Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Graph-based Gossiping for Communication Efficiency in Decentralized Federated Learning

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Replacing flooding broadcast with a minimum-spanning-tree gossip schedule cuts decentralized federated learning bandwidth by up to about 8x and communication time by up to about 4.4x, on real routers and across several network topologies.

desk verdict The physical testbed and scheduling idea are real, but the headline “8x bandwidth reduction” does not survive contact with the data: the method saves wall-clock time through reduced contention, not bytes. read the letter →

arxiv 2506.10607 v1 pith:OKFUJKTD submitted 2025-06-12 cs.DC

Huong Nguyen , Hong-Tri Nguyen , Praveen Kumar Donta , Susanna Pirttikangas , Lauri Lovén This is my paper
classification cs.DC
keywords FederatedLearningDecentralizedGossipingGraphColoringMinimumSpanningTreeCommunicationEfficiencyNetworkTopologyLatency
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

Decentralized federated learning lets nodes exchange model updates directly instead of through a central server, but naive broadcasting floods the network and slows training. This paper proposes a gossip schedule built from a minimum spanning tree and a two-color graph coloring: nodes of one color transmit while the other color waits, so adjacent nodes never send at the same time. On a testbed of ten physical devices behind three routers, the method reports up to about 8x lower bandwidth use and about 4.4x faster communication rounds than flooding broadcast, with the largest gains on the biggest models. The authors conclude that communication cost in decentralized learning can be cut substantially without changing the learning algorithm, only by scheduling who talks when.

What carries the argument

The load-bearing machinery is a minimum spanning tree plus a two-coloring. The MST (the cheapest set of links that still connects every node) removes redundant edges from the dense overlay, and because every tree is bipartite, a two-coloring guarantees that communicating nodes are never adjacent. The moderator schedules each color in a fixed time slot, with slot length computed as $\text{slot} = \frac{\text{ping}_{\max} \times M_{\text{size}} \times 1000}{\text{ping}_{\text{size}}}$ seconds, so the schedule scales with both the measured worst-case ping and the model size. Each node keeps a FIFO queue of model updates and forwards its oldest entry when its color's slot arrives; leaf nodes, having degree 1, clear their queue immediately after sending their own model. This scheduling discipline is what converts a congested all-to-all broadcast into staged, collision-free gossip.

What would settle it

Run a single transfer of the 48 MB EfficientNet-B3 model between two adjacent nodes on the three-router testbed, measure the real FTP transfer time, and compare it with the slot length computed from the paper's formula using the measured ping latency; if the real transfer takes longer than the slot, the schedule's core assumption is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that communication in decentralized federated learning can be made cheaper by replacing all-to-all flooding with a two-phase gossip schedule on a minimum spanning tree. The moderator first measures connection costs (ping latencies), builds an MST using a standard greedy algorithm, colors the tree with two colors via breadth-first traversal, and assigns each color a slot whose length is proportional to the largest ping among that color's nodes scaled by model size. Nodes sharing a color transmit in parallel; nodes of the other color stay silent, so no adjacent nodes transmit simultaneously. In experiments with seven MobileNet and EfficientNet variants on a ten-node, three-router physical setup, the method raises bandwidth from as low as 0.767 MB/s to over 6 MB/s in the best case and reduces communication-round time by up to about 4.4x compared with flooding. The gain is consistent across complete, random, small-world, and scale-free topologies and grows with model size.

Load-bearing premise

The schedule assumes that sending a model takes exactly the ping latency scaled by model size, computed as $\text{slot} = \frac{\text{ping}_{\max} \times M_{\text{size}} \times 1000}{\text{ping}_{\text{size}}}$ seconds, so if actual FTP transfer times are longer than this linear estimate, transmissions will overrun their slots and the collision-free schedule breaks.

