acceptodds
Under review as a conference paper at ICLR 2027

Who Knows What? One-Shot Federated Graph Learning via Diffusion

Abstract

One-shot federated graph learning is well suited to communication-constrained environments because each client communicates with the server only once. However, this setting is particularly challenging when clients hold graphs that differ in their node features, label distributions, connectivity patterns, and local topology. Conventional aggregation treats each client uniformly across all prediction classes, overlooking variations in class-specific expertise caused by heterogeneous local data and graph structure. We propose CDRA, a lightweight Class-wise Diffusion Reliability Aggregation framework for one-shot federated graph learning under severe client heterogeneity. Each client trains a local graph neural network and constructs compact class-conditioned diffusion signatures that capture label-specific structural patterns within its graph. The server uses Hellinger distance to compare these signatures with federation-level class references and estimate each client’s reliability for different prediction classes. These reliability estimates, together with client size and class support, determine class-wise aggregation weights for combining local model predictions. Experiments on seven graph benchmarks under severe non-IID heterogeneity show that CDRA improves predictive performance while substantially reducing runtime, peak GPU memory, and upload communication relative to reconstruction-based one-shot federated graph learning methods. Anonymous code is available at: https://anonymous.4open.science/r/CDRA/README.md

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Federated Learning, Graph Neural Networks, Federated Graph Learning, One-Shot Federated Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_zgMLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies one-shot federated graph learning (OFGL) for node classification: KK clients each hold a private subgraph, train a local GNN, and communicate with the server exactly once. The proposed method, CDRA, has each client upload (i) its trained model, (ii) for every class with at least τ\tau local training nodes a "class-conditioned diffusion signature" Sk,c∈RH×CS_{k,c}\in\mathbb{R}^{H\times C} — the class-cc rows of the row-normalised class-mixing matrices Yk⊤PkhYkY_k^\top P_k^h Y_k for h=1..Hh=1..H, i.e. how label mass propagated from class-cc training nodes is distributed over the classes present on that client (Eqs. 5–8) — and (iii) the class-support counts nk,cn_{k,c}. The server forms leave-one-client-out class references (Eq. 14), scores each (client, class) pair by a Hellinger margin (distance to the nearest competing-class reference minus distance to the same-class reference, median-normalised; Eqs. 15–20), combines it with a client-size prior αk1/2\alpha_k^{1/2} and a class-support prior nk,cγn_{k,c}^{\gamma} into per-class client weights wk,cw_{k,c} (Eqs. 21–25), and fuses the KK local predictions class-wise (Eq. 26). The payload is HC2HC^2 floats plus CC counts, independent of graph size and feature dimension; no generator, proxy graph or server-side training is needed. Experiments on seven node-classification datasets with K=10K=10 clients and Dirichlet label skew αDir=0.01\alpha_{\mathrm{Dir}}=0.01 report the best accuracy on all datasets (average 33.3% vs. 23.9% for the best baseline; Table 1) and large reductions in runtime, peak GPU memory and upload size relative to the reconstruction-based OFGL methods GHOST and OASIS (Fig. 2c, Table 7). The appendices add a short theoretical analysis (the signature distance is a bounded metric; signature error is O(nk,c−1/2)O(n_{k,c}^{-1/2}); a triangle-inequality certificate that the margin is positive when population class profiles are well separated), an ablation on Cora and OGBN-Arxiv (Table 4), GAT/GraphSAGE backbones (Table 5), a client-scaling study up to K=320K=320 in a milder partition (Table 6), and sensitivity to the diffusion depth HH (Fig. 3).

Critical strengths

1. A lightweight, practical design for a setting where existing OFGL methods are heavy. The structural payload is HC2HC^2 floats plus CC counts regardless of graph size or feature dimension, the server does no training, and Table 7 shows roughly 50–60× lower runtime than GHOST/OASIS on the two largest graphs, 30–50× lower peak memory on OGBN-Arxiv, and the smallest upload on every dataset. This is consequential for my assessment because removing generator training and proxy-graph materialisation addresses the main practical objection to current one-shot federated graph learning pipelines, and the efficiency numbers are concrete and reproducible (anonymised code is provided).

2. Class-wise rather than client-wise weighting is the right inductive bias for label-skewed federations, and the paper shows it is the operative ingredient. Collapsing the same reliability estimates into a single weight per client ("Global CDRA") drops Cora accuracy from 45.95% to 30.90% (Table 4). The paper is clearly written, states its threat model and limitations honestly (Apps. J–K), and the method is simple enough to serve as a baseline for future OFGL work.

