One-Shot Federated Graph Learning via High-Fidelity Proxies and Transferability-Guided Collaboration
Authors: Jia Jiang, Zihan Tan, Wenke Huang, Bin Yang, Mang Ye
Organizations: School of Computer Science, Wuhan University, Wuhan, China. · Guangdong Laboratory of Artificial Intelligence and Digital Economy, Shenzhen, China. · College of Computing and Data Science, Nanyang Technological University, Singapore.
One-shot Federated Graph Learning (OFGL) has emerged as a communication-efficient paradigm for federated graph learning under strict bandwidth constraints, compressing collaboration into a single transmission round. Nevertheless, this extreme setting suffers from a fundamental Single-Transmission Bottleneck, which limits both the quality and utility of exchanged knowledge. First, while existing methods attempt to distill private graphs into shareable proxies, their aggressive one-shot compression inherently triggers a fidelity crisis. This compression distorts local graph topology and semantics across clients, leading to externalized knowledge underexpression. Second, although the server aggregates these proxies to update global models, current integration schemes suffer from transferability agnosticism. The server cannot reliably identify cross-client collaboration potential, thus resulting in integrated knowledge underutilization. To tackle these dual challenges, we propose FedHFT, an effective One-Shot Personalized Federated Graph Learning (OPFGL) framework that addresses this bottleneck from two perspectives, namely High-Fidelity Knowledge Externalization and Transferability-Ware Knowledge Integration. We use Fidelity-Driven Proxy Refinement to preserve client-side knowledge fidelity, and Transferability-Steered Personalized Collaboration to enable adaptive server-side collaboration. The superiority of FedHFT is validated through extensive experiments on both homophilic and heterophilic graphs. The code is available at https://github.com/Jumpy-Juice/FedHFT.
Figures & tables
Figure 1: Single-Transmission Bottleneck. With increasing training steps, the optimization gradients induced by existing methods progressively deviate from the true optimization direction, highlighting the bottleneck in the OFGL setting.
Figure 2: Problem Illustration. The Single-Transmission Bottleneck in OFGL impedes knowledge transfer via two challenges. (a) Externalized Knowledge Underexpression: Aggressive compression damages graph topology, causing a fidelity crisis . (b) Integrated Knowledge Underutilization: Compressed capacity obscures cross-client collaboration potential, driving transferability agnosticism .
Figure 3: Architecture of FedHFT for one-shot federated graph learning. Top: Client-side FPR. Clients generate high-fidelity proxy knowledge via Fisher-guided gradient matching and top- M class-semantic feature refinement prior to the single upload. Bottom: Server-side TPC. The server performs personalized aggregation using inter-client transferability matrices derived from per-class NLL of the uploaded prototypes.
Category
Methods
Cora
CiteSeer
PubMed
Actor
Amz-Photo
Roman-Empire
Ogbn-Arxiv
FL
FedAvg [ASTAT17]
34.09
43.00
55.45
24.00
29.87
21.41
23.83
FedProx [MLSys20]
33.31 ↓0.78
41.84 ↓1.16
43.04 ↓12.40
22.69 ↓1.30
27.44 ↓2.42
19.32 ↓2.09
23.58 ↓0.25
FedNova [NeurIPS20]
15.50 ↓18.59
16.45 ↓26.55
35.33 ↓20.12
20.01 ↓3.98
13.00 ↓16.86
6.64 ↓14.77
3.20 ↓20.64
FGL
FedPub [ICML23]
31.06 ↓3.03
24.63 ↓18.37
41.29 ↓14.15
24.19 ↑0.19
25.69 ↓4.18
17.06 ↓4.35
14.34 ↓9.49
FedTAD [IJCAI24]
34.62 ↑0.53
46.37 ↑3.36
47.40 ↓8.04
21.71 ↓2.28
28.67 ↓1.20
20.34 ↓1.07
11.50 ↓12.34
FedGTA [VLDB24]
15.94 ↓18.14
14.15 ↓28.86
8.27 ↓47.18
14.20 ↓9.79
9.33 ↓20.53
13.29 ↓8.12
8.01 ↓15.82
Table 1: Comparison with state-of-the-art baselines on seven datasets. We report node classification accuracy (%) for downstream performance evaluation. Green arrows ↑ indicate improvements over FedAvg, while red arrows ↓ denote performance drops relative to FedAvg. OOM means out-of-memory error. The best and second-best results in each column are marked in bold and underline, respectively. Among the baselines, FedPub and FedGTA are personalized methods, while the others learn a single global model.
Table 5
Figure 4: Analysis on hyper-parameters in FedHFT. Node classification results under different refinement weights (a) and temperature coefficient (b). The results illustrate the performance sensitivity across representative datasets, demonstrating that FedHFT maintains stable accuracy over a broad range of parameter values. The dashed lines represent the baseline test accuracy for each setting.