Editorial extensions

If this is right

  • Communication savings layer onto any decentralized learning algorithm, because the method changes only the message schedule, not the aggregation or training rule.
  • Larger models benefit the most: the bandwidth multiplier rises from about 2.4x for the smallest MobileNet to about 8x for the largest EfficientNet.
  • The method works across all four tested topologies, including complete, random, small-world, and scale-free networks, so the gains are not tied to a particular network structure.
  • Because the moderator rotates each round and the graph is only recomputed when nodes join or leave, the extra control overhead is kept out of the steady-state communication cost.
  • The approach is compatible with real physical routers and FTP-based transfer, suggesting it can be applied in cross-silo settings without specialized hardware.

Reading between the lines

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

  • The slot-length formula assumes linear scaling between ping delay and file transfer time; we infer this is the least validated piece of the design, since FTP throughput on real links is often sub-linear in payload size, so a model transfer could overrun its slot and reintroduce collisions.
  • Because the MST has only two colors, at any moment roughly half the nodes are silent; we infer that on dense, well-connected networks the schedule may leave bandwidth on the table, and a degree-aware or k-coloring schedule could improve utilization further.
  • The paper measures communication only and cites prior work for accuracy equivalence; we infer that the true test of the method is a full training run where the gossip schedule changes the order in which models mix, which could affect convergence dynamics in ways not captured by transfer-time benchmarks.
  • The moderator, though rotated, still holds a global map of the network each round; we infer that in adversarial or highly dynamic settings, moderator selection and the cost of re-synchronization could become the bottleneck, an issue the paper does not address.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes a gossip-based communication scheme for decentralized federated learning: a moderator builds a minimum spanning tree of the overlay network, two-colors it with BFS, and schedules transmissions in alternating color slots whose length is derived from ping latency and model size. Experiments on a 10-node physical testbed using FTP and seven MobileNet/EfficientNet variants report bandwidth, per-transfer time, and round-completion time across four topologies. The authors claim up to roughly 8x bandwidth reduction and 4.4x transfer-time reduction versus naive flooding broadcast.

Significance. If the communication-time gains are confirmed, the paper contributes a simple, practical scheduling mechanism and one of the few physical testbed evaluations in this area; the use of seven real model sizes and four topologies is a strength. However, the headline metric conflates throughput with aggregate bandwidth, the key ratios compare different topologies, and the slot-length model is unvalidated, so the significance as stated is not currently established. No code or data release is mentioned, which limits reproducibility.

major comments (3)
  1. [§V, Tables III–V] The abstract's claim of 'reducing bandwidth and transfer time by up to circa 8 and 4.4 times' is not supported by the reported metrics. Table III's 'Bandwidth (MB/s)' is model size divided by the average single-transfer time in Table IV (e.g., ER/b3: 48/62.576 ≈ 0.767 and 48/7.971 ≈ 6.022), so it measures throughput, not bandwidth consumption. In both flooding on the complete graph and gossiping on the MST, each of the N models traverses each of the N−1 edges exactly once per round, so the aggregate number of bytes transmitted is N(N−1)M in both schemes; the proposed method does not reduce total bytes transferred. Moreover, the 'up to 8x' and 'up to 4.4x' ratios pair the broadcast values for the Erdős–Rényi topology with the proposed values for Watts–Strogatz (e.g., 0.767 MB/s vs. 6.146 MB/s for b3, and 55 s vs. 12.571 s for b1), rather than comparing the same topology. The manuscript should either re-state the claims in terms of per-transfer time and round-completion time with same-topology baselines, or provide a genuine byte-accounting comparison if a bandwidth-reduction claim is intended.
  2. [§V, Tables III–V] The broadcast baseline is reported only for the Erdős–Rényi row; the Watts–Strogatz, Barabási–Albert, and complete rows contain only proposed-method values. Because topology is one of the paper's main experimental variables, this missing baseline makes the claim that the method is 'compatible with different topologies' non-comparable. The tables also state that all figures are averaged but do not report the number of runs, standard deviations, or any error/confidence measure; with N=10 and FTP over real routers, run-to-run variance is likely non-negligible. Please provide full baseline tables for every topology and include repeated-run statistics.
  3. [§III-C, Eq. (1)] The slot-length model slot = ping_max × M_size × 1000 / ping_size assumes that an FTP transfer of a model of size M_size completes in exactly the same scaled time as a ping exchange, with transfer time scaling linearly in payload size. This is not validated anywhere in the paper; FTP throughput is governed by TCP congestion control, connection setup, and router buffering, and real transfer times can deviate substantially from ping-latency scaling. If the computed slot is shorter than the actual transfer time, transmissions from one color overrun into the other color's slot, breaking the two-color schedule and invalidating the round-time measurements in Table V. The authors should either calibrate slot length against measured FTP transfer times on the testbed or derive it from per-link throughput measurements.
minor comments (5)
  1. [§III-C] The paper correctly notes that any proper coloring of an MST uses exactly two colors, so the comparison of BFS, DSatur, LDF, and Welsh-Powell is moot for this application; a sentence stating that a simple bipartition suffices would avoid the misleading algorithm-selection discussion.
  2. [§V-A] The statement that the total time reduces from 10 s to 3.16 s is a 68.4% reduction, not 'nearly 70%', and the phrase '3.16 times' should be phrased as a reduction factor rather than a time ratio.
  3. [Tables III–V] The table formatting makes it difficult to determine which entries are broadcast and which are proposed values, especially since only one broadcast row is populated; separate labeled sub-tables or explicit column headers per topology would improve readability.
  4. [§I and §V] The term 'bandwidth' is used inconsistently: the abstract says the method 'reduce[s] bandwidth', but Table III labels higher MB/s values as improvements. Please rename the metric to 'throughput' and phrase the contribution in terms of transfer time and round-completion time.
  5. [§III-A] The moderator, although periodically rotated, holds the full adjacency matrix and computes the MST and coloring centrally; the paper should clarify that this graph-computation time is excluded from the reported round times (as stated in §V) and discuss the cost of moderator rotation and failure recovery.