Critical weaknesses

1. The headline gains are measured against baselines that are degenerate in this regime, while the most natural baseline appears only in an appendix — and on the largest dataset it essentially matches CDRA. With αDir=0.01\alpha_{\mathrm{Dir}}=0.01 and K=10K=10, almost every method in Table 1 sits at the majority-class or chance level: on Cora, FedAvg, FedProx, FedPub, FGSSL and FENS all land at 30.2–30.8%, i.e. the frequency of Cora's largest class; on the 5-class Chameleon and Actor graphs every baseline is within a few points of the 20% chance level; on OGBN-Arxiv, FedNova (0.81%) and FedSD2C (0.95%) are below random guessing over 40 classes. Appendix D confirms that this is an artefact of forcing multi-round methods into a single round. Against this backdrop, Table 4 (App. E) shows that a plain Uniform Ensemble of the same local models already reaches 30.54% on OGBN-Arxiv — 11 points above every Table 1 baseline — versus 32.50 ± 6.0% for full CDRA, and that on Cora the Support-Only variant (class-count weighting, no diffusion signatures) reaches 42.50 ± 4.5% of CDRA's 45.95 ± 2.3%. The diffusion-specific contribution is therefore within one standard deviation on both ablation datasets; H=1H=1 and H=3H=3 give identical accuracy (Fig. 3); the backbone study (Table 5) shows the same pattern on Cora (39.24 ± 8.6 vs. OASIS 37.77 ± 2.6 with GAT; 36.22 ± 5.7 vs. 32.00 ± 4.2 with GraphSAGE); and the number of seeds behind every "±" is never stated. The advantage is also not consistent across skew levels: reading Fig. 2(a), OASIS is ahead of CDRA at αDir∈{0.1,10,100}\alpha_{\mathrm{Dir}}\in\{0.1, 10, 100\} (about 73% vs. 62% at αDir=100\alpha_{\mathrm{Dir}}=100), so "remains competitive across the full range" (Sec. 6.2) is generous, and CDRA has no global model to fall back on when skew is mild. As written, Table 1 establishes that CDRA beats baselines that are broken in this protocol; it does not establish that the proposed mechanism (diffusion signatures + Hellinger margins) adds anything beyond class-count-weighted ensembling, which is the paper's central claim. To be convincing, Table 1 needs the Uniform-Ensemble and Support-Only rows (ideally plus a centralised upper bound) on all seven datasets with the same seeds, the seed count, and a significance assessment.

2. What the signature actually measures is under-examined, and the protocol details that determine its meaning are missing. Row cc of RowNorm(Y⊤PhY)\mathrm{RowNorm}(Y^\top P^h Y) is the distribution of label mass propagated from class-cc training nodes over the training labels present on the client. Under αDir=0.01\alpha_{\mathrm{Dir}}=0.01, a client that holds (almost) only class cc obtains Sk,c≈ecS_{k,c}\approx e_c for every hh regardless of its topology, while a client where cc is a minority spreads mass over its majority classes; the Hellinger margin then largely compares local label marginals, information the server already receives as nk,cn_{k,c}, rather than "label-specific structural patterns" (abstract). No control separates these explanations (e.g. a signature built from the label histogram alone, or signatures computed on degree-preserving randomly rewired edges). Coverage (Eqs. 58/60) is reported only for the αDir=5, τ=2\alpha_{\mathrm{Dir}}=5,\ \tau=2 scaling study; in the main setting it may be low, in which case most margins are neutral (Eq. 20) and the method reduces to size/support weighting. In addition, the paper never states (a) how edges are assigned once nodes are Dirichlet-partitioned (induced subgraphs? cross-client edges dropped?) and whether client subgraphs are connected; (b) where test nodes live and who computes pk(v)p_k(v) for a node held by client j≠kj\neq k — this requires every client (or the server) to hold all KK models and the test graph, a download/inference cost absent from the communication accounting and from the "one communication" narrative; (c) whether checkpoint selection uses local or global validation accuracy; and (d) how the seven hyperparameters (τ,H,γ,λ,Td,Tg,β\tau,H,\gamma,\lambda,T_d,T_g,\beta; App. B) and the baselines' hyperparameters were chosen, only HH has a sensitivity study. Without (a)–(d) the results cannot be interpreted or reproduced from the paper alone.

Questions for authors
  1. How many seeds underlie each "±" in Tables 1, 4–6 and Figs. 2–3? Please add Uniform Ensemble and Support-Only rows to Table 1 on all seven datasets with the same seeds, together with a significance assessment.
  2. Does a label-histogram-only signature, or a signature computed on degree-preserving randomly rewired edges, reproduce the gains? This directly tests the structural interpretation of the signatures.
  3. How are edges assigned to clients after the Dirichlet partition of nodes? Are the client subgraphs connected? (Label-skew partitions of a homophilous graph produce highly fragmented induced subgraphs.)
  4. Inference protocol: where are the test nodes, on which graph is pk(v)p_k(v) computed, and does each client download all KK models? What are the resulting download size and inference cost as KK grows?
  5. How were τ,γ,λ,Td,Tg,β\tau,\gamma,\lambda,T_d,T_g,\beta and the baselines' hyperparameters selected, and which validation split (local or global) is used for checkpoint selection? Please report sensitivity to τ\tau, β\beta and λ\lambda.
  6. For classes with 0<nk,c<τ0<n_{k,c}<\tau, is nk,cn_{k,c} still uploaded and used in Eq. (21), or treated as zero?
  7. Which dataset does Table 6 use? Which methods appear in Fig. 2(a)/(b), and were they re-tuned for each αDir\alpha_{\mathrm{Dir}} and KK?
  8. FedTAD on Cora (31.42 ± 1.4) has exactly the same mean as the Uniform Ensemble (31.42 ± 5.3). Is this a coincidence, or does the one-shot FedTAD implementation reduce to an ensemble?
4 comments

Dear reviewer zgML,

We sincerely thank you for your critical and constructive evaluation of our work. We address all comments below with detailed explanations and additional experimental evidence. The corresponding results and clarifications will be incorporated into the final version of the paper.

Response to Q1: Random Seeds, Additional Baselines, and Statistical Significance

Random Seeds. All results in Tables 1, 4–6 and Figures 2–3 are obtained using three random seeds (0, 1, and 2), with results reported as mean ± standard deviation.

Additional Baselines. We additionally evaluate Uniform Ensemble and Support Only on all seven datasets using the same random seeds and experimental settings as CDRA. The results are presented below.

Table R1: Test accuracy (%) across seven datasets (mean ± SD).

