%.
  • Scaling from 50 to 500 workers grows completion time and network traffic more slowly under DUPLEX than under any baseline tested.
  • Communication cost to reach a target accuracy is the right comparison metric: DUPLEX saves 16–39% depending on dataset, so the joint mechanism is not just an accuracy trick.
  • Under a fixed communication budget, DUPLEX reaches higher accuracy than all baselines, with the largest gains appearing when both topology and sampling are adaptive.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The authors do not pursue, but the same coupling argument likely applies to decentralized federated learning beyond graphs: model-exchange topology and per-worker data/update sampling interact in any peer-to-peer training system, so a counterpart of DUPLEX could be tested on image or text tasks.
  • The reliance on a logical coordinator means the bottleneck moves rather than disappears; a fully decentralized consensus-based extension would trade adaptation speed for fault tolerance, a testable frontier the paper only sketches.
  • Consensus distance as a proxy for distribution skew assumes model-parameter distance correlates with data heterogeneity; a direct comparison on tasks where this correlation fails would delimit the method's applicability.
  • Because the paper measures communication cost to a fixed target accuracy, an even more direct test is to hold total bytes constant and compare final accuracy, which the authors also report for a few budget points.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes DUPLEX, a decentralized federated graph learning (DFGL) framework that jointly optimizes the peer-to-peer network topology and per-worker graph sampling ratios via a DDPG-based coordinator. The stated goal is to reduce communication time and traffic for exchanging node embeddings and model parameters while maintaining or improving accuracy under non-IID graph data and dynamic bandwidth. The authors formulate a TOMAS optimization problem, define a consensus-distance-based reward, present a DRL algorithm, and report experiments on three datasets with up to 50 Jetson devices. They claim 20.1%-48.8% reductions in completion time, 16.7%-37.6% reductions in communication cost, and 3.3%-7.9% accuracy improvements over selected baselines, with additional scalability experiments up to 500 simulated workers. The paper also includes an ablation study and sensitivity analysis.

Significance. If the central claims held, the paper would make a useful contribution: it addresses a real bottleneck in DFGL (cross-worker node-embedding communication) and proposes a plausible mechanism for coupling topology construction and graph sampling. The empirical setup with physical Jetson devices and multiple large graph datasets is a strength, and the ablation study is informative. However, the paper currently contains a load-bearing flaw in the definition of the consensus-distance estimator that feeds the DRL reward, as well as unresolved issues about the coordinator's communication cost and the discrete-action DDPG implementation. These issues mean the mechanism is not yet convincingly validated, and the claimed improvements rest on an experiment protocol with no statistical replication and some inconsistent baseline selection. The contribution is potentially valuable but needs substantial revision before the claims can be accepted.

