acceptodds
Under review as a conference paper at ICLR 2027

Distribution-Calibrated Diffusion for One-Shot Federated Graph Classification

Abstract

Federated learning enables multiple clients to collaboratively train models without sharing their private data. However, conventional approaches typically require repeated client–server communication, leading to substantial communication overhead and repeated exposure of client model parameters. One-shot federated learning has therefore attracted increasing attention by restricting client–server interaction to a single communication round. Recently, this paradigm has been extended from image data to graph data, giving rise to one-shot federated graph learning. However, existing one-shot federated graph learning studies mainly focus on node-level tasks, while one-shot learning for graph-level classification remains largely unexplored. Under strict one-shot communication, the server must reconstruct reliable graph distributions and reconcile heterogeneous client data without any subsequent client feedback. To bridge this gap, we propose a one-shot federated graph classification framework that leverages diffusion models to transfer and reconstruct client graph distributions. Specifically, clients upload diffusion-based generative knowledge and compact graph distribution summaries only once, after which the server generates, validates, and reweights synthetic graphs to reconstruct and fuse client graph distributions. Extensive experiments demonstrate that our framework consistently outperforms state-of-the-art FGL and one-shot FGL baselines.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
One-Shot Federated Learning, Federated Graph Classification, Graph Diffusion Models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_ppZKOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes OSFGC, a one-shot federated graph classification framework that enables multiple clients to collaboratively train a global graph classifier. The core pipeline is: client trains a class and size conditioned graph diffusion model to generate graphs and constructs a Graph Distribution Sketch that includes class-conditional structural statistics, histograms, and random-feature moments of its local graph data. Both the generator and GDS are uploaded to the server. Then, the server generates synthetic graphs using the uploaded diffusion models, screens candidates based on structural consistency and diversity, and employs CR-KMM to assign weights to the retained samples. Finally, a global classifier is trained on the weighted synthetic dataset, with an additional GDS-based predictive calibration mechanism at inference time.

Critical strengths
  1. The problem itself is meaningful.
  2. Promising empirical results across many heterogeneous graph classification settings. The paper evaluates the proposed approach on eight federated graph classification settings, including both single-dataset and cross-dataset scenarios. The reported improvements, particularly in Macro-F1, are encouraging. The authors also provide several component ablations and analyses of structural alignment, which offer preliminary evidence that the proposed mechanisms can improve classification performance.
  3. The weighting mechanism for each generated is reasonable and is believed to better match the original data distribution.
Critical weaknesses
  1. The whole pipeline consists of many components and the necessity of some is not clear. For example, the uploaded GDS contains rich class-conditional information, including structural histograms, descriptor means, second moments, and class counts. The predictive calibration mechanism in Eq. (29) already defines a classifier-like scoring function based directly on these statistics. It’s not sure how well a classifier would use only the uploaded GDS perform. Is the other component necessary?
  2. The paper does not quantify the communication costs of the proposed method. Although OSFGC requires only one communication round, each client must upload a diffusion generator together with a collection of class-conditional distribution statistics. A single communication round does not necessarily imply low communication overhead. And high communication costs would trade the benefit of the one-shot setting.
  3. The training effectiveness of graph diffusion models under extreme label-distribution skews or situations where each client has very limited local training data is not clear. Especially when the proposed method chooses to train a generator with fine-grained control.
  4. The performance and robustness of proposed method under privacy protection strategy is not provided. Each client uploads both a locally trained diffusion generator and class-conditional GDS statistics. These statistics contain information about local class distributions, graph sizes, node attributes, connectivity patterns, and other structural properties. Together with the generator parameters, they may expose information about the underlying private training graphs. The Appendix discussed optional Gaussian perturbation and DP-SGD, but how the proposed method perform under such protection is not clear.
  5. Some one-shot federated learning baselines (e.g., DENSE, Co-Boost) should be included, if applicable. They share a similar paradigm that clients train a data generator and train a classifier on the server.
Questions for authors
  1. Some implementation choices require clarification. Appendix B.1 describes the server-side classifier as an MLP, whereas several baselines use GIN-based classifiers. The paper does not clearly specify how variable-sized graphs are converted into the MLP input representation.
  2. How effective is direct classification using only the uploaded GDS (or the variable derived from it)?
  3. The framework introduces several hyperparameters across candidate screening, distribution matching, and predictive calibration, including λs\lambda_s, λd\lambda_d, λ2\lambda_2, λr\lambda_r, λe\lambda_e, ρ\rho , and λcal\lambda_{\mathrm{cal}}. The selection and sensitiveness of these hyper-parameters need more illustration.
4 comments
AuthorsOpenReview ↗

Response to Reviewer ppZK (1/4)

We thank the reviewer for the valuable and constructive comments. We provide direct component comparisons, communication measurements, stress tests, privacy-noise experiments, and additional one-shot baselines below.