MethodCoraCiteSeerChameleonActorRoman-EmpireOGBN-ArxivCoauthor-CS
Uniform Ensemble31.42 ± 5.323.69 ± 2.433.59 ± 6.428.60 ± 1.618.93 ± 6.330.29 ± 7.828.31 ± 3.9
Support Only42.50 ± 4.542.19 ± 12.932.61 ± 7.629.01 ± 2.317.88 ± 3.827.15 ± 2.728.39 ± 3.7
CDRA (Ours)45.95 ± 2.340.40 ± 3.933.48 ± 7.528.57 ± 1.420.81 ± 3.332.50 ± 6.031.12 ± 4.1

Statistical Significance. We conduct two-sided paired t-tests using the seed-wise accuracies of CDRA, Uniform Ensemble, and Support Only. The resulting unadjusted p-values are reported below.

Table R2: Paired t-test p-values for CDRA against the additional baselines.

DatasetCDRA vs. UniformCDRA vs. Support Only
Cora0.0530.541
CiteSeer0.0530.741
Chameleon0.9760.566
Actor0.9770.594
Roman-Empire0.3550.143
OGBN-Arxiv0.5220.436
Coauthor-CS0.3910.046

Discussion. CDRA achieves higher mean accuracy than Uniform Ensemble on five of seven datasets and than Support Only on five of seven datasets. The comparison with Support Only on Coauthor-CS reaches nominal significance (p=0.046p=0.046), whereas the remaining comparisons do not meet the conventional p<0.05p<0.05 threshold. Given the limited statistical power of three seeds and the multiple comparisons performed, we interpret these findings cautiously and do not claim statistically established superiority across all datasets.

These additional experiments and statistical results provide a more transparent assessment of CDRA's accuracy improvements and variability.

Response to Q2: Assessing the Contribution of Graph Structure through Label-Histogram and Degree-Preserving Edge-Rewiring Ablations

To investigate whether CDRA's improvements arise from meaningful graph structural information rather than label statistics or degree distributions alone, we conduct two additional experiments across all seven datasets:

  1. Label Histogram: A label-only baseline that computes class-wise ensemble weights using clients' local training-label histograms, without using graph topology for aggregation.
  2. Degree-Preserving Edge Rewiring: We randomly rewire each client's graph using degree-preserving double-edge swaps, targeting three successful swaps per eligible edge, subject to graph constraints. We then reconstruct the diffusion signatures using the rewired topology while retaining the original graph for local GNN training and evaluation. The remaining reliability estimation and aggregation procedure is unchanged.

The results (mean ± standard deviation over three different seeds, accuracy in %) are:

DatasetCDRA (Ours)Label HistogramEdge Rewired
Cora45.95 ± 2.3031.88 ± 2.3247.73 ± 1.45
CiteSeer40.40 ± 3.9029.68 ± 9.0135.19 ± 3.42
Chameleon33.48 ± 7.5028.44 ± 2.4733.11 ± 8.82
Actor28.57 ± 1.4026.72 ± 1.8428.23 ± 1.44
Roman-Empire20.81 ± 3.3017.99 ± 5.5420.23 ± 4.68
OGBN-Arxiv32.50 ± 6.0028.09 ± 1.8732.37 ± 7.79
Coauthor-CS31.12 ± 4.1030.20 ± 8.9630.14 ± 4.08

Findings. CDRA consistently outperforms the label-histogram baseline across all seven datasets, with improvements ranging from 0.92 to 14.07 percentage points, indicating that label-frequency-based aggregation alone does not reproduce its performance. Compared with degree-preserving rewiring, CDRA achieves higher mean accuracy on six of seven datasets, with the largest improvement on CiteSeer (+5.21 percentage points). However, the rewired variant performs better on Cora, and the differences on several other datasets are modest. These results suggest that the benefit of preserving the original graph connectivity is dataset-dependent, rather than universal.

Response to Q3: Edge Assignment and Connectivity of Client Subgraphs

Edge Assignment. We employ a class-wise Dirichlet node-partitioning strategy to distribute nodes among clients. Following node assignment, each client's local graph is constructed as an induced subgraph of the original graph. Specifically, an edge is retained only if both endpoints belong to the same client, while cross-client edges are discarded. Node indices are subsequently remapped to the local graph, and no synthetic edges are introduced.

Subgraph Connectivity. We do not explicitly enforce connectivity during partitioning. Consequently, client subgraphs may contain multiple connected components or isolated nodes, particularly under extreme label-skew settings. CDRA operates directly on these induced subgraphs without requiring connectivity. The diffusion operator is constructed using the available within-client edges, with self-loops added during normalization to accommodate isolated nodes.

We acknowledge that graph fragmentation can reduce the structural information available locally and potentially limit the informativeness of diffusion signatures. We will clarify the partitioning protocol, connectivity assumptions, and their implications in the final version of the paper.

Response to Q4: Inference Protocol, Model Downloads, and Scalability

Inference Protocol. We follow a server-side inference protocol using a held-out global test set, consistently applied to CDRA and all competing baselines. Test nodes are excluded from local model training. Each locally trained model fkf_k is evaluated on the same server-accessible evaluation graph to obtain its class probabilities pk(v)p_k(v) for the test nodes. The server then combines these predictions using CDRA's learned class-wise reliability weights.

Model Downloads and Communication. Clients do not download the other KK models for inference. Instead, each client uploads its trained model and compact signature information to the server once. The server performs ensemble inference centrally. Consequently, there is no client-side model-download overhead associated with ensemble inference. However, the server receives and maintains KK local models, resulting in model-upload volume and server-side storage that scale linearly with K.K.