Circularity Check

2 steps flagged · score 4.0 of 10

Headline '8x bandwidth reduction' is a definitional restatement of the per-transfer time speedup, not a reduction in bytes; accuracy support relies on a self-citation.

  1. self definitional [Abstract; Section V indicator definitions; Table III vs Table IV (ER-b3 row)]
    "Experimental results demonstrate that our method significantly improves communication, compatible with different topologies and data sizes, reducing bandwidth and transfer time by up to circa 8 and 4.4 times, respectively, compared to naive flooding broadcasting methods. ... the report on bandwidth (MB/s), details the network’s data transmission capabilities, specifically the rate (in megabytes) at which data can be transmitted or received over the network within a specific amount of time (one second)."

    Bandwidth is defined as MB/s, i.e. model capacity divided by per-transfer time. The reported ratios are therefore not independent: for ER-b3, Table III gives broadcast 0.767 MB/s and proposed 6.022 MB/s, while Table IV gives 62.576 s and 7.971 s; 48/62.576 = 0.767 and 48/7.971 = 6.022, so the 7.85x 'bandwidth reduction' is exactly the inverse of the 7.85x per-transfer speedup in the same cell. Both flooding and tree gossip transmit N(N-1) model copies per round (complete overlay: N nodes x (N-1) peers; tree gossip: N models x (N-1) edges), so there is no aggregate byte reduction. The headline thus reports the same wall-clock speedup twice, once as 'bandwidth' and once as 'transfer time', with the direction of 'reducing bandwidth' inverted (throughput increases).

  2. self citation load bearing [Section V, opening paragraph (accuracy support)]
    "Besides, this work does not include the evaluation of training performance, relying instead on findings from prior studies [4], [16] and a practical report by Nguyen et al. [15]. These sources demonstrated that DFL can maintain comparable accuracy levels to CFL in broadcast or gossip mode."