major comments (4)
  1. [Section 3.2.3, Eq. (15)] The estimator \hat C^(k) in Eq. (15) is not a computable or valid estimator of the global consensus distance defined in Eq. (6). First, \hat C_{i,j} = min_q(C_{i,q}+C_{j,q}) requires pairwise distances C_{i,q} and C_{j,q} for every q, but Algorithm 2 (line 6) only computes distances between current neighbors. Thus the minimization is not well-defined for non-neighbor pairs. Second, even if all pairwise distances were available, the formula is structurally wrong: for a complete topology all (1-a_{i,j}) factors vanish, so \hat C^{(k)}=0 regardless of actual model divergence, whereas the true C^{(k)} is generally positive. Third, by the triangle inequality min_q(C_{i,q}+C_{j,q}) is an upper bound on ||\omega_i-\omega_j||, not an estimate of it. Because the reward in Eq. (12) uses \hat C^{(k)}, the DRL signal conflates topology density with consensus and does not measure what Eq. (6) defines
  2. [Section 3.1 and Section 6] The paper repeatedly contrasts DUPLEX with a parameter server, arguing that the logical coordinator does not become a bottleneck because the state information it collects is much smaller than model parameters. However, the coordinator receives global state from all workers every round and broadcasts the new adjacency matrix and sampling ratios. To compute Eq. (15) as written, it would need pairwise model distances for essentially all worker pairs, not just neighbors, which is O(m^2) communication per round. The assumption that this traffic is negligible (100-300 KB) is not demonstrated for the claimed 500-worker scale, and Section 6 acknowledges that coordinator failure 'could temporarily disrupt adaptation.' The scalability claims in Section 4.6 therefore do not account for the control traffic that DUPLEX itself introduces. This needs either a concrete cost model or a revised architectu
  3. [Section 3.2.4, Eq. (16)] The action space in Section 3.2.2 includes the binary adjacency matrix A^{(k)} \in {0,1}^{m\times m}, while the paper selects DDPG because it 'can efficiently and effectively handle continuous action spaces.' The paper does not describe how the actor network outputs a binary matrix or how gradients are backpropagated through the discrete topology decision. A sigmoid threshold, Gumbel-softmax, or other discretization would need to be specified and its effect on the DDPG update analyzed. As written, the algorithm is not reproducible and the training dynamics of the discrete-continuous hybrid action space are unexplained.
  4. [Section 4.2, Figs. 8-10, Table 4] The empirical evaluation has several issues that affect the strength of the reported improvements. (1) No error bars, standard deviations, or multiple seeds are reported for any accuracy, time, or communication-cost result, although the system involves stochastic DRL, random bandwidth fluctuation, and random non-IID partitioning; the claimed 3.3%-7.9% accuracy improvements could be within run-to-run noise. (2) The baseline selection is post hoc: baselines are chosen because their final accuracy is 'close' to DUPLEX, and at least one inconsistency appears. The text says that on Reddit 'D-FedGraph(sparse)' is selected, but Table 4 reports 83.51% final accuracy for that baseline, below the 85% target used in Figs. 9(b) and 10(b); the figure captions list D-FedGraph(dense). Either the text or the figure labels are wrong, and a baseline that cannot reach the target accuracy cannot be used for
minor comments (6)
  1. [Abstract/Introduction] Typo: 'Prameter Server' should be 'Parameter Server'.
  2. [Section 2.3.3] Language: 'jointly optimization' should be 'joint optimization'; 'challenge' should be 'challenging' in the sentence 'jointly optimization of network topology and graph sampling ratios is crucial yet challenge in DFGL.'
  3. [Fig. 4 caption] The caption contains garbled text ('/s945/s61/s49/s48/s46/s48 /s945/s61/s49/s46/s48 /s945/s61/s48/s46/s49') that appears to be an artifact; the figure labels are unclear and should be redrawn.
  4. [Section 3.4, Eq. (24)] The mixing weight formula in Eq. (24) gives only the off-diagonal weights; to make the update in Eq. (23) correspond to a doubly stochastic matrix, the self-weight P^{(k)}_{i,i} should be defined explicitly. Please clarify how the consensus update is normalized.
  5. [Section 3.5] The privacy argument that aggregation is 'irreversible' is overstated; the paper should cite concrete reconstruction attacks or limit the claim to 'difficult to recover exactly.' This does not affect the main technical claims.
  6. [Section 4.2, text vs. figures] The labels in Figs. 8, 9, and 10 should be checked against the baseline selection described in the text. For example, Fig. 8(b) lists D-FedGraph(dense) whereas Section 4.2 says D-FedGraph(sparse) is selected for Reddit. Please ensure consistency.

Circularity Check

0 steps flagged

No significant circularity: reported gains are measured on a physical testbed; the DRL reward is the optimization objective itself, and the consensus-distance estimator concern is a correctness gap, not a circular reduction.

full rationale

The paper's central claims (20.1–48.8% time reduction, 16.7–37.6% communication reduction, 3.3–7.9% accuracy gain) are evaluated on a 50-Jetson physical platform against S-Glint, TDGE, D-FedPNS, and D-FedGraph with matched final accuracy; they are measurements, not predictions derived from a fitted model. The TOMAS problem (Eq. 11) is solved by a DDPG agent whose reward (Eq. 12) is built directly from the optimization objectives—round time, consensus distance, and training loss. Using the objective as the RL reward is a standard formulation, not a fit-then-predict loop; the hand-set reward weights (χ=2, ϱ=1, φ=10) are hyperparameters, not calibrated to the reported outcomes. The consensus-distance estimator in Eq. (15) is imported from prior work [24] and is indeed suspect: by construction it depends on the adjacency matrix A and is not equivalent to the true C(k) in Eq. (6), and Algorithm 2 only reports neighbor distances, so the min over q in Eq. (15) may be undefined. This is a correctness/validation gap that undermines the claimed guidance mechanism, but it does not make the paper's empirical results circular: the reported improvements are not constructed to equal Eq. (15), and no quantity in the derivation chain is defined in terms of the predicted outcome. Self-citations exist (e.g., [24], [32], [36]) but are not load-bearing for the central comparison; key external anchors are Boyd's mixing weights [42], consensus-distance literature [34][35], and standard DDPG [39]. Overall no identifiable circular step.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central claim depends on the domain modeling assumptions and the heuristic consensus-distance estimator listed above. The reward weights are hand-tuned, not derived. No new physical or conceptual entities are introduced.