Inference Cost and Scalability. Since the server evaluates all KK local models on the held-out test graph, the inference complexity is O(K⋅TGNN)O(K \cdot T_{\mathrm{GNN}}), where TGNNT_{\mathrm{GNN}} denotes the cost of one GNN forward pass on the test graph. Thus, inference cost grows linearly with the number of clients, while the learned reliability weights introduce only a small additional aggregation overhead.

Response to Q5: Hyperparameter Selection, Validation Protocol, and Sensitivity Analysis

Hyperparameter Configuration. We use fixed hyperparameter values across the main experiments: τ=10\tau=10, λ=1.0\lambda=1.0, γ=0.5\gamma=0.5, Tg=Td=1.5T_g=T_d=1.5, and β=0.8\beta=0.8. These values were set as fixed experimental configurations rather than selected through dataset-specific hyperparameter tuning. The baselines also use fixed method-specific configurations, without re-tuning for each Dirichlet concentration or client count.

Checkpoint Selection. Each client's checkpoint is selected using its local validation accuracy, not a global validation set. The held-out test set is used exclusively for final evaluation.

Sensitivity Analysis. We additionally evaluate sensitivity to τ\tau, β\beta, and λ\lambda on Cora and Chameleon, varying one parameter at a time while keeping all remaining settings fixed.

Table 1: Sensitivity to the minimum class-support threshold τ\tau. Test accuracy (%), mean ± standard deviation.

τ\tauCoraChameleon
147.64 ± 1.133.11 ± 7.2
247.36 ± 0.932.75 ± 7.2
547.02 ± 0.933.26 ± 7.8
10 (default)45.95 ± 2.333.48 ± 7.5
2046.59 ± 3.032.57 ± 6.8

Table 2: Sensitivity to the class-wise blending coefficient β\beta. Test accuracy (%), mean ± standard deviation.

β\betaCoraChameleon
0.030.90 ± 0.531.88 ± 2.0
0.234.99 ± 3.533.22 ± 2.2
0.543.90 ± 8.533.88 ± 6.4
0.8 (default)45.95 ± 2.333.48 ± 7.5
1.046.72 ± 4.133.08 ± 7.3

Table 3: Sensitivity to the structural reliability coefficient λ\lambda. Test accuracy (%), mean ± standard deviation.

λ\lambdaCoraChameleon
0.046.57 ± 0.732.90 ± 7.5
0.546.10 ± 1.633.33 ± 7.5
1.0 (default)45.95 ± 2.333.48 ± 7.5
1.545.25 ± 3.132.71 ± 7.1
2.045.07 ± 4.033.37 ± 6.8

Observations. Performance remains relatively stable across the evaluated τ\tau and λ\lambda values, while β\beta has a stronger effect on Cora, highlighting the importance of class-wise aggregation. The best mean performance does not always occur at the default settings, and differences are generally modest relative to the reported variability. These results and the hyperparameter-selection protocol will be incorporated into the final manuscript.

Response to Q6: Treatment of Class Support Below the Minimum Threshold

For classes with 0<nk,c<τ0 < n_{k,c} < \tau, the exact support count is not uploaded and is instead reported as zero. The corresponding class-conditioned diffusion signature is also suppressed. Consequently, these client–class pairs do not contribute to the raw class-specific reliability score in Eq. (21), i.e., rk,c=0r_{k,c}=0.

Only classes satisfying nk,c≥τn_{k,c}\geq\tau contribute their support counts and diffusion signatures to reliability estimation. This prevents insufficiently supported classes from producing unreliable structural estimates. We will explicitly clarify this thresholding behavior in the final manuscript.

Response to Q7: Dataset, Baseline Selection, and Hyperparameter Consistency

Table 6. The extended client-scaling experiments in Table 6 are conducted on the Cora dataset, with Dirichlet concentration αDir=5\alpha_{Dir}=5 and minimum support threshold τ=2\tau=2.

Methods in Figures 2(a) and 2(b). For clarity, we select the strongest-performing baseline from each methodological category based on its average accuracy across all seven datasets in Table 1. Specifically, the selected methods are FedProx (traditional FL), FedTAD (traditional FGL), FENS (one-shot FL), and OASIS (one-shot FGL), alongside CDRA (Ours). These methods are identified in the figure legends. Figure 2(a) evaluates label heterogeneity on Cora, whereas Figure 2(b) evaluates client scaling on Coauthor-CS.

Hyperparameter Consistency. The methods were not independently re-tuned for each Dirichlet concentration or client count. In Figure 2(a), only αDir\alpha_{Dir} is varied, while in Figure 2(b), only the number of clients KK is varied. All other experimental configurations and method-specific hyperparameters remain fixed, ensuring consistent comparisons across settings.

We will explicitly clarify the dataset used in Table 6, the baseline selection criteria, and the fixed-hyperparameter evaluation protocol in the final manuscript.

Response to Q8: Clarification of FedTAD and Uniform Ensemble Results

The identical mean accuracy of FedTAD and Uniform Ensemble on Cora (31.42%) is a numerical coincidence at the reported precision, rather than an equivalence between the two methods.

Uniform Ensemble directly averages the predictions of locally trained models without additional server-side optimization. In contrast, FedTAD employs topology-aware, data-free knowledge distillation to transfer knowledge from local models into a global model. In our one-shot adaptation, client-server communication is restricted to a single round, while FedTAD's server-side distillation procedure is retained.

Furthermore, the two methods exhibit substantially different standard deviations (1.4 vs. 5.3 percentage points), indicating different variability across random seeds despite their identical rounded means.