1. Is Classification from GDS Alone Sufficient? (W1, Q2)

Direct Comparison: We evaluated three prediction rules: GDS-only, classifying solely from the uploaded GDS; Classifier-only, using the classifier trained on CR-KMM-weighted synthetic graphs with predictive calibration disabled; and Full OSFGC, combining both. Mean ACC and Macro-F1 in percent over three repeated runs are reported below.

SettingMetric (%)GDS-onlyClassifier-onlyStrong shared-head baseline (by ACC, Table 1)Full OSFGC
PROTEINSACC64.1972.4171.55 (Co-Boost)76.68
PROTEINSMacro-F148.9366.0064.84 (Co-Boost)73.36
BIO-SN-CVACC31.7533.0630.62 (MOON)49.97
BIO-SN-CVMacro-F121.9519.2317.75 (MOON)42.88

Results:

  • The synthetic-graph pipeline alone outperforms comparable one-shot baselines: Even with test-time GDS calibration disabled (λcal=0\lambda_{\mathrm{cal}}=0), Classifier-only outperforms all baselines with reported results on PROTEINS in Table 1 in both mean ACC and Macro-F1, including Co-Boost (71.55 / 64.84) and DENSE (63.28 / 45.78). On BIO-SN-CV, its ACC of 33.06% exceeds that of every global shared-head and native one-shot baseline with a reported result, including MOON (30.62%), GCFL+ (28.81%), FedAvg (28.62%), DENSE (24.14%), Co-Boost (23.20%), GHOST (16.97%), and OASIS (13.99%). Only FedStar reports higher ACC (47.68%) on this dataset. It retains client-specific classification heads.
  • The two branches complement each other: Classifier-only learns instance-level decision boundaries from synthetic graphs. On BIO-SN-CV, nine clients across three domains share class IDs whose meanings can differ across datasets. GDS calibration provides distributional evidence that can suppress implausible out-of-domain classes. Combining the branches raises ACC / Macro-F1 from 33.06 / 19.23 for Classifier-only to 49.97 / 42.88 for Full OSFGC, compared with 31.75 / 21.95 for GDS-only.
  • Contribution of generation and fusion: In Figure 5 of the main paper, removing each of the following components lowers both metrics even when GDS predictive calibration is retained. Mean ACC / Macro-F1 on PROTEINS and BIO-SN-CV, respectively, are: w/o DG (62.75 / 39.41 and 30.63 / 17.36), w/o D-GDS-G (72.19 / 65.11 and 33.63 / 18.21), w/o CS (73.78 / 70.61 and 36.61 / 27.13), and w/o CR-KMM (73.30 / 67.85 and 34.22 / 21.64). These results show that each stage contributes to performance in the tested pipeline.

2. Measured Communication Payload. (W2)

We measured OSFGC's complete upload files, including generator parameters, GDS, and metadata. For CeFGC and OASIS, we count float32 parameter bytes in the uploaded objects received by their official server implementations: CeFGC uploads one label-conditioned EDP-GNN diffusion generator per client, using its communication-saving variant, while each OASIS client uploads a synthesizer and a teacher model with a codebook. All values below are summed across clients and reported in MiB. The OSFGC generator and GDS columns count float32 tensor bytes. Metadata and file-format overhead is estimated as the total upload minus the generator and GDS values, using the rounded figures shown.

DatasetsCeFGCOASISOSFGC TotalOSFGC GeneratorOSFGC GDSOSFGC Metadata & File-format Overhead
PTC_MR1.100.810.670.600.020.05
NCI113.099.076.996.330.190.47
PROTEINS3.492.442.161.940.060.16
DD7.153.922.592.370.070.15
CHEM3.562.311.711.530.070.11
BIO-SN-CV63.669.864.463.930.310.22
BIO-CHEM7.153.952.612.370.080.16
BIO-CHEM-SN-CV113.1817.537.696.980.380.33

Results: OSFGC has a smaller upload payload than CeFGC and OASIS in all eight settings. GDS accounts for only 0.02–0.38 MiB across all clients, approximately 3%–7% of OSFGC's uploaded tensor bytes, with generator parameters accounting for the remainder. The gap to CeFGC is largest in cross-domain settings, where its generator size grows substantially with node-feature dimensionality.