free parameters (5)
  • Reward weights χ, ϱ, φ = χ=2, ϱ=1, φ=10 (default; sensitivity tested in Section 4.5)
    Hand-tuned weights in Eq. (12) balancing round time, consensus distance, and training loss; no principled selection method.
  • EMA coefficient β for C_max = not specified
    Coefficient in Eq. (14) for the exponential moving average of gradient norms used as consensus threshold; no value is given in the paper.
  • Target network update coefficient ξ = not specified
    Soft-update coefficient for DDPG target networks in Eq. (21); omitted from the parameter settings.
  • Discount factor γ = not specified
    Discount factor in Eq. (17); never instantiated.
  • Moving average smoothing Υ = not specified
    Smoothing factor in Eq. (13) for round-time moving average; never instantiated.
axioms (5)
  • domain assumption Communication time is modeled as the max over neighbors of (sampling ratio × embedding size / link bandwidth) plus the max over neighbors of (model size / link bandwidth), with all links to a worker sharing bandwidth equally (Eqs. 8-10).
    The TOMAS objective and reward rest on this bandwidth/communication model, which is stated without empirical validation of the equal-sharing assumption.
  • domain assumption Consensus distance, the Euclidean distance between local model parameters, correlates with data distribution skewness and training performance (Section 3.2.1, Fig. 5).
    Used to guide topology construction and sampling and to shape the reward; the correlation is empirically asserted on one setup and may not transfer.
  • ad hoc to paper The global consensus distance can be approximated by the pairwise estimator in Eq. (15) using two-hop path distances between non-neighbor workers.
    No derivation or bound is provided; this estimator feeds the reward function and is specific to DUPLEX.
  • domain assumption A logical coordinator can collect global state and send topology/sampling actions with negligible communication cost (Section 3.1).
    Load-bearing for the claim of avoiding a PS bottleneck; the paper assumes 100-300 KB control traffic is ignorable and acknowledges in Section 6 that the coordinator is a transient failure point.
  • domain assumption Sharing only post-aggregation embeddings from layers 2 through L prevents inference of raw node features (Section 3.5).
    The irreversibility of the aggregation function is asserted without a privacy analysis; not central to performance but part of the privacy claim.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Communication-Efficient Decentralized Federated Graph Learning over Non-IID Data." pith.science (2026). https://pith.science/paper/J5JRG5CU

@misc{pith2026250908409,
  author       = {Pith},
  title        = {Pith review of: Towards Communication-Efficient Decentralized Federated Graph Learning over Non-IID Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J5JRG5CU}},
  note         = {Machine review of arXiv:2509.08409}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Decentralized Federated Graph Learning (DFGL) overcomes potential bottlenecks of the parameter server in FGL by establishing a peer-to-peer (P2P) communication network among workers. However, while extensive cross-worker communication of graph node embeddings is crucial for DFGL training, it introduces substantial communication costs. Most existing works typically construct sparse network topologies or utilize graph neighbor sampling methods to alleviate the communication overhead in DFGL. Intuitively, integrating these methods may offer promise for doubly improving communication efficiency in DFGL. However, our preliminary experiments indicate that directly combining these methods leads to significant training performance degradation if they are jointly optimized. To address this issue, we propose Duplex, a unified framework that jointly optimizes network topology and graph sampling by accounting for their coupled relationship, thereby significantly reducing communication cost while enhancing training performance in DFGL. To overcome practical DFGL challenges, eg, statistical heterogeneity and dynamic network environments, Duplex introduces a learning-driven algorithm to adaptively determine optimal network topologies and graph sampling ratios for workers. Experimental results demonstrate that Duplex reduces completion time by 20.1%--48.8% and communication costs by 16.7%--37.6% to achieve target accuracy, while improving accuracy by 3.3%--7.9% under identical resource budgets compared to baselines.

Figures

Figures reproduced from arXiv: 2509.08409 by Chenxia Tang, Hongli Xu, Jianchun Liu, Liusheng Huang, Qianpiao Ma, Shilong Wang.