Reviewer_K6CHOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies one-shot federated node classification under severe client heterogeneity and proposes Class-wise Diffusion Reliability Aggregation (CDRA). Clients construct multi-hop class-conditioned diffusion signatures from local graph structures and training labels, then upload them alongside local models and class-support counts. The server constructs leave-one-client-out class references and estimates structural reliability through Hellinger-distance margins between matching and competing classes. These estimates are combined with client size and class support to produce class-specific weights for prediction fusion. Experiments on seven graph benchmarks report higher accuracy than the selected baselines, with lower training time, peak GPU memory, and upload communication than reconstruction-based methods.

Critical strengths
  1. The method is simple and offers meaningful resource advantages. CDRA constructs class-level structural summaries through sparse propagation, avoiding auxiliary generator training, proxy-graph reconstruction, and additional server-side optimization. The reported reductions in training time, GPU memory, and upload size relative to reconstruction-based competitors are practically relevant to resource-constrained one-shot federations.

  2. The evaluation has reasonable breadth. The seven benchmarks cover homophilous and heterophilous graphs, including a larger graph, and CDRA achieves the highest mean accuracy across the main table. The appendices provide additional backbone, ablation, and client-scaling experiments, while explicitly acknowledging limitations related to cross-client class support and ensemble deployment.

Critical weaknesses
  1. The connection between structural consistency and predictive reliability lacks direct evidence. CDRA estimates reliability from the distance margin between a client’s class-conditioned signature and matching versus competing class references. However, structural agreement with other clients does not necessarily imply better classification, and structurally different clients may provide useful complementary knowledge. Appendix C establishes distance properties, concentration bounds, and sufficient conditions for positive margins, but does not connect the margin to class-specific prediction error or aggregation gains. Direct validation is needed to establish whether the margin predicts client competence after controlling for class support and client size.

  2. The technical advance over existing topology-aware aggregation needs clearer justification. FedGTA [1] already uses multi-step propagation to extract structural statistics and combines them with local smoothing confidence to guide server-side aggregation. CDRA introduces distinct mechanisms through class-conditioned signatures, Hellinger margins, and class-wise prediction fusion, but the additional information provided by these mechanisms is not sufficiently demonstrated. Although FedGTA is included in the main comparison, restricting its parameter aggregation to one round does not isolate the contribution of the new reliability measure. Comparing propagated-prediction confidence against the proposed margin within the same prediction-fusion framework would be more informative.

  3. The ablations do not sufficiently isolate the diffusion component. Table 4 lacks a direct ablation that removes only the diffusion contribution while retaining client size, class support, class-wise fusion, and global mixing. On OGBN-Arxiv, uniform ensembling achieves 30.54 ± 6.8%, compared with 32.50 ± 6.0% for full CDRA and 28.08 ± 1.1% for the diffusion-only variant. These results do not establish a consistent incremental benefit from diffusion. Paired comparisons across identical seeds and a more controlled ablation are needed to distinguish the gains from class-wise weighting from those attributable to diffusion-based reliability estimation.

  4. The efficiency evaluation does not cover the complete deployment cost. CDRA retains all client models for prediction, whereas the efficiency comparison primarily reports training time, peak GPU memory, and upload size. Equation (26) requires multiple models to predict the same target node, but it remains unclear where these models run, whether they must be downloaded to clients, and who performs fusion when raw graphs remain local. Without inference time, total model storage, and any necessary download costs, the overall resource advantage over methods that produce a single model is difficult to assess.

Reference

[1] Xunkai Li, Zhengyu Wu, Wentao Zhang, Yinlin Zhu, Rong-Hua Li, and Guoren Wang. FedGTA: Topology-aware Averaging for Federated Graph Learning. Proceedings of the VLDB Endowment, 17(1): 41–50, 2023.

Questions for authors
  1. After controlling for class support and client size, does the diffusion margin consistently correlate with class-specific client accuracy on the target test distribution? How does the method handle structurally consistent but inaccurate clients, or structurally different but accurate clients?

  2. Could you report an ablation that sets the diffusion contribution to zero while keeping all other components unchanged, together with paired results against this variant and uniform ensembling? How does FedGTA-style propagated-prediction confidence perform when substituted for the Hellinger margin within the same fusion framework?

  3. Where are the multiple models in Equation (26) executed on the target graph? If graphs remain local, must all models be downloaded to each client? Please clarify the complete communication protocol and report model storage, inference cost, and any necessary download costs as the number of clients increases.

  4. Appendix J acknowledges that the method does not protect against inference attacks. Nevertheless, class-support counts and diffusion signatures may additionally reveal client label distributions and structural properties. What information do these summaries expose beyond model-only uploads, and what analysis or experiments support the stated disclosure boundary?

3 comments

Dear reviewer K6CH,

We sincerely thank you for your thoughtful and constructive evaluation of our work. We address the comments below with additional explanations.

Response to Q1: Diffusion Margins and Class-Specific Predictive Accuracy

Correlation with Predictive Accuracy. As reported in our response to reviewer UWgH (Table R1), we conducted additional experiments on Chameleon (αDir=0.05\alpha_{Dir}=0.05) across three random seeds. The diffusion margin exhibits a positive Spearman correlation with class-specific held-out recall (0.642 ± 0.073). Furthermore, the pooled margin-recall correlation (ρ=0.593\rho=0.593) exceeds the class-support-recall correlation (ρ=0.204\rho=0.204), suggesting that diffusion margins capture information beyond local class frequency. Positive correlations are also observed under the evaluated feature-distribution shifts and local training variations.

Structural Consistency vs. Predictive Accuracy. We acknowledge that structural consistency does not necessarily imply predictive accuracy. A structurally consistent but inaccurate client may receive an overly favorable diffusion score, while a structurally different but accurate client may be downweighted. CDRA mitigates reliance on diffusion alone by incorporating class support and client size in Eq. (21) and blending class-wise and global reliability weights in Eq. (25). These mechanisms cannot completely eliminate such mismatches.