AuthorsOpenReview ↗

Response to Reviewer ppZK (2/4)

In the following, we clarify weaknesses 3 and 4.


3. Generator Utility with Scarce Data and Severe Label Skew. (W3)

Protocol: We retrained local generators after reducing or redistributing PROTEINS training data, retaining the original validation and test partitions. Predictive calibration is disabled, so the classifier uses only the learned predictions from synthetic training graphs. This directly tests whether the generated data retain useful class information. Results from a single run are reported below.

Training conditionACC (%)Macro-F1 (%)
100% original training data (reference)73.5368.87
Retain 25% of each client's training data71.9364.65
Retain 10% of each client's training data65.8648.09
Dirichlet label skew (α = 1)71.8566.52
Dirichlet label skew (α = 0.1)71.7867.83
Dirichlet label skew (α = 0.03)68.3052.35
At most 10 graphs per class per client71.8566.15
At most 5 graphs per class per client71.7069.77
α = 0.1 and retain 25%66.4765.42

Results: All eight scarcity and label-skew conditions exceed the majority-class ACC reference of 57.8%, demonstrating useful class information in the generated training set under limited local data or severe label imbalance.

Why It Remains Effective: For scarce data: Fine-grained structural control is applied during generation and subsequent candidate screening using empirical histogram guidance, without requiring the denoiser to learn additional complex control parameters during local training. For label skew: Each generator models its locally supported classes, and the server combines their synthetic samples without requiring every client to cover the global label space.


4. Performance under Privacy-oriented Perturbations. (W4)

Protection Tested: We jointly perturb GDS statistics with Gaussian noise and train the diffusion generator using DP-SGD, clipping each graph's gradient to an ℓ2\ell_2 norm of at most C=1.0C=1.0 before adding Gaussian noise with standard deviation σC\sigma C, where σ\sigma is the noise multiplier. GDS perturbations cover histograms, counts, moments, and reliability scores. We add Gaussian noise independently to each component according to its upload scale. We use σ∈{0.5,0.6,0.75,1.0}\sigma\in\{0.5, 0.6, 0.75, 1.0\}, progressively increasing the noise strength. The differential privacy budget ε\varepsilon quantifies privacy protection: smaller values provide stronger protection against inferring from the released generator parameters whether an individual graph participated in training. The reported generator-only ε\varepsilon estimates decrease from 10.0 to 4.4 (δ=10−5\delta=10^{-5}, with neighboring datasets differing by the addition or removal of one graph), within the approximately 1–10 range commonly studied in DP-SGD research. Values below are three-seed mean ACC / Macro-F1 (%); the no-noise reference is the original Table 1 result.

Generator σGenerator ε (reported estimate)PROTEINSBIO-SN-CV
0 (no-noise reference)—76.68 / 73.3649.97 / 42.88
0.5010.076.43 / 73.0649.78 / 42.30
0.608.075.91 / 72.3947.38 / 39.82
0.756.175.96 / 71.9546.70 / 38.94
1.004.475.16 / 70.2744.59 / 36.60

Results: The method retains substantial predictive utility under simultaneous perturbations of both uploaded information sources, with greater degradation under stronger noise and cross-domain heterogeneity.

AuthorsOpenReview ↗

Response to Reviewer ppZK (3/4)

In the following, we address weakness 5 on the additional one-shot baselines.


5. Additional DENSE and Co-Boost Baselines. (W5)

We added DENSE and Co-Boost with GIN teachers trained on local original graphs and uploaded once, followed by server-side soft-graph generation and GIN student training, using the same GIN backbone as FedAvg, FedProx, and GCFL+ to accommodate graph data. Their core algorithms directly call the official implementations, and training hyperparameters follow the official configurations. The following results are mean ± standard deviation (%), with OSFGC results taken directly from the original Table 1.