    The only support for the accuracy-preservation premise is prior work, one item of which ([15], Nguyen et al.) shares authors with this paper (H. Nguyen, L. Lovén, S. Pirttikangas). Since the paper explicitly omits accuracy experiments, the claim that the proposed communication scheme does not hurt learning performance rests on this self-citation. It is not load-bearing for the measured communication-efficiency results, but it is a self-referential dependency in the paper's overall narrative.

full rationale

The MST-plus-two-coloring mechanism is not circular: it constructively builds a tree, colors it with BFS, schedules slots, and measures wall-clock round time against an external flooding baseline. Those total-time numbers (Table V) are independent empirical evidence and are not fitted to the claimed outcome. The slot-length formula is an untested linear-scaling assumption, but it is not fitted to the reported improvements, so it is a validation gap rather than circularity. The one genuine definitional reduction is the 'bandwidth (MB/s)' metric: because it is computed as model size divided by per-transfer time, the 8x headline is the inverse of the measured per-transfer speedup, not a separate reduction in bytes consumed; aggregate byte count is identical for flooding and tree gossip. A minor self-citation ([15], same authorship group) supports the omitted accuracy evaluation, but is not load-bearing for the communication results. Overall, the central empirical claim survives with independent content; the score of 4 reflects the headline metric conflation and the self-referential accuracy support.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the validity of the slot-length scaling, the correctness of the gossip schedule, and the transferability of prior accuracy results. No new physical entities are introduced.

free parameters (1)
  • Color slot length = computed from ping_max, M_size, ping_size
    The duration of each color slot is set by an unvalidated linear scaling of ping time. It determines the total round time, and the reported gains are sensitive to it.
assumptions (4)
  • domain assumption The moderator can collect accurate connection information (ping latency, IP) for all nodes at the start of each round.
    Section III-A assumes every node reports its connection costs to the moderator; if costs are stale or asymmetric, the MST and coloring are suboptimal.
  • ad hoc to paper A model of size M can be transferred over a link in exactly slot = ping_max x M_size x 1000 / ping_size seconds.
    Section III-C introduces this linear scaling without empirical support; the schedule relies on it to avoid collisions and idle time.
  • domain assumption The FIFO gossip schedule on a 2-colored MST delivers every model to every node in a finite number of slots.
    Section III-D demonstrates this by example for N=10 (Table I) but provides no proof; deadlocks or starvation would invalidate the results.
  • domain assumption Prior findings [4], [15], [16] establish that DFL with gossip maintains accuracy comparable to CFL, so training evaluation can be omitted.
    Section V states this reliance; reference [15] is co-authored by the current authors, so the support is partly self-citation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Graph-based Gossiping for Communication Efficiency in Decentralized Federated Learning." pith.science (2026). https://pith.science/paper/OKFUJKTD

@misc{pith2026250610607,
  author       = {Pith},
  title        = {Pith review of: Graph-based Gossiping for Communication Efficiency in Decentralized Federated Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OKFUJKTD}},
  note         = {Machine review of arXiv:2506.10607}
}
read the original abstract

Federated learning has emerged as a privacy-preserving technique for collaborative model training across heterogeneously distributed silos. Yet, its reliance on a single central server introduces potential bottlenecks and risks of single-point failure. Decentralizing the server, often referred to as decentralized learning, addresses this problem by distributing the server role across nodes within the network. One drawback regarding this pure decentralization is it introduces communication inefficiencies, which arise from increased message exchanges in large-scale setups. However, existing proposed solutions often fail to simulate the real-world distributed and decentralized environment in their experiments, leading to unreliable performance evaluations and limited applicability in practice. Recognizing the lack from prior works, this work investigates the correlation between model size and network latency, a critical factor in optimizing decentralized learning communication. We propose a graph-based gossiping mechanism, where specifically, minimum spanning tree and graph coloring are used to optimize network structure and scheduling for efficient communication across various network topologies and message capacities. Our approach configures and manages subnetworks on real physical routers and devices and closely models real-world distributed setups. Experimental results demonstrate that our method significantly improves communication, compatible with different topologies and data sizes, reducing bandwidth and transfer time by up to circa 8 and 4.4 times, respectively, compared to naive flooding broadcasting methods.

Figures

Figures reproduced from arXiv: 2506.10607 by the authors.