Overall, the results support diffusion margins as a complementary indicator of class-specific predictive reliability, rather than a guaranteed measure of model accuracy. We will clarify this distinction in the final manuscript.

Response to Q2: Zero-Diffusion Ablation and FedGTA-Style Confidence

Zero-Diffusion Ablation. We set λ=0\lambda=0 in Eq. (21), eliminating the diffusion contribution while retaining the class-support prior, client-size prior, reliability normalization, and class-wise/global fusion mechanism.

FedGTA-Style Confidence Substitution. We additionally replace the Hellinger margin with FedGTA-style propagated-prediction confidence. Specifically, each client generates soft predictions using its locally trained GNN, propagates them over its local graph, and computes degree-weighted prediction-entropy confidence. These scores are aggregated class-wise, centered, and normalized before entering the original CDRA reliability equation.

All remaining components, including local training, minimum support threshold, client-size and class-support priors, temperature parameters, and final ensemble fusion, are kept unchanged. This is a controlled substitution of the reliability signal, rather than an implementation of the complete FedGTA framework.

Table R: Test accuracy (%) across three datasets (mean ± SD over three seeds).

MethodCoraChameleonCoauthor-CS
Uniform Ensemble31.42 ± 5.333.59 ± 6.428.31 ± 3.9
CDRA (λ=0\lambda=0)46.57 ± 0.732.90 ± 7.530.62 ± 4.0
Full CDRA (Hellinger)45.95 ± 2.333.48 ± 7.531.12 ± 4.1
CDRA + FedGTA-style Confidence41.90 ± 3.525.21 ± 5.734.97 ± 2.2

Compared with λ=0\lambda=0, Full CDRA changes mean accuracy by −0.62, +0.58, and +0.50 percentage points on Cora, Chameleon, and Coauthor-CS, respectively, indicating that the additional benefit of the Hellinger margin is modest under these settings.

When compared with FedGTA-style confidence, Full CDRA achieves 4.05 and 8.27 percentage-point higher accuracy on Cora and Chameleon, respectively, whereas FedGTA-style confidence outperforms Full CDRA by 3.85 percentage points on Coauthor-CS.

Overall, these results indicate that the effectiveness of reliability signals is dataset-dependent. The Hellinger margin performs better than propagated-prediction confidence on two of the three evaluated datasets, but its contribution beyond class-support and client-size information remains limited. We will clarify this distinction and the associated limitations in the revised manuscript.

Response to Q3: Inference Location, Communication Protocol, and Scalability

Inference Protocol. We follow a centralized evaluation protocol, consistently applied to CDRA and all baselines. Each client trains its model locally and uploads the trained parameters, class-conditioned diffusion signatures, and support counts to the server in a single communication round. The server then executes all KK local models on the held-out benchmark evaluation graph and computes pk(v)p_k(v) for the test nodes. The predictions are combined using the class-wise weights in Eq. (26). Clients do not download the other clients' models.

Model Storage and Inference Cost. Let SθS_\theta denote the size of one local model and TGNNT_{\mathrm{GNN}} the cost of one forward pass on the evaluation graph. Server-side model storage scales as O(KSθ)O(KS_\theta), and ensemble inference requires KK forward passes, giving O(KTGNN)O(KT_{\mathrm{GNN}}) computational complexity. The additional signature storage and weighted prediction aggregation are lightweight. Since inference is performed centrally, client-side model download cost is zero.

For example, on Cora with K=10K=10, the total client-to-server upload is approximately 7.04 MB, dominated by the trained models. With the same architecture, model storage and inference computation increase approximately linearly with KK, while no additional client downloads are required.

Evaluation Assumption. The held-out benchmark evaluation graph is available to the central evaluator. This is distinct from the clients' private training subgraphs and represents the evaluation setting used in our experiments. Deployments in which the target graph is available only on clients would require a different inference arrangement, which is outside the evaluated protocol.

We will clarify these assumptions and the associated communication and inference costs in the final manuscript.

Response to Q4: Information Disclosure and Privacy Boundaries

Information Exposed by Client Summaries. Compared with model-only uploads, CDRA additionally communicates class-support counts and class-conditioned diffusion signatures. The support counts reveal the exact number of local training nodes for classes satisfying nk,c≥τn_{k,c}\geq\tau, while smaller counts are suppressed. The diffusion signatures summarize multi-hop class-mixing patterns and may therefore reveal aggregate structural properties, including class homophily and heterophily. We acknowledge that these summaries introduce additional information exposure beyond model parameters alone.

Disclosure Boundary. CDRA does not transmit raw node features, individual labels, adjacency matrices, or node-level representations. The additional structural payload contains at most HC2HC^2 floating-point values and CC class-support counts per client, independently of graph size. For Cora (H=3H=3, C=7C=7), this corresponds to at most 154 scalar values per client, with fewer values when classes do not meet the support threshold.

Privacy Scope. Our disclosure boundary is based on the information explicitly communicated by the protocol, rather than a formal privacy guarantee. Although the summaries are aggregated and compact, they may still support inference about local label distributions or structural characteristics. We do not claim protection against such inference attacks, and no formal leakage bound or attack-based privacy evaluation is established.

We will clarify these additional disclosures and the distinction between communication efficiency and privacy protection in Appendix J of the final manuscript.