SettingDENSE (ACC / Macro-F1)Co-Boost (ACC / Macro-F1)OSFGC (ACC / Macro-F1)
PTC_MR57.07 ± 0.00 / 36.33 ± 0.0052.78 ± 4.29 / 34.48 ± 1.8468.18 ± 0.32 / 64.76 ± 0.25
NCI159.16 ± 6.31 / 48.73 ± 12.4765.13 ± 1.38 / 64.01 ± 1.6367.37 ± 0.18 / 66.45 ± 0.15
PROTEINS63.28 ± 4.19 / 45.78 ± 5.5771.55 ± 2.83 / 64.84 ± 0.3076.68 ± 0.01 / 73.36 ± 0.02
DD50.00 ± 10.57 / 32.94 ± 4.7361.56 ± 0.99 / 45.75 ± 8.0765.58 ± 0.23 / 56.99 ± 0.20
CHEM54.53 ± 3.07 / 40.32 ± 2.9355.60 ± 1.22 / 42.07 ± 1.1974.44 ± 0.10 / 65.24 ± 0.05
BIO-SN-CV24.14 ± 0.00 / 13.81 ± 0.0023.20 ± 9.05 / 12.12 ± 5.0349.97 ± 0.35 / 42.88 ± 0.35
BIO-CHEM54.62 ± 0.22 / 42.12 ± 0.5951.92 ± 4.08 / 38.17 ± 4.9666.11 ± 0.11 / 59.12 ± 0.22
BIO-CHEM-SN-CV34.81 ± 0.00 / 21.50 ± 0.0041.15 ± 0.48 / 25.84 ± 1.8456.24 ± 2.02 / 50.13 ± 0.77

Results: OSFGC achieves higher mean ACC and Macro-F1 than both DENSE and Co-Boost in all eight settings.

AuthorsOpenReview ↗

Response to Reviewer ppZK (4/4)

In the following, we address questions 1 and 3 on graph inputs and hyperparameters.


6. Server MLP and Variable-size Graph Inputs. (Q1)

Input Representation: Each graph, whether synthetic or a real test graph, is first mapped by the same deterministic function to a fixed-length graph-level descriptor. This vector consists of (i) bounded structural statistics, including normalized graph size, edge density, degree statistics, clustering, connectivity, triangle density, and low-order spectral moments of the normalized Laplacian; and (ii) channel-wise means and standard deviations of the selected node features, with node features first zero-padded to a common width across datasets. Graphs of different sizes therefore receive the same form of representation before MLP classification.

Design Rationale: (i) The server trains not on a massive collection of real graphs, but on a limited number of screened synthetic graphs generated for each eligible class by each client's generator and assigned fractional weights by CR-KMM, while cross-domain zero-padding brings node-feature dimensionality to nearly 500. Training a deep message-passing GNN from scratch on this limited synthetic pool could overfit sampling artifacts, including discrete edge perturbations. Mapping these graphs to bounded graph-level statistics consistent with GDS and training a lightweight MLP is therefore intended to limit overfitting. (ii) The w/o DG ablation uses exactly the same descriptor inputs and MLP architecture and retains GDS predictive calibration, but replaces conditional diffusion with GDS-based random graph sampling. As shown in Figure 5 of the main paper, retaining the same MLP architecture while removing only diffusion reduces PROTEINS ACC from 76.68% to 62.75% and Macro-F1 from 73.36% to 39.41%, while BIO-SN-CV ACC falls from 49.97% to 30.63%. Because the representation and classifier are held fixed, these gains cannot be attributed to an advantage from the descriptor inputs or MLP architecture. Instead, they support the contribution of conditional graph diffusion to the utility of the synthetic graphs, consistent with better preservation of joint structural and attribute information.


7. Hyperparameter Configuration and Sensitivity. (Q3)

Thank you for the valuable comment. Across all eight benchmarks, we fix the candidate screening and CR-KMM weights without per-dataset tuning: λs=1.0\lambda_s=1.0, λd=0.2\lambda_d=0.2, λ2=0.25\lambda_2=0.25, λr=0.05\lambda_r=0.05, λe=0.001\lambda_e=0.001, and ρ=0\rho=0. The default settings are amax⁡=0.2a_{\max}=0.2 and vmin⁡=0.002v_{\min}=0.002. We use λcal∈{5,10}\lambda_{\mathrm{cal}}\in\{5,10\} in the main experiments. Figure 6 reports sensitivity over [0,25][0,25]. A value of λcal≥5\lambda_{\mathrm{cal}}\geq 5 does not by itself mean that predictions rely on GDS alone: (i) in prototype calibration (Appendix A.5), λcal\lambda_{\mathrm{cal}} scales standardized GDS scores before they are added to unstandardized classifier logits, so its effect depends on the score differences between competing classes. (ii) In cross-domain calibration, both the classifier and GDS scores are standardized. Very low GDS scores can suppress implausible out-of-domain classes while leaving the classifier to distinguish among plausible alternatives. At a fixed variance floor, Figure 6 shows little change in ACC or Macro-F1 over λcal∈[5,25]\lambda_{\mathrm{cal}}\in[5,25]. (iii) Full OSFGC outperforms GDS-only in both metrics on both datasets (Table 2). Even with calibration disabled (λcal=0\lambda_{\mathrm{cal}}=0), Classifier-only achieves higher mean ACC and Macro-F1 on both PROTEINS and BIO-SN-CV (Table 2) than the four global shared-head one-shot baselines DENSE, Co-Boost, GHOST, and OASIS (Table 1).


