- 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: 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 local training nodes a "class-conditioned diffusion signature" — the class- rows of the row-normalised class-mixing matrices for , i.e. how label mass propagated from class- training nodes is distributed over the classes present on that client (Eqs. 5–8) — and (iii) the class-support counts . 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 and a class-support prior into per-class client weights (Eqs. 21–25), and fuses the local predictions class-wise (Eq. 26). The payload is floats plus 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 clients and Dirichlet label skew 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 ; 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 in a milder partition (Table 6), and sensitivity to the diffusion depth (Fig. 3).
Critical strengths
1. A lightweight, practical design for a setting where existing OFGL methods are heavy. The structural payload is floats plus 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 and , 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; and 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 (about 73% vs. 62% at ), 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 of is the distribution of label mass propagated from class- training nodes over the training labels present on the client. Under , a client that holds (almost) only class obtains for every regardless of its topology, while a client where is a minority spreads mass over its majority classes; the Hellinger margin then largely compares local label marginals, information the server already receives as , 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 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 for a node held by client — this requires every client (or the server) to hold all 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 (; App. B) and the baselines' hyperparameters were chosen, only has a sensitivity study. Without (a)–(d) the results cannot be interpreted or reproduced from the paper alone.
Questions for authors
- 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.
- 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.
- 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.)
- Inference protocol: where are the test nodes, on which graph is computed, and does each client download all models? What are the resulting download size and inference cost as grows?
- How were and the baselines' hyperparameters selected, and which validation split (local or global) is used for checkpoint selection? Please report sensitivity to , and .
- For classes with , is still uploaded and used in Eq. (21), or treated as zero?
- Which dataset does Table 6 use? Which methods appear in Fig. 2(a)/(b), and were they re-tuned for each and ?
- 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
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 is evaluated on the same server-accessible evaluation graph to obtain its class probabilities 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 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 local models, resulting in model-upload volume and server-side storage that scale linearly with
Inference Cost and Scalability. Since the server evaluates all local models on the held-out test graph, the inference complexity is , where 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: , , , , and . 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 , , and 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 . Test accuracy (%), mean ± standard deviation.
| Cora | Chameleon | |
|---|---|---|
| 1 | 47.64 ± 1.1 | 33.11 ± 7.2 |
| 2 | 47.36 ± 0.9 | 32.75 ± 7.2 |
| 5 | 47.02 ± 0.9 | 33.26 ± 7.8 |
| 10 (default) | 45.95 ± 2.3 | 33.48 ± 7.5 |
| 20 | 46.59 ± 3.0 | 32.57 ± 6.8 |
Table 2: Sensitivity to the class-wise blending coefficient . Test accuracy (%), mean ± standard deviation.
| Cora | Chameleon | |
|---|---|---|
| 0.0 | 30.90 ± 0.5 | 31.88 ± 2.0 |
| 0.2 | 34.99 ± 3.5 | 33.22 ± 2.2 |
| 0.5 | 43.90 ± 8.5 | 33.88 ± 6.4 |
| 0.8 (default) | 45.95 ± 2.3 | 33.48 ± 7.5 |
| 1.0 | 46.72 ± 4.1 | 33.08 ± 7.3 |
Table 3: Sensitivity to the structural reliability coefficient . Test accuracy (%), mean ± standard deviation.
| Cora | Chameleon | |
|---|---|---|
| 0.0 | 46.57 ± 0.7 | 32.90 ± 7.5 |
| 0.5 | 46.10 ± 1.6 | 33.33 ± 7.5 |
| 1.0 (default) | 45.95 ± 2.3 | 33.48 ± 7.5 |
| 1.5 | 45.25 ± 3.1 | 32.71 ± 7.1 |
| 2.0 | 45.07 ± 4.0 | 33.37 ± 6.8 |
Observations. Performance remains relatively stable across the evaluated and values, while 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 , 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., .
Only classes satisfying 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 and minimum support threshold .
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 is varied, while in Figure 2(b), only the number of clients 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.
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).
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.
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=0.046), whereas the remaining comparisons do not meet the conventional p<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:
The results (mean ± standard deviation over three different seeds, accuracy in %) are:
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.