Figure 1
Figure 1. Figure 1: Illustration of local subgraphs on workers, where the [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Impact of network topology/graph sampling ratio on accuracy and resource consumption of model training. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Performance comparison of four methods on Reddit. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Harm of Non-IID data on model accuracy. compared to S-Glint or FedSample, significant optimization space remains. By jointly optimizing network topology and graph sam￾pling ratios, DUPLEX reduces network traffic by 22.1% compared to S-Glint+FedSample. These results highlight the importance of joint optimization in leveraging the benefits of both network topology construction and graph sampling in DFGL. 2.4… view at source ↗
Figure 5
Figure 5. Figure 5: Global consensus distance and model accuracy on the [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Illustration of graph sampling and aggregation of a 2-layer GCN on a single worker. The green and blue nodes are in the 0-th [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Illustration of graph sampling in DUPLEX with sampling ratio r1 = 0.5. The black dotted lines indicate external edges among different subgraphs, while the orange and red lines repre￾sent internal and external graph neighbor sampling, respectively. the GC operation on the top layer L, the training loss Bi for the mini-batch Bi is computed using Eqs. (2) and (3) (Line 16). Finally, the local GCN model parame… view at source ↗
Figure 8
Figure 8. Figure 8: Time-to-Accuracy of DUPLEX and the baselines on the three datasets. For a fair comparison, the baselines are selected based on the results in [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Training time of DUPLEX and the selected baselines to reach the target accuracy on the three datasets. and 256 for GraphSage. Besides, we adopt the Adam optimizer [50] to optimize the models, with an initial learning rate of 0.01 and the weight decay of 3e −4 . The number of local updates and the training batch size are fixed to 5 and 64 for Reddit, and 10 and 128 for ogbn-arxiv and ogbn-products, respecti… view at source ↗
Figure 10
Figure 10. Figure 10: Communication cost of DUPLEX and the baselines to reach the target accuracy on the three datasets. TABLE 5: Test accuracy (%) of DUPLEX and the baselines on ogbn-arxiv, Reddit and ogbn-products under the communication resource constraint of 1.5GB, 10GB and 20GB, respectively. Datasets ogbn-arxiv Reddit ogbn-products D-FedGraph 49.32 83.98 61.37 D-FedPNS 49.89 84.32 62.19 Glint 51.53 85.72 62.91 TDGE 50.36… view at source ↗
Figure 11
Figure 11. Figure 11: Impact of non-IID data on model performance. [PITH_FULL_IMAGE:figures/full_fig_p012_11.png] view at source ↗
Figure 13
Figure 13. Figure 13: Training time of different versions of D [PITH_FULL_IMAGE:figures/full_fig_p013_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Communication cost of different versions of D [PITH_FULL_IMAGE:figures/full_fig_p013_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: DUPLEX outperforms baselines with different parameters. Training Time and Communication Cost. As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p014_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Completion time and network traffic on different worker [PITH_FULL_IMAGE:figures/full_fig_p014_16.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

69 extracted references · 55 canonical work pages · 3 internal anchors

  1. [1]

    A review of recommender systems based on knowledge graph embedding,

    J.-C. Zhang, A. M. Zain, K.-Q. Zhou, X. Chen, and R.-M. Zhang, “A review of recommender systems based on knowledge graph embedding,” Expert Systems with Applications, p. 123876, 2024

  2. [2]

    Predicting social anxiety disorder based on communication logs and social network data from a massively multiplayer online game: Using a graph neural network,

    K. Yokotani, M. Takano, N. Abe, and T. A. Kato, “Predicting social anxiety disorder based on communication logs and social network data from a massively multiplayer online game: Using a graph neural network,”Psychiatry and Clinical Neurosciences, 2025

  3. [3]

    Pain-fl: Personalized privacy-preserving incentive for federated learning,

    P. Sun, H. Che, Z. Wang, Y . Wang, T. Wang, L. Wu, and H. Shao, “Pain-fl: Personalized privacy-preserving incentive for federated learning,”IEEE Journal on Selected Areas in Communications, vol. 39, no. 12, pp. 3805– 3820, 2021

  4. [4]

    Optimizing e-commerce recommender systems: A comprehen- sive review of techniques and future directions,

    Z. Wu, “Optimizing e-commerce recommender systems: A comprehen- sive review of techniques and future directions,”Applied and Computa- tional Engineering, vol. 97, pp. 96–101, 2024

  5. [5]

    Semi-supervised classification with graph convolutional networks,

    T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,”arXiv preprint arXiv:1609.02907, 2016

  6. [6]

    Inductive representation learn- ing on large graphs,

    W. Hamilton, Z. Ying, and J. Leskovec, “Inductive representation learn- ing on large graphs,”Advances in neural information processing systems, vol. 30, 2017

  7. [7]

    Phase space graph convolutional net- work for chaotic time series learning,

    W. Ren, N. Jin, and L. OuYang, “Phase space graph convolutional net- work for chaotic time series learning,”IEEE Transactions on Industrial Informatics, 2024

  8. [8]

    Keystrokesniffer: An off-the-shelf smartphone can eavesdrop on your privacy from anywhere,

    J. Huang, J.-X. Bai, X. Zhang, Z. Liu, Y . Feng, J. Liu, X. Sun, M. Dong, and M. Li, “Keystrokesniffer: An off-the-shelf smartphone can eavesdrop on your privacy from anywhere,”IEEE Transactions on Information Forensics and Security, 2024

  9. [9]

    Adaptive local update and neural composition for accelerating federated learning in heterogeneous edge networks,