Finally, thanks again for your time and helpful comments. We hope your concerns are well addressed. For any other questions, we are happy to discuss them and make further clarifications.

Reviewer_wSg7OpenReview ↗
Rating
4
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes a framework, OSFGC, to increase accuracy in strict one-shot federated learning settings. To achieve this, the authors have clients upload a graph diffusion generator and a graph distribution sketch. The server first generates synthetic graphs, then selects, weights, and fuses them with statistical features from the real data. It then trains a graph classifier. The authors use 16 public graph classification datasets, which are combined into 8 federated settings covering chemistry, bioinformatics, social networks, and computer vision to prove the high performance.

Critical strengths

This paper could make a significant algorithmic contribution for the following reasons:

The proposed method effectively improves the accuracy of strict one-shot federated graph classification. The method is novel because it combines graph generation, structure-guided verification, and constrained distribution matching.

Critical weaknesses

None

Questions for authors

If the authors clarify the experiments in their response, I would be willing to increase the score.

The following issues should be addressed in the experiments:

  1. Do the same class IDs in different datasets represent the same semantics? If not, what learning task does sharing these class labels actually correspond to? This is a substantive question that affects how the cross-domain results should be interpreted.
  2. The method uses the generator’s training reconstruction loss to construct a reliability score and introduces the parameter (\rho) to control its mixture with uniform client weighting. However, Appendix A.3 states that the default is (\rho=0), meaning that each client has equal target mass within the same class. If the main experiments indeed use this default value, the reliability score would not affect client–class mass. It would be helpful to clarify which value of (\rho) was used in the experiments.
3 comments
AuthorsOpenReview ↗

Response to Reviewer wSg7

We appreciate the reviewer's support and constructive feedback. Thank you for highlighting the importance of interpreting the shared-label task and the actual weighting configuration correctly. We address these two questions below.


1. Meaning of Shared Class IDs across Datasets. (Q1)

No, identical class IDs across datasets do not necessarily have the same semantics. In our cross-dataset and cross-domain settings, each dataset retains its original class IDs while sharing a common output space.

Task Interpretation: This is joint classification of heterogeneous tasks through a shared output layer, meaning that the model predicts a graph's class ID within its source dataset. This requires the classifier to implicitly distinguish the relevant domain from graph characteristics and then discriminate among classes within that domain. Differences in graph structure and attributes, such as graph size and connectivity patterns in molecular versus social-network graphs, can provide cues for this implicit domain distinction. At test time, each client evaluates only on its own test graphs using its own labels, so the meaning of classification accuracy remains well defined within each client. The cross-domain results therefore measure how well one global model serves participating clients with heterogeneous tasks and label semantics.

Following the GCFL/FedStar data-partitioning setup, we assign one dataset to each client. Under the strict one-shot protocol, the server trains one global model without subsequent downloads during training or local fine-tuning. A shared classification head serves all clients in this protocol. OSFGC and the evaluated native one-shot baselines train and evaluate in the same shared label space using a single shared classification head. In contrast, methods originally designed for multi-round learning with local classification heads preserve their official heads, avoiding cross-dataset label conflicts at the output layer. Accordingly, the shared-label protocol is applied consistently across one-shot comparisons and does not favor OSFGC over these local-head baselines: OSFGC must additionally accommodate heterogeneous label semantics within a single shared head.


2. Value of ρ\rho and the Role of Reliability. (Q2)

Thank you for the valuable question. The setting ρ=0\rho=0 is used in all eight main experiments, consistent with Appendix A.3. As you noted, the diffusion-loss reliability scores therefore do not affect client–class total weights in these experiments. To maintain client coverage while allowing generation quality to influence client contributions, Eq. (26) in revised Appendix A.3 provides a general scheme combining equal within-class client weights with reliability-based weights. In our implementation, we uniformly set ρ=0\rho=0, as described in Appendix A.3, to keep the method simple and reduce hyperparameter tuning. Positive ρ\rho remains an optional extension.