Figure 1
Figure 1. Adjacency matrix [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Redundancy removal and coloring vertices on an undirected graph [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Experimental setting computational configurations align with or are even below the hardware specifications commonly found in contemporary mobile devices. Besides, we utilized File Transfer Protocol (FTP)1 for transmitting information and models. To emulate varying geographical distances and inter-node 1https://www.rfc-editor.org/info/rfc959 communication scenarios, we used three routers, each config￾ured to govern a… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Graph experiments (black edge is interconnection, dashed-blue is local connections) [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Constructed MST (black edge is interconnection, [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Colored MST (black edge is interconnection, [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

Discussion (0). Sign in 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. A Comprehensive Data-centric Overview of Federated Graph Learning

    cs.LG 2025-07 conditional novelty 5.0 of 10

    A data-centric taxonomy for Federated Graph Learning that classifies 79 studies by data characteristics and data utilization, plus a discussion of integration with pre-trained large models.

Reference graph

Works this paper leans on

32 extracted references · 21 canonical work pages · cited by 1 Pith paper

  1. [1]

    Decentralized directed collaboration for personalized federated learning,

    Y . Liu, Y . Shi, Q. Li, B. Wu, X. Wang, and L. Shen, “Decentralized directed collaboration for personalized federated learning,” inProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2024, pp. 23 168–23 178

    work page 2024

  2. [2]

    Communication-efficient learning of deep networks from decentralized data,

    B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication-efficient learning of deep networks from decentralized data,” inProceedings of the 20th International Conference on Artificial Intelligence and Statistics, AISTATS 2017, 20-22 April 2017, Fort Lauderdale, FL, USA, ser. Proceedings of Machine Learning Research, A. Singh and X. J...

    work page 2017

  3. [3]

    Federated learning: Challenges, methods, and future directions,

    T. Li, A. K. Sahu, A. Talwalkar, and V . Smith, “Federated learning: Challenges, methods, and future directions,”IEEE signal processing magazine, vol. 37, no. 3, pp. 50–60, 2020

    2020

  4. [4]

    Fedstellar: A platform for decentralized federated learning,

    E. T. M. Beltr ´an, ´A. L. P. G ´omez, C. Feng, P. M. S. S ´anchez, S. L. Bernal, G. Bovet, M. G. P ´erez, G. M. P ´erez, and A. H. Celdr ´an, “Fedstellar: A platform for decentralized federated learning,”Expert Systems with Applications, vol. 242, p. 122861, 2024

    work page 2024

  5. [5]

    Advances and open problems in federated learning,

    P. Kairouz, H. B. McMahan, B. Avent, A. Bellet, M. Bennis, A. N. Bhagoji, K. Bonawitz, Z. Charles, G. Cormode, R. Cummingset al., “Advances and open problems in federated learning,”Foundations and Trends® in Machine Learning, vol. 14, no. 1–2, pp. 1–210, 2021

    2021

  6. [6]

    Can decentralized algorithms outperform centralized algorithms? a case study for decentralized parallel stochastic gradient descent,

    X. Lian, C. Zhang, H. Zhang, C.-J. Hsieh, W. Zhang, and J. Liu, “Can decentralized algorithms outperform centralized algorithms? a case study for decentralized parallel stochastic gradient descent,” inAdvances in Neural Information Processing Systems, I. Guyon, U. V . Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, Eds., vol. 3...

    work page 2017

  7. [7]

    When federated learning meets blockchain: A new distributed learning paradigm,

    C. Ma, J. Li, L. Shi, M. Ding, T. Wang, Z. Han, and H. V . Poor, “When federated learning meets blockchain: A new distributed learning paradigm,”IEEE Computational Intelligence Magazine, vol. 17, no. 3, pp. 26–33, 2022

    2022

  8. [8]

    Blockchain- empowered decentralized horizontal federated learning for 5g-enabled uavs,

    C. Feng, B. Liu, K. Yu, S. K. Goudos, and S. Wan, “Blockchain- empowered decentralized horizontal federated learning for 5g-enabled uavs,”IEEE Transactions on Industrial Informatics, vol. 18, no. 5, pp. 3582–3592, 2021

    work page 2021

Show all 32 references
  1. [9]

    A survey on scalable lorawan for massive iot: Recent advances, potentials, and challenges,