    J. Liu, J. Yan, J. Qi, H. Xu, S. Wang, C. Qiao, and L. Huang, “Adaptive local update and neural composition for accelerating federated learning in heterogeneous edge networks,”IEEE Transactions on Networking, 2025

  10. [10]

    Federated graph classification over non-iid graphs,

    H. Xie, J. Ma, L. Xiong, and C. Yang, “Federated graph classification over non-iid graphs,”Advances in neural information processing systems, vol. 34, pp. 18 839–18 852, 2021

  11. [11]

    Subgraph federated learning with missing neighbor generation,

    K. Zhang, C. Yang, X. Li, L. Sun, and S. M. Yiu, “Subgraph federated learning with missing neighbor generation,”Advances in Neural Infor- mation Processing Systems, vol. 34, pp. 6671–6682, 2021

  12. [12]

    Finch: Enhancing federated learning with hierarchical neural architecture search,

    J. Liu, J. Yan, H. Xu, Z. Wang, J. Huang, and Y . Xu, “Finch: Enhancing federated learning with hierarchical neural architecture search,”IEEE Transactions on Mobile Computing, vol. 23, no. 5, pp. 6012–6026, 2023

  13. [13]

    Enhancing semi-supervised federated learning with progressive training in heteroge- neous edge computing,

    J. Liu, J. Liu, H. Xu, Y . Liao, Z. Yao, M. Chen, and C. Qian, “Enhancing semi-supervised federated learning with progressive training in heteroge- neous edge computing,”IEEE Transactions on Mobile Computing, 2024

  14. [14]

    Decen- tralized federated graph neural networks,

    Y . Pei, R. Mao, Y . Liu, C. Chen, S. Xu, F. Qiang, and B. E. Tech, “Decen- tralized federated graph neural networks,” inInternational Workshop on Federated and Transfer Learning for Data Sparsity and Confidentiality in Conjunction with IJCAI, 2021

  15. [15]

    Glint: Decentralized federated graph learning with traffic throttling and flow scheduling,

    T. Liu, P. Li, and Y . Gu, “Glint: Decentralized federated graph learning with traffic throttling and flow scheduling,” in2021 IEEE/ACM 29th International Symposium on Quality of Service (IWQOS). IEEE, 2021, pp. 1–10

  16. [16]

    Acceler- ating decentralized federated learning with probabilistic communication in heterogeneous edge computing,

    J. Liu, J. Yan, H. Xu, L. Wang, Z. Wang, J. Huang, and C. Qiao, “Acceler- ating decentralized federated learning with probabilistic communication in heterogeneous edge computing,”IEEE Transactions on Networking, 2025

  17. [17]

    S-glint: Secure federated graph learning with traffic throttling and flow scheduling,

    T. Liu, P. Li, Y . Gu, and Z. Su, “S-glint: Secure federated graph learning with traffic throttling and flow scheduling,”IEEE Transactions on Green Communications and Networking, 2022

  18. [18]

    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,”Advances in neural information processing systems, vol. 33, pp. 22 118–22 133, 2020

  19. [19]

    Accelerating federated learning via parameter selection and pre-synchronization in mobile edge- cloud networks,

    H. Zhou, M. Li, P. Sun, B. Guo, and Z. Yu, “Accelerating federated learning via parameter selection and pre-synchronization in mobile edge- cloud networks,”IEEE Transactions on Mobile Computing, vol. 23, no. 11, pp. 10 313–10 328, 2024

  20. [20]

    Accurate, large minibatch sgd: Training imagenet in 1 hour,

    P. Goyal, P. Dollár, R. Girshick, P. Noordhuis, L. Wesolowski, A. Kyrola, A. Tulloch, Y . Jia, and K. He, “Accurate, large minibatch sgd: Training imagenet in 1 hour,”arXiv preprint arXiv:1706.02677, 2017

  21. [21]

    Fedgraph: Federated graph learning with intelligent sampling,

    F. Chen, P. Li, T. Miyazaki, and C. Wu, “Fedgraph: Federated graph learning with intelligent sampling,”IEEE Transactions on Parallel and Distributed Systems, vol. 33, no. 8, pp. 1775–1786, 2021

  22. [22]

    Federated graph learning with periodic neighbour sampling,

    B. Du and C. Wu, “Federated graph learning with periodic neighbour sampling,” in2022 IEEE/ACM 30th International Symposium on Quality of Service (IWQoS). IEEE, 2022, pp. 1–10

  23. [23]

    Historical embedding-guided efficient large-scale federated graph learning,

    A. Li, Y . Chen, J. Zhang, M. Cheng, Y . Huang, Y . Wu, A. T. Luu, and H. Yu, “Historical embedding-guided efficient large-scale federated graph learning,”Proceedings of the ACM on Management of Data, vol. 2, no. 3, pp. 1–24, 2024