Clarification in the Revised Manuscript: We appreciate you highlighting this point, which helped us make this distinction clearer. To prevent misunderstanding, we have revised Section 2.4 to describe the implemented equal within-class client allocation βk,c=1/∣Kc∣\beta_{k,c}=1/|\mathcal K_c| directly. The loss-based weighting formulation with ρ\rho has been relocated to Appendix A.3 as an optional extension.

CR-KMM Quality Control Remains Active: With this choice, each eligible client receives an equal share within class cc, namely βk,c=1/∣Kc∣\beta_{k,c}=1/|\mathcal K_c|. The distinction is that CR-KMM optimizes individual synthetic-example weights ac,ja_{c,j}, rather than within-class client total weights βk,c\beta_{k,c}. The equality ∑j:oj=kac,j=βk,c\sum_{j:o_j=k}a_{c,j}=\beta_{k,c} only preserves each group's assigned total, leaving its allocation among synthetic graphs to be learned. In Eq. (11), the first two terms guide this allocation by matching weighted synthetic means and second moments to targets derived from the uploaded original-data statistics. CR-KMM therefore continues to learn how much each synthetic graph contributes, reducing statistical discrepancies even when the within-class client shares are fixed.


Finally, thanks again for your time and helpful comments. We hope your concerns are well addressed. For any other questions, we are happy to discuss them and make further clarifications.

Thank you for the detailed response. It addresses my questions.

Thank you again for your time and encouraging feedback. We are glad that our responses and manuscript revisions have addressed your questions.

Reviewer_5Hh7OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework for one-shot federated graph classification. In the proposed method, each client trains a conditional graph diffusion model and extracts graph distribution summaries, which are uploaded to the server in a single communication round. The server then generates synthetic graphs, evaluates their structural quality, and performs distribution matching to train a global classifier. Experiments across eight federated settings demonstrate promising performance in terms of accuracy and Macro-F1.

Critical strengths

The paper is generally well organized, with a clear problem formulation and detailed descriptions of the main components.

The paper addresses the important problem of one-shot federated graph classification, which has practical significance.

The proposed method utilizes structural screening and distribution matching to improve the quality of generated graphs, which provides a reasonable approach to integrating heterogeneous client data.

Critical weaknesses

Equation (1) defines the objective using client weights proportional to their data sizes, whereas the final training objective employs different class and client weighting schemes. These weights do not necessarily represent the original data distribution. The authors should better justify this difference and discuss its implications for global learning.

CR-KMM estimates generation reliability using local diffusion training losses, which may not be directly comparable across heterogeneous clients or accurately reflect generation quality. Moreover, the default setting of rou = 0 removes the influence of these reliability scores on client weighting.

Several multi-round federated learning methods are restricted to a single communication round, while GHOST and OASIS are adapted from node classification to graph classification. These modifications may weaken the baselines and exaggerate the reported improvements.

The ablation study does not fully isolate the contributions of GDS and CR-KMM. Since GDS is involved in multiple stages, it remains unclear which components are responsible for the improvements.

In the cross-dataset experiments, class indices from different datasets are directly mapped into a shared label space. However, identical class indices may represent different meanings across datasets.

Questions for authors

Equation (1) defines the objective using client weights proportional to their data sizes, whereas the final training objective employs different class and client weighting schemes. These weights do not necessarily represent the original data distribution. The authors should better justify this difference and discuss its implications for global learning.

CR-KMM estimates generation reliability using local diffusion training losses, which may not be directly comparable across heterogeneous clients or accurately reflect generation quality. Moreover, the default setting of rou = 0 removes the influence of these reliability scores on client weighting.

Several multi-round federated learning methods are restricted to a single communication round, while GHOST and OASIS are adapted from node classification to graph classification. These modifications may weaken the baselines and exaggerate the reported improvements.

The ablation study does not fully isolate the contributions of GDS and CR-KMM. Since GDS is involved in multiple stages, it remains unclear which components are responsible for the improvements.

In the cross-dataset experiments, class indices from different datasets are directly mapped into a shared label space. However, identical class indices may represent different meanings across datasets.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 5Hh7 (1/4)

We thank the reviewer for the careful assessment and constructive suggestions. In this part, we address weakness 1 and question 1 concerning the relationship between the population objective and synthetic training weights.