Thank you for the author's additional experiments and clarifications regarding the information that the server can obtain. My main concern remains whether the federated learning method can require the client to disclose additional local information while improving the prediction performance. In the "honest but curious" server scenario set in this paper, the server can directly obtain the categories that have reached the threshold and the exact number of training samples for each category in each client. Even if a certain category does not reach the reporting threshold, as long as its component corresponding to the step diffusion feature in another category is positive, the server can confirm the existence of this category in the local training set and know that there is at least one edge between these two training nodes. Therefore, the basic bottom line of the privacy-preserving federated learning method cannot be achieved by sacrificing privacy to gain performance improvement.

The supplementary ablation experiments by the author are helpful in determining the benefits brought by this additional disclosure. Under the condition of retaining the number of category samples and other aggregation components, after adding the diffusion term, the average accuracy rates on Cora, Chameleon, and Coauthor-CS increased by −0.62, +0.58, and +0.50 percentage points respectively. The change amplitudes on the three datasets are all small and the directions are inconsistent. Given that sharing the diffusion feature will disclose the client's category-level graph structure, and the existing results have not shown that this part of information can bring stable performance improvement, I believe that this paper needs to consider modifying the existing optimization algorithm under the premise of privacy protection. Therefore, I cannot raise my score based on the current technical solution of the paper, and I still maintain a score of 2.

Reviewer_UWgHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CDRA (Class-wise Diffusion Reliability Aggregation), a lightweight framework for one-shot federated graph learning under severe client heterogeneity. Rather than assigning each client a single aggregation weight across all classes, CDRA estimates class-specific client reliability using structural information from local graphs. Specifically, each client independently trains a local GNN and constructs class-conditioned diffusion signatures from multi-hop propagation of training-label indicators. The server builds leave-one-client-out class references and measures the agreement and separation of local signatures using Hellinger distances. The resulting reliability margins, combined with client size and class support, determine class-wise weights for aggregating local model predictions. The framework requires only one client-to-server communication round and avoids synthetic graph generation, proxy-graph reconstruction, and additional server-side model training. Experiments on seven graph datasets report improved classification accuracy and substantially lower computational, memory, and communication costs compared with several existing federated graph learning approaches.

Critical strengths

1.The paper addresses an important problem in one-shot federated graph learning and proposes a lightweight framework that avoids graph reconstruction and additional server-side training. 2.The class-wise aggregation strategy is well motivated, as different clients may have varying expertise across classes under severe non-IID settings. 3.The experiments cover seven graph benchmarks and report promising accuracy and overall efficiency improvements compared with existing methods, particularly GHOST and OASIS.

Critical weaknesses

1.Insufficient validation of class-wise reliability estimation. CDRA assumes that diffusion-based structural reliability margins serve as a proxy for clients' class-specific predictive reliability, but the ablation study does not directly validate this assumption. The authors should provide correlation analyses between reliability margins and actual per-class accuracy, along with controlled experiments under varying feature distributions and model quality. 2.Insufficient evaluation of the performance-cost trade-off. Although CDRA demonstrates substantial efficiency gains over GHOST and OASIS, its ablation study reports only accuracy, with relatively modest improvements over simpler aggregation variants on some datasets. The authors should compare accuracy, runtime, memory, and communication costs across Support Only, Diffusion Only, and Full CDRA to justify the additional complexity. 3.Insufficient evaluation of structural coverage under client heterogeneity. Although Appendix G defines Cov_sig and Cov_LOO and claims to report them, Table 6 only presents accuracy results. The authors should report these metrics under both the main extreme non-IID setting and the client-scaling experiments to clarify how structural coverage affects the effectiveness of diffusion-based reliability estimation.

Questions for authors

1.Can the authors demonstrate that diffusion-based reliability margins correlate with actual per-class prediction accuracy? Does this relationship remain consistent under varying feature distributions and local model quality? 2.Can the authors provide an accuracy–efficiency comparison of Support Only, Diffusion Only, and Full CDRA, including runtime, GPU memory, and communication costs, to justify the additional complexity of diffusion-based reliability estimation? 3.Can the authors report Cov_sig and Cov_LOO under both the main extreme non-IID setting and the client-scaling experiments? How does structural coverage affect the actual performance gains from diffusion-based reliability estimation?

2 comments

Dear reviewer UWgH,

We sincerely appreciate your thoughtful and constructive evaluation of our work. We address each concern below with detailed explanations and additional experimental evidence. All additional experiments and their corresponding analyses will be incorporated into the final version of the paper.

Response to Q1: Correlation Between Diffusion Reliability and Per-Class Prediction Accuracy

To empirically validate the proposed diffusion-based reliability mechanism, we conducted additional experiments on the heterophilous Chameleon benchmark across three random seeds. Specifically, we measured the Spearman rank correlation between the raw diffusion reliability margin mk,cm_{k,c} (Eq. 18) and the corresponding local model's class-specific recall on held-out test nodes. We further evaluated the consistency of this relationship under varying client feature distributions and local training conditions.

Table R1: Spearman correlation (mean ± standard deviation across three seeds) under different experimental conditions.

Experimental settingSpearman ρ (Mean ± SD)
Original setting0.642 ± 0.073
Feature distribution heterogeneity
Mean shift (0.1)0.642 ± 0.073
Mean shift (0.3)0.642 ± 0.073
Mean shift (0.5)0.593 ± 0.131
Local model training variation
10 epochs0.621 ± 0.191
30 epochs0.642 ± 0.073
100 epochs0.642 ± 0.073
Mixed (10/30/100 epochs)0.684 ± 0.081

Correlations are calculated using client–class pairs with valid diffusion margins.