  24. [24]

    Accelerating de- centralized federated learning in heterogeneous edge computing,

    L. Wang, Y . Xu, H. Xu, M. Chen, and L. Huang, “Accelerating de- centralized federated learning in heterogeneous edge computing,”IEEE Transactions on Mobile Computing, 2022

  25. [25]

    Enhancing federated learning with intelligent model migration in heterogeneous edge computing,

    J. Liu, Y . Xu, H. Xu, Y . Liao, Z. Wang, and H. Huang, “Enhancing federated learning with intelligent model migration in heterogeneous edge computing,” in2022 IEEE 38th International Conference on Data Engineering (ICDE). IEEE, 2022, pp. 1586–1597

  26. [26]

    Mobility- aware fog computing in dynamic environments: Understandings and implementation,

    M. Waqas, Y . Niu, M. Ahmed, Y . Li, D. Jin, and Z. Han, “Mobility- aware fog computing in dynamic environments: Understandings and implementation,”IEEE Access, vol. 7, pp. 38 867–38 879, 2018

  27. [27]

    Matcha: Speeding up decentralized sgd via matching decomposition sampling,

    J. Wang, A. K. Sahu, Z. Yang, G. Joshi, and S. Kar, “Matcha: Speeding up decentralized sgd via matching decomposition sampling,” in2019 Sixth Indian Control Conference (ICC). IEEE, 2019, pp. 299–300

  28. [28]

    Mergesfl: Split federated learning with feature merging and batch size regulation,

    Y . Liao, Y . Xu, H. Xu, L. Wang, Z. Yao, and C. Qiao, “Mergesfl: Split federated learning with feature merging and batch size regulation,” in 2024 IEEE 40th International Conference on Data Engineering (ICDE). IEEE, 2024, pp. 2054–2067

  29. [29]

    Parallelsfl: A novel split federated learning framework tackling heterogeneity issues,

    Y . Liao, Y . Xu, H. Xu, Z. Yao, L. Huang, and C. Qiao, “Parallelsfl: A novel split federated learning framework tackling heterogeneity issues,” IEEE TRANSACTIONS ON XXX, VOL., NO., NOV. 2024 17 inProceedings of the 30th Annual International Conference on Mobile Computing and Networking, 2024, pp. 845–860

  30. [30]

    Fedgraphnn: A federated learn- ing system and benchmark for graph neural networks,

    C. He, K. Balasubramanian, E. Ceyani, C. Yang, H. Xie, L. Sun, L. He, L. Yang, P. S. Yu, Y . Ronget al., “Fedgraphnn: A federated learn- ing system and benchmark for graph neural networks,”arXiv preprint arXiv:2104.07145, 2021

  31. [31]

    Communication-efficient decentralized machine learning over heteroge- neous networks,

    P. Zhou, Q. Lin, D. Loghin, B. C. Ooi, Y . Wu, and H. Yu, “Communication-efficient decentralized machine learning over heteroge- neous networks,” in2021 IEEE 37th International Conference on Data Engineering (ICDE). IEEE, 2021, pp. 384–395

  32. [32]

    Decentralized machine learning through experience-driven method in edge networks,

    H. Xu, M. Chen, Z. Meng, Y . Xu, L. Wang, and C. Qiao, “Decentralized machine learning through experience-driven method in edge networks,” IEEE Journal on Selected Areas in Communications, vol. 40, no. 2, pp. 515–531, 2021

  33. [33]

    Optimal online data partitioning for geo-distributed machine learning in edge of wireless networks,

    X. Lyu, C. Ren, W. Ni, H. Tian, R. P. Liu, and E. Dutkiewicz, “Optimal online data partitioning for geo-distributed machine learning in edge of wireless networks,”IEEE Journal on Selected Areas in Communications, vol. 37, no. 10, pp. 2393–2406, 2019

  34. [34]

    Consensus control for decentralized deep learning,

    L. Kong, T. Lin, A. Koloskova, M. Jaggi, and S. Stich, “Consensus control for decentralized deep learning,” inInternational Conference on Machine Learning. PMLR, 2021, pp. 5686–5696

  35. [35]

    Decentralized deep learning with arbitrary communication compression,

    A. Koloskova, T. Lin, S. U. Stich, and M. Jaggi, “Decentralized deep learning with arbitrary communication compression,” inProceedings of the 8th International Conference on Learning Representations, 2019

  36. [36]

    Adaptive configuration for heterogeneous participants in decentralized federated learning,