1. Relationship between the Population Objective and Synthetic Training Weights. (W1, Q1)

Thank you for the valuable comment. Our synthetic-training objective indeed differs from the conventional sample-size-weighted population objective in Eq. (1). It uses three levels of weighting: class proportions from the original data, client shares, and synthetic-example weights learned from original-data statistics to correct generation bias. Thus, it reflects the original data distribution through both class proportions and descriptor statistics. Specifically, the training weight of a synthetic graph from client kk and class cc decomposes as

wc,j=γc⏟class total weight×βk,c⏟within-class client total weight×ac,jβk,c⏟within-client-group weight allocation,k=oj.\displaystyle w_{c,j}=\underbrace{\gamma_c}_{\text{class total weight}}\times\underbrace{\beta_{k,c}}_{\text{within-class client total weight}}\times\underbrace{\frac{a_{c,j}}{\beta_{k,c}}}_{\text{within-client-group weight allocation}},\qquad k=o_j.

Equation (11) constrains ∑j:oj=kac,j=βk,c\sum_{j:o_j=k}a_{c,j}=\beta_{k,c}, so the last factor allocates weight within each client–class group. The expression simplifies to wc,j=γcac,jw_{c,j}=\gamma_c a_{c,j}, as in Eq. (12).

The three factors have distinct roles:

  • Class total weight γc\gamma_c follows the class counts from clients' original data: γc=∑kN~k,c/∑c′∈Yact∑kN~k,c′\gamma_c=\sum_k\tilde N_{k,c}/\sum_{c'\in\mathcal Y_{\mathrm{act}}}\sum_k\tilde N_{k,c'}. This is the sum of uploaded class-cc counts divided by the total count over active classes, so class proportions in the training objective come directly from the original data.
  • Within-class client total weight βk,c\beta_{k,c} is assigned equally among clients with data from the corresponding class in our implementation to preserve distinct distributions on smaller clients that sample-count weighting could underrepresent under heterogeneity. This does not imply equal overall client weights: client kk receives ∑c:k∈Kcγc/∣Kc∣\sum_{c:k\in\mathcal K_c}\gamma_c/|\mathcal K_c|, determined jointly by the original data's global class proportions and the classes represented on that client; total weights can therefore differ across clients.
  • Within-client-group weight allocation is obtained by optimizing ac,ja_{c,j} in Eq. (11), subject to the group-total constraint above. CR-KMM brings the weighted synthetic random-feature means and second moments close to mcfm_c^f and vcfv_c^f, the mixtures of uploaded original-data statistics defined using βk,c\beta_{k,c} in Eq. (10). It thus uses original-data distributional statistics to assign synthetic-example weights, bringing the weighted synthetic data closer to the original data in these statistics.

Effect on Global Learning: We replaced equal within-class client weights with count-proportional weights while retaining the remaining settings. On BIO-SN-CV, count-proportional weighting reduces mean client ACC by 0.79 percentage points and worst-client ACC by 2.30 points. This supports the chosen weighting for heterogeneous clients.

AuthorsOpenReview ↗

Response to Reviewer 5Hh7 (2/4)

In the following, we address weakness 2 and question 2 on reliability weighting.


2. Reliability Scores and the Actual Choice of ρ\rho. (W2, Q2)

Thank you for the valuable comment. All eight main experiments use ρ=0\rho=0, as reported in the paper. To preserve client coverage while allowing generation quality to influence client contributions, Eq. (26) in revised Appendix A.3 provides a general scheme mixing equal within-class client weights with reliability-based weights. Reconstruction losses are averaged over nodes, edges, and local class examples. Clipped inverse-loss scores are normalized across active clients within each class. Because reconstruction losses can still differ in scale across heterogeneous clients, we set ρ=0\rho=0 in all main experiments to keep the method simple and avoid extra hyperparameter tuning (Appendix A.3). When local loss scales are comparable, ρ>0\rho>0 remains an optional extension.

Clarification in the Revised Manuscript: We appreciate your careful reading and your comment, which helped us make this distinction clearer. To avoid misunderstanding, revised Section 2.4 now directly presents the equal within-class client allocation βk,c=1/∣Kc∣\beta_{k,c}=1/|\mathcal K_c| actually used in our experiments. The loss-based weighting formulation involving ρ\rho is described in Appendix A.3 as an optional extension.