Correlation with predictive accuracy. The original setting exhibits consistently positive correlations across all three seeds, achieving a mean Spearman correlation of 0.642 ± 0.073. Furthermore, the pooled margin–recall correlation (ρ = 0.593) exceeds that of local class support and recall (ρ = 0.204), suggesting that diffusion-based structural agreement captures information not simply reflected by class frequency.

Robustness to feature heterogeneity. Under client-specific feature mean shifts of increasing magnitude, the correlation remains positive and largely stable, ranging from 0.593 to 0.642. This indicates that diffusion-based structural reliability retains its association with class-specific predictive performance under the evaluated feature-distribution perturbations.

Robustness to local training variation. We additionally evaluate different local training budgets (10, 30, and 100 epochs) and heterogeneous client training budgets (10/30/100 epochs). Positive correlations persist across all configurations, with the mixed-budget setting achieving 0.684 ± 0.081. These findings demonstrate that the observed margin–recall association is maintained across the tested differences in local training conditions.

These results provide direct empirical evidence that diffusion reliability margins are positively associated with class-specific predictive performance, with this relationship remaining positive under the evaluated feature-distribution shifts and local training variations. This supports their use as a complementary structural reliability signal for class-wise aggregation in CDRA.

Response to Q2: Accuracy–Efficiency Trade-off of CDRA Components

We provide an additional accuracy–efficiency comparison of Support Only, Diffusion Only, and Full CDRA on the Cora dataset, reporting runtime, peak GPU memory, and total upload communication.

Table 1: Accuracy–efficiency comparison on Cora. Accuracy (%) is reported as mean ± standard deviation across seeds. Runtime (s), peak GPU memory (GB), and upload size (MB) are measured for a fixed experimental run.

MethodAccuracy (%)Time (s) ↓GPU Mem. (GB) ↓Upload (MB) ↓
Support Only42.50 ± 4.513.870.077.04
Diffusion Only43.08 ± 7.314.970.077.04
Full CDRA45.95 ± 2.313.370.077.04

Note: Absolute runtimes may differ from the original paper due to different GPU hardware. All variants here were evaluated under identical conditions.

Accuracy–Efficiency Trade-off. Full CDRA improves mean accuracy by 3.45 percentage points over Support Only and 2.87 percentage points over Diffusion Only, demonstrating the complementary contributions of class support and structural reliability information.

Importantly, these improvements incur negligible additional resource overhead. All three variants exhibit identical reported peak GPU memory (0.07 GB) and upload size (7.04 MB). This is because communication is dominated by the uploaded local GNN models, while CDRA's additional class-conditioned diffusion signatures and support counts are extremely compact. Their contribution to the total payload is negligible at the reported precision.

Similarly, diffusion signature construction and reliability estimation involve lightweight operations rather than additional model training or iterative server optimization. The small runtime differences among variants are not indicative of substantial computational overhead.

Overall, the results indicate that Full CDRA improves accuracy with negligible additional computational and communication overhead.

Response to Q3: Structural Coverage and Its Impact on Predictive Performance

We additionally report the signature coverage CovsigCov_{sig} and leave-one-client-out reference coverage CovLOOCov_{LOO}, as defined in Eqs. (58)–(60), under both the main extreme non-IID setting and the extended client-scaling experiments.

Table R1: Structural coverage under the main extreme non-IID setting αDir=0.01\alpha_{Dir}=0.01, τ=10\tau=10.

ClientsCDRA accuracy (%)CovsigCov_{sig}(%)CovLOOCov_{LOO}(%)
1045.95 ± 2.312.38 ± 2.1819.04 ± 16.49

Table R2: Structural coverage and predictive accuracy under client scaling αDir=5\alpha_{Dir}=5, τ=2\tau=2.

ClientsCDRA accuracy (%)CovsigCov_{sig}(%)CovsigCov_{sig}(%)
576.45 ± 0.491.43 ± 2.60100.00 ± 0.00
1067.13 ± 4.889.99 ± 5.15100.00 ± 0.00
2047.32 ± 0.974.76 ± 2.70100.00 ± 0.00
4051.76 ± 5.750.71 ± 0.71100.00 ± 0.00
8055.97 ± 7.023.57 ± 0.31100.00 ± 0.00
16047.83 ± 6.86.16 ± 0.5095.24 ± 8.25
32036.45 ± 2.60.37 ± 0.0719.05 ± 8.24

Structural coverage under extreme non-IID heterogeneity. As shown in Table R1, the severe label imbalance results in limited signature coverage (12.38%) and independent LOO reference coverage (19.04%). Nevertheless, CDRA achieves 45.95% accuracy. This illustrates that CDRA remains applicable even when structural evidence is sparse, as unavailable diffusion references contribute neutrally while the aggregation mechanism retains class-support and client-size information.

Relationship between coverage and performance. Table R2 reveals a clear decline in signature coverage as the number of clients increases, from 91.43% at 5 clients to just 0.37% at 320 clients. LOO coverage remains at 100% through 80 clients before dropping to 19.05% at 320 clients. This distinction is important: although multiple clients may collectively provide references for each class, the number of individual client–class pairs contributing structural evidence can already be substantially reduced.

Correspondingly, predictive accuracy decreases overall from 76.45% to 36.45% as the federation grows. Although the relationship is not strictly monotonic, these observations are consistent with the finite-support limitation: partitioning a fixed graph among more clients reduces local class support and the availability of informative diffusion-based reliability comparisons.

Contribution to predictive performance. Our existing ablation study (Table 4) further shows that Full CDRA improves over Support-Only aggregation on Cora (45.95% versus 42.50%, a 3.45 percentage-point improvement). These results indicate that structural evidence can complement class-support information, while its availability places practical constraints on reliability estimation.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.