    Y . Liao, Y . Xu, H. Xu, L. Wang, and C. Qian, “Adaptive configuration for heterogeneous participants in decentralized federated learning,” in IEEE INFOCOM 2023-IEEE Conference on Computer Communications. IEEE, 2023, pp. 1–10

  37. [37]

    The non-iid data quagmire of decentralized machine learning,

    K. Hsieh, A. Phanishayee, O. Mutlu, and P. Gibbons, “The non-iid data quagmire of decentralized machine learning,” inInternational Confer- ence on Machine Learning. PMLR, 2020, pp. 4387–4398

  38. [38]

    The complexity of facets (and some facets of complexity),

    C. H. Papadimitriou and M. Yannakakis, “The complexity of facets (and some facets of complexity),” inProceedings of the fourteenth annual ACM symposium on Theory of computing, 1982, pp. 255–260

  39. [39]

    Continuous control with deep reinforcement learning,

    T. P. Lillicrap, J. J. Hunt, A. Pritzel, N. Heess, T. Erez, Y . Tassa, D. Silver, and D. Wierstra, “Continuous control with deep reinforcement learning,” arXiv preprint arXiv:1509.02971, 2015

  40. [40]

    Experience-driven networking: A deep reinforcement learning based approach,

    Z. Xu, J. Tang, J. Meng, W. Zhang, Y . Wang, C. H. Liu, and D. Yang, “Experience-driven networking: A deep reinforcement learning based approach,” inIEEE INFOCOM 2018-IEEE conference on computer communications. IEEE, 2018, pp. 1871–1879

  41. [41]

    Human-level control through deep reinforcement learning,

    V . Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovskiet al., “Human-level control through deep reinforcement learning,”nature, vol. 518, no. 7540, pp. 529–533, 2015

  42. [42]

    Fast linear iterations for distributed averaging,

    L. Xiao and S. Boyd, “Fast linear iterations for distributed averaging,” Systems & Control Letters, vol. 53, no. 1, pp. 65–78, 2004

  43. [43]

    Nvidia jetson devices

    “Nvidia jetson devices.” https://docs.nvidia.com/jetson/

  44. [44]

    Docker: lightweight linux containers for consistent development and deployment,

    D. Merkelet al., “Docker: lightweight linux containers for consistent development and deployment,”Linux j, vol. 239, no. 2, p. 2, 2014

  45. [45]

    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,”Advances in neural information processing systems, vol. 32, 2019

  46. [46]

    socket library

    “socket library.” https://github.com/python/cpython/blob/3.10/Lib/ socket.py

  47. [47]

    Semi-supervised decentralized machine learning with device-to-device cooperation,

    Z. Jiang, Y . Xu, H. Xu, Z. Wang, J. Liu, and C. Qiao, “Semi-supervised decentralized machine learning with device-to-device cooperation,”IEEE Transactions on Mobile Computing, 2024

  48. [48]

    Yoga: Adaptive layer- wise model aggregation for decentralized federated learning,

    J. Liu, J. Liu, H. Xu, Y . Liao, Z. Wang, and Q. Ma, “Yoga: Adaptive layer- wise model aggregation for decentralized federated learning,”IEEE/ACM Transactions on Networking, 2023

  49. [49]

    Topology design and graph embedding for decentralized federated learning,

    Y . Duan, X. Li, and J. Wu, “Topology design and graph embedding for decentralized federated learning,”Intelligent and Converged Networks, vol. 5, no. 2, pp. 100–115, 2024

  50. [50]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014

  51. [51]

    Graph neural networks,

    G. Corso, H. Stark, S. Jegelka, T. Jaakkola, and R. Barzilay, “Graph neural networks,”Nature Reviews Methods Primers, vol. 4, no. 1, p. 17, 2024

  52. [52]

    A survey of graph neural networks for social recommender systems,

    K. Sharma, Y .-C. Lee, S. Nambi, A. Salian, S. Shah, S.-W. Kim, and S. Kumar, “A survey of graph neural networks for social recommender systems,”ACM Computing Surveys, vol. 56, no. 10, pp. 1–34, 2024

  53. [53]

    A review of convolutional neural networks in computer vision,

    X. Zhao, L. Wang, Y . Zhang, X. Han, M. Deveci, and M. Parmar, “A review of convolutional neural networks in computer vision,”Artificial Intelligence Review, vol. 57, no. 4, p. 99, 2024

  54. [54]

    Federated graph neural networks: Overview, techniques, and challenges,

    R. Liu, P. Xing, Z. Deng, A. Li, C. Guan, and H. Yu, “Federated graph neural networks: Overview, techniques, and challenges,”IEEE Transactions on Neural Networks and Learning Systems, 2024

  55. [55]