    M. Jouhari, N. Saeed, M.-S. Alouini, and E. M. Amhoud, “A survey on scalable lorawan for massive iot: Recent advances, potentials, and challenges,”IEEE Communications Surveys & Tutorials, 2023

    2023

  2. [10]

    Decentralized federated learning: Fundamentals, state of the art, frameworks, trends, and challenges,

    E. T. M. Beltr ´an, M. Q. P ´erez, P. M. S. S ´anchez, S. L. Bernal, G. Bovet, M. G. P ´erez, G. M. P ´erez, and A. H. Celdr ´an, “Decentralized federated learning: Fundamentals, state of the art, frameworks, trends, and challenges,”IEEE Communications Surveys & Tutorials, 2023

    2023

  3. [11]

    Decentralized opti- mization over noisy, rate-constrained networks: Achieving consensus by communicating differences,

    R. Saha, S. Rini, M. Rao, and A. J. Goldsmith, “Decentralized opti- mization over noisy, rate-constrained networks: Achieving consensus by communicating differences,”IEEE Journal on Selected Areas in Communications, vol. 40, no. 2, pp. 449–467, 2021

    2021

  4. [12]

    Quantized decentralized stochastic learning over directed graphs,

    H. Taheri, A. Mokhtari, H. Hassani, and R. Pedarsani, “Quantized decentralized stochastic learning over directed graphs,” inInternational Conference on Machine Learning. PMLR, 2020, pp. 9324–9333

    2020

  5. [13]

    On the benefits of multiple gossip steps in communication-constrained decentralized federated learning,

    A. Hashemi, A. Acharya, R. Das, H. Vikalo, S. Sanghavi, and I. Dhillon, “On the benefits of multiple gossip steps in communication-constrained decentralized federated learning,”IEEE Transactions on Parallel and Distributed Systems, vol. 33, no. 11, pp. 2727–2739, 2021

    2021

  6. [14]

    Gossipfl: A decentralized federated learning framework with sparsified and adaptive communication,

    Z. Tang, S. Shi, B. Li, and X. Chu, “Gossipfl: A decentralized federated learning framework with sparsified and adaptive communication,”IEEE Transactions on Parallel and Distributed Systems, vol. 34, no. 3, pp. 909–922, 2022

    2022

  7. [15]

    Wait or not to wait: Evaluating trade-offs between speed and precision in blockchain- based federated aggregation,

    H. Nguyen, T. Nguyen, L. Lov ´en, and S. Pirttikangas, “Wait or not to wait: Evaluating trade-offs between speed and precision in blockchain- based federated aggregation,” in2024 IEEE 44th International Confer- ence on Distributed Computing Systems Workshops (ICDCSW), 2024, pp. 83–92

    2024

  8. [16]

    Tram-fl: Routing-based model training for decentralized federated learning,

    K. Maejima, T. Nishio, A. Yamazaki, and Y . Hara-Azumi, “Tram-fl: Routing-based model training for decentralized federated learning,” in 2024 IEEE 21st Consumer Communications & Networking Conference (CCNC). IEEE, 2024, pp. 1038–1039

    2024

  9. [17]

    Peer-to-peer mem- bership management for gossip-based protocols,

    A. J. Ganesh, A.-M. Kermarrec, and L. Massouli ´e, “Peer-to-peer mem- bership management for gossip-based protocols,”IEEE transactions on computers, vol. 52, no. 2, pp. 139–149, 2003

    2003

  10. [18]

    Decentralized federated learning: Balancing communication and computing costs,

    W. Liu, L. Chen, and W. Zhang, “Decentralized federated learning: Balancing communication and computing costs,”IEEE Transactions on Signal and Information Processing over Networks, vol. 8, pp. 131–143, 2022

    2022

  11. [19]

    Cross-silo federated learning for multi-tier networks with vertical and horizontal data par- titioning,

    A. Das, T. Castiglia, S. Wang, and S. Patterson, “Cross-silo federated learning for multi-tier networks with vertical and horizontal data par- titioning,”ACM Transactions on Intelligent Systems and Technology (TIST), vol. 13, no. 6, pp. 1–27, 2022