Figure 5: Proxy Data Analysis. Our method consistently performs best across all four proxy-quality metrics.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Notation
Definition
K
The number of clients.
k,i,j
The client indices.
Nk
The number of nodes on client k .
C
The total number of classes.
dk
The feature dimension of client k .
fθ
The initial global model.
Appendix
Table 4: Notation and Definitions
Figure 6: Sensitivity analysis of the Top- M selection ratio.
Figure 7: Correlation between estimated transferability and realized transfer accuracy.
Dataset
#Nodes
#Edges
#Classes
#Features
Cora
2,708
5,278
7
1,433
CiteSeer
3,327
4,552
6
3,703
PubMed
19,717
44,324
3
500
Amz-Photo
7,650
287,326
8
745
OGBN-Arxiv
169,343
1,166,243
40
128
Actor
7,600
30,019
5
932
Appendix
Table 5: Statistics of datasets used in experiments.
Datasets ( → )
Cora
CiteSeer
Category
Methods ( ↓ )
5 Clients
10 Clients
20 Clients
5 Clients
10 Clients
20 Clients
FL
FedAvg [AISTATS17]
41.27
34.09
30.46
47.21
43.00
38.13
FedProx [MLSys20]
40.16 ↓1.11
33.31 ↓0.78
31.60 ↑1.14
44.20 ↓3.01
41.84 ↓1.16
39.17 ↑1.04
FedNova [NeurIPS20]
16.25 ↓25.02
15.50 ↓18.59
17.44 ↓13.02
16.48 ↓30.73
16.45 ↓26.55
17.26 ↓20.87
FGL
FedPub [ICML23]
36.79 ↓4.48
31.06 ↓3.03
29.88 ↓0.58
35.62 ↓11.59
24.63 ↓18.37
28.45 ↓9.68
FedTAD [IJCAI24]
40.50 ↓0.77
34.62 ↑0.53
31.54 ↑1.08
44.63 ↓2.58
46.37 ↑3.37
42.21 ↑4.08
Appendix
Table 6: Comparison with state-of-the-art baselines under different client numbers (5, 10, 20) on Cora, CiteSeer, Actor, and Roman-Empire datasets. Green arrows ↑ indicate improvements over FedAvg, while red arrows ↓ denote performance drops relative to FedAvg. The best and second-best results in each column are marked in bold and underline, respectively.
Category
Methods
Cora
CiteSeer
PubMed
Actor
Amz-Photo
Roman-Empire
Ogbn-Arxiv
FL
FedAvg [AISTATS17]
38.56
36.11
56.58
22.80
32.64
23.06
26.04
FedProx [MLSys20]
36.12 ↓2.44
46.59 ↑10.48
40.36 ↓16.22
22.38 ↓0.42
27.88 ↓4.76
22.76 ↓0.30
25.70 ↓0.34
FedNova [NeurIPS20]
15.31 ↓23.25
16.95 ↓19.16
35.69 ↓20.89
19.81 ↓2.99
12.29 ↓20.35
6.54 ↓16.52
3.23 ↓22.81
FGL
FedPub [ICML23]
30.77 ↓7.79
28.41 ↓7.70
44.68 ↓11.90
23.14 ↑0.34
32.06 ↓0.58
20.82 ↓2.24
20.01 ↓6.03
FedTAD [IJCAI24]
38.26 ↓0.30
40.83 ↑4.72
52.27 ↓4.31
21.14 ↓1.66
26.67 ↓5.97
22.31 ↓0.75
15.30 ↓10.74
FedGTA [VLDB24]
12.98 ↓25.58
14.84 ↓21.27
7.41 ↓49.17
19.50 ↓3.30
5.65 ↓26.99
14.23 ↓8.83
9.06 ↓16.98
Appendix
Table 7: Comparison with state-of-the-art FL, FGL, OFL and OFGL baselines on seven datasets ( α=0.1 ). We report node classification accuracy (%) for downstream performance evaluation. Green arrows ↑ indicate improvements over FedAvg, while red arrows ↓ denote performance drops relative to FedAvg. OOM means out-of-memory error. The best and second-best results in each column are marked in bold and underline, respectively.