    Federated Graph Semantic and Structural Learning

    W. Huang, G. Wan, M. Ye, and B. Du, “Federated graph semantic and structural learning,”arXiv preprint arXiv:2406.18937, 2024

  56. [56]

    Hierarchical federated learning with local model embedding,

    Y . He, D. Yan, and F. Chen, “Hierarchical federated learning with local model embedding,”Engineering Applications of Artificial Intelligence, vol. 123, p. 106148, 2023

  57. [57]

    Fedgl: Federated graph learning framework with global self-supervision,

    C. Chen, Z. Xu, W. Hu, Z. Zheng, and J. Zhang, “Fedgl: Federated graph learning framework with global self-supervision,”Information Sciences, vol. 657, p. 119976, 2024

  58. [58]

    Federated graph learning under domain shift with generalizable prototypes,

    G. Wan, W. Huang, and M. Ye, “Federated graph learning under domain shift with generalizable prototypes,” inProceedings of the AAAI Confer- ence on Artificial Intelligence, vol. 38, no. 14, 2024, pp. 15 429–15 437

  59. [59]

    Federated learning over coupled graphs,

    R. Lei, P. Wang, J. Zhao, L. Lan, J. Tao, C. Deng, J. Feng, X. Wang, and X. Guan, “Federated learning over coupled graphs,”IEEE Transactions on Parallel and Distributed Systems, vol. 34, no. 4, pp. 1159–1172, 2023

  60. [60]

    Federated hetero- geneous contrastive distillation for molecular representation learning,

    J. Feng, Z. Wang, Z. Wei, Y . Li, B. Ding, and H. Xu, “Federated hetero- geneous contrastive distillation for molecular representation learning,” in Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, 2024, pp. 1038–1048

  61. [61]

    Decentralized federated learning: Fundamentals, state of the art, frameworks, trends, and chal- lenges,

    E. T. M. Beltrán, M. Q. Pérez, P. M. S. Sánchez, S. L. Bernal, G. Bovet, M. G. Pérez, G. M. Pérez, and A. H. Celdrán, “Decentralized federated learning: Fundamentals, state of the art, frameworks, trends, and chal- lenges,”IEEE Communications Surveys & Tutorials, 2023

  62. [62]

    Spreadgnn: Decentralized multi-task federated learning for graph neural networks on molecular data,

    C. He, E. Ceyani, K. Balasubramanian, M. Annavaram, and S. Aves- timehr, “Spreadgnn: Decentralized multi-task federated learning for graph neural networks on molecular data,” inProceedings of the AAAI conference on artificial intelligence, vol. 36, no. 6, 2022, pp. 6865–6873

  63. [63]

    Dci-pfgl: Decentralized cross-institutional personalized federated graph learning for iot service recommendation,

    B. Xie, C. Hu, H. Huang, J. Yu, and H. Xia, “Dci-pfgl: Decentralized cross-institutional personalized federated graph learning for iot service recommendation,”IEEE Internet of Things Journal, 2023

  64. [64]

    Federated Graph Learning with Adaptive Importance-based Sampling

    A. Li, Y . Chen, C. Ren, W. Wang, M. Hu, T. Li, H. Yu, and Q. Chen, “Federated graph learning with adaptive importance-based sampling,” arXiv preprint arXiv:2409.14655, 2024

  65. [65]

    p2pgnn: A de- centralized graph neural network for node classification in peer-to-peer networks,

    E. Krasanakis, S. Papadopoulos, and I. Kompatsiaris, “p2pgnn: A de- centralized graph neural network for node classification in peer-to-peer networks,”IEEE Access, vol. 10, pp. 34 755–34 765, 2022

  66. [66]

    Towards personalized federated learning,

    A. Z. Tan, H. Yu, L. Cui, and Q. Yang, “Towards personalized federated learning,”IEEE transactions on neural networks and learning systems, vol. 34, no. 12, pp. 9587–9603, 2022

  67. [67]

    Improving Federated Relational Data Modeling via Basis Alignment and Weight Penalty

    Y . Lin, C. Chen, C. Chen, and L. Wang, “Improving federated relational data modeling via basis alignment and weight penalty,”arXiv preprint arXiv:2011.11369, 2020

  68. [68]

    An information theoretic perspective for heterogeneous subgraph federated learning,

    J. Guo, S. Li, and Y . Zhang, “An information theoretic perspective for heterogeneous subgraph federated learning,” inInternational Conference on Database Systems for Advanced Applications. Springer, 2023, pp. 745–760

  69. [69]

    How powerful are graph neural networks?

    K. Xu, W. Hu, J. Leskovec, and S. Jegelka, “How powerful are graph neural networks?”arXiv preprint arXiv:1810.00826, 2018

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.