    2022

  12. [20]

    Intelligent neighbor selection for efficient query routing in unstructured p2p networks using q-learning,

    M. Shoab and S. A. Jubayrin, “Intelligent neighbor selection for efficient query routing in unstructured p2p networks using q-learning,”Applied Intelligence, pp. 1–10, 2022

    2022

  13. [21]

    Deep reinforcement learning based transmission policy enforcement and multi-hop routing in qos aware lora iot networks,

    M. S. A. Muthanna, A. Muthanna, A. Rafiq, M. Hammoudeh, R. Alkan- hel, S. Lynch, and A. A. Abd El-Latif, “Deep reinforcement learning based transmission policy enforcement and multi-hop routing in qos aware lora iot networks,”Computer Communications, vol. 183, pp. 33– 50, 2022

    2022

  14. [22]

    Sparsfa: Towards robust and communication-efficient peer-to-peer fed- erated learning,

    H. Wang, L. Mu ˜noz-Gonz´alez, M. Z. Hameed, D. Eklund, and S. Raza, “Sparsfa: Towards robust and communication-efficient peer-to-peer fed- erated learning,”Computers & Security, vol. 129, p. 103182, 2023

    2023

  15. [23]

    Decentralized federated learning: A segmented gossip approach,

    C. Hu, J. Jiang, and Z. Wang, “Decentralized federated learning: A segmented gossip approach,”arXiv preprint arXiv:1908.07782, 2019

    1908 arXiv

  16. [24]

    Beyond exponential graph: Communication-efficient topologies for decentralized learning via finite-time convergence,

    Y . Takezawa, R. Sato, H. Bao, K. Niwa, and M. Yamada, “Beyond exponential graph: Communication-efficient topologies for decentralized learning via finite-time convergence,”Advances in Neural Information Processing Systems, vol. 36, 2024

    2024

  17. [25]

    Reliable federated learning based on dual-reputation reverse auction mechanism in internet of things,

    Y . Tang, Y . Liang, Y . Liu, J. Zhang, L. Ni, and L. Qi, “Reliable federated learning based on dual-reputation reverse auction mechanism in internet of things,”Future Generation Computer Systems, vol. 156, pp. 269–284, 2024

    2024

  18. [26]

    On the shortest spanning subtree of a graph and the traveling salesman problem,

    J. B. Kruskal, “On the shortest spanning subtree of a graph and the traveling salesman problem,”Proceedings of the American Mathematical society, vol. 7, no. 1, pp. 48–50, 1956

    1956

  19. [27]

    Shortest connection networks and some generalizations,

    R. C. Prim, “Shortest connection networks and some generalizations,” The Bell System Technical Journal, vol. 36, no. 6, pp. 1389–1401, 1957

    1957

  20. [28]

    O jist ´em probl ´emu minim ´aln´ım,

    O. Boruvka, “O jist ´em probl ´emu minim ´aln´ım,”Pr ´ace Mor. Prırodved. Spol. v Brne (Acta Societ. Scienc. Natur. Moravicae), vol. 3, no. 3, pp. 37–58, 1926

    1926

  21. [29]

    On random graphs i,

    P. Erd ¨os and A. R ´enyi, “On random graphs i,”Publicationes Mathemat- icae Debrecen, vol. 6, pp. 290–297, 1959

    1959

  22. [30]

    Collective dynamics of ‘small- world’networks,

    D. J. Watts and S. H. Strogatz, “Collective dynamics of ‘small- world’networks,”nature, vol. 393, no. 6684, pp. 440–442, 1998

    1998

  23. [31]

    Emergence of scaling in random net- works,

    A.-L. Barab ´asi and R. Albert, “Emergence of scaling in random net- works,”Science, vol. 286, no. 5439, pp. 509–512, 1999

    1999

  24. [32]

    Flooding in wireless ad hoc networks,

    H. Lim and C. Kim, “Flooding in wireless ad hoc networks,”Computer communications, vol. 24, no. 3-4, pp. 353–363, 2001

    2001

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.