Figure 14
Methods ( ↓ )
Cora
CiteSeer
PubMed
Actor
Amz-Photo
Roman-Empire
Ogbn-Arxiv
FedAvg [AISTATS17]
0.3326
0.5411
0.5366
0.2561
0.3380
0.2190
0.2376
FedProx [MLSys20]
0.3417
0.5350
0.5397
0.2542
0.4402
0.2183
0.2348
FedTAD [IJCAI24]
0.3468
0.5136
0.5196
0.2534
0.3634
0.2249
0.1011
FGGP [AAAI24]
0.3318
0.4965
0.4978
0.2489
0.3948
0.2231
OOM
FedHFT (ours)
0.5224
0.5602
0.5698
0.4151
0.4931
0.4082
0.4450
Appendix
Table 8: Comparison under matched total communication budgets: one-round FedHFT versus multi-round baselines on seven datasets. We report node classification accuracy (%) for downstream performance evaluation. OOM means out-of-memory error. The best and second-best results in each column are marked in bold and underline, respectively. On Ogbn-Arxiv, FedAvg, FedProx, and FedTAD are allowed three rounds and FGGP two rounds; all baselines use two rounds on the other datasets; all ten FGGP runs on Ogbn-Arxiv failed with OOM. FedHFT uses one round in all cases.
One-shot federated graph learning generally aims to train Graph Neural Networks (GNNs) across clients with disconnected subgraphs in a single communication round. Existing methods predominantly design advanced optimization strategies under the premise that local GNN training is indispensable. However, empirical observations reveal that under extreme non-IID conditions, local GNN training suffers from severe cross-client representation misalignment, becoming a major source of error rather than a remedy. Motivated by this, we reformulate one-shot FGL as a statistical estimation problem. We propose SPEAR (Statistical Prototype Estimation with Adaptive Reliability), a completely training-free framework that directly computes topology-smoothed class prototypes from local graphs in the original feature space. The server then aggregates these prototypes using a sample-size-adaptive shrinkage estimator that down-weights unreliable local estimates, producing robust global class prototypes. Extensive experiments across seven benchmarks demonstrate that SPEAR consistently achieves state-of-the-art accuracy under extreme heterogeneity. Moreover, SPEAR delivers at least an order-of-magnitude speedup over all baselines, reaching several orders of magnitude against generative and distillation-based methods. Our findings suggest that training-free statistical estimation, rather than local GNN optimization, provides the key to robust and efficient one-shot federated graph learning. The code is available at https://github.com/Yodeesy/SPEAR .
Shutong Zheng, Sijia Chen
Sun Yat-sen University · The Hong Kong University of Science and Technology (Guangzhou)
One-shot federated graph learning (FGL) requires the server to estimate client contributions from highly compressed information, yet conventional volume-based weighting captures the amount of client data while overlooking how its connectivity is organized. In this paper, we propose SPIRE, a Structural Entropy-Driven Graph Diffusion Generation method that introduces topology-aware client differentiation into one-shot FGL. Specifically, we employ first-order degree-distribution structural entropy as a compact descriptor of degree-mass dispersion and use it to derive structural client weights, providing an inductive bias that accounts for differences in graph topology beyond data volume. On the generation side, a graph diffusion model on the server synthesizes pseudographs conditioned on the weighted client prototypes, capturing both semantic and structural information without requiring additional client-side training. The generated pseudographs are then assembled via disjoint union fusion to train a global graph neural network. Extensive experiments on seven real-world graph datasets demonstrate that SPIRE consistently outperforms conventional and one-shot FGL methods, with particularly strong gains under highly heterogeneous (non-IID) and graph-perturbed settings.
Shutong Zheng, Lele Fu, Sheng Huang +2
Sun Yat-sen University Guangzhou, China · Nanyang Technological University Singapore
Client heterogeneity creates both an opportunity and a risk in personalized federated graph learning. Knowledge held by other subgraphs may complement a receiver's Local model, but an incompatible transfer can override reliable predictions. One-shot communication sharpens this tension because an unsuitable server return cannot be corrected later. We introduce PACE, which treats collaborative knowledge as a compact correction to a complete Local predictor rather than as its replacement. Each client uploads a rank-r update carrier and a diagonal sketch of propagated message moments. The server uses them to construct a propagation-aware, receiver-anchored correction, while the receiver retains its full Local model. Convex negative-log-likelihood calibration (CNLL) then selects one coefficient between Local and External logits using validation nodes; model parameters remain fixed and no feedback is sent. At Rank-6, personalized returns occupy 9.6-17.6% of dense tensor bytes across the six evaluated datasets. The correction receives nonzero weight and improves both Accuracy and weighted-F1 over Local on five datasets; on ogbn-arxiv, CNLL assigns zero predictive weight to the correction and preserves Local predictions exactly. Applying the same CNLL rule to matched baselines on three citation datasets does not account for these gains. The central result is therefore that a small transported correction can augment a complete Local model when receiver evidence supports it while leaving the Local prediction unchanged otherwise.
Ruizhe Huang, Chengran Li, Xiaochuan Shi
School of Cyber Science and Engineering, Wuhan University