CR-KMM Quality Control Remains Active: Setting ρ=0\rho=0 makes client total weights uniform within each class, i.e., βk,c=1/∣Kc∣\beta_{k,c}=1/|\mathcal K_c|. Notably, CR-KMM optimizes individual synthetic-example weights ac,ja_{c,j}, rather than within-class client total weights βk,c\beta_{k,c}. The latter are held fixed, with ∑j:oj=kac,j=βk,c\sum_{j:o_j=k}a_{c,j}=\beta_{k,c} constraining each group's total weight. Specifically, the first two terms in Eq. (11) optimize weights ac,ja_{c,j} so that the weighted synthetic means and second moments approach the targets constructed from uploaded original-data statistics. Thus, even with fixed within-class client shares, CR-KMM still adjusts the contributions of individual synthetic graphs as usual to reduce statistical discrepancies from the original data.

AuthorsOpenReview ↗

Response to Reviewer 5Hh7 (3/4)

In the following, we address weakness 3 and question 3 on baseline comparisons.


3. Baseline Protocol and Additional One-Shot Comparisons. (W3, Q3)

CeFGC is our closest counterpart and, among the one-shot FGL baselines considered, the only one targeting graph classification, yet its original protocol still requires three communication rounds. We therefore adapt multi-round federated learning methods and one-shot FGL methods designed for node classification to enable comparisons under the same strict one-shot graph-classification setting.

Adaptation Based on Official Settings: Starting from their official implementations, we restrict FedAvg, FedProx, MOON, FedOpt, GCFL+, FedStar, and FedSSP to one communication round with 50 local epochs and one upload per client. For GHOST and OASIS, we retain their official generation, alignment, and distillation pipelines and replace only node-level prediction with graph pooling and graph-level supervision. CeFGC is evaluated using its official generative one-shot variant.

Additional Training-Budget Comparison: To examine whether limited local training weakens the one-shot baselines, we compare FedAvg, GCFL+, FedStar, and FedSSP on PROTEINS under three configurations: one communication round with 50 local epochs (Table 1), one communication round with 200 local epochs, and multi-round federated learning with 200 communication rounds and one local epoch per round. All three use the same data split and GIN settings, with results averaged over three repeated runs. Mean ACC / Macro-F1 (%) are:

MethodOne-shot (1 round × 50 local epochs, Table 1)One-shot (1 round × 200 local epochs)Multi-round (200 rounds × 1 local epoch)
FedAvg60.26 / 49.4356.02 / 48.8474.59 / 67.57
GCFL+63.14 / 55.4357.21 / 50.1476.77 / 70.91
FedStar62.16 / 50.5664.19 / 57.7373.75 / 67.21
FedSSP65.45 / 49.3765.15 / 57.0373.14 / 64.97
OSFGC (original Table 1 configuration)76.68‾\underline{76.68} / 73.36——

Increasing local training from 50 to 200 epochs does not consistently improve the one-shot baselines, with both ACC and Macro-F1 declining for FedAvg and GCFL+. These results support the reasonableness of the one-round, 50-local-epoch baseline setting used in our comparison. With one upload per client, OSFGC also achieves ACC comparable to the strongest multi-round result and higher Macro-F1 than all four multi-round baselines on PROTEINS.

Additional Baseline Experiments: We added DENSE and Co-Boost with GIN teachers trained on local original graphs and uploaded once, followed by server-side soft-graph generation and GIN student training, using the same GIN backbone as FedAvg, FedProx, and GCFL+ to accommodate graph data. Their core algorithms directly call the official implementations, and training hyperparameters follow the official configurations. The following ACC / Macro-F1 results are mean ± standard deviation (%), with OSFGC results taken directly from the original Table 1.

SettingDENSE (ACC / Macro-F1)Co-Boost (ACC / Macro-F1)OSFGC (ACC / Macro-F1)
PTC_MR57.07 ± 0.00 / 36.33 ± 0.0052.78 ± 4.29 / 34.48 ± 1.8468.18 ± 0.32 / 64.76 ± 0.25
NCI159.16 ± 6.31 / 48.73 ± 12.4765.13 ± 1.38 / 64.01 ± 1.6367.37 ± 0.18 / 66.45 ± 0.15
PROTEINS63.28 ± 4.19 / 45.78 ± 5.5771.55 ± 2.83 / 64.84 ± 0.3076.68 ± 0.01 / 73.36 ± 0.02
DD50.00 ± 10.57 / 32.94 ± 4.7361.56 ± 0.99 / 45.75 ± 8.0765.58 ± 0.23 / 56.99 ± 0.20
CHEM54.53 ± 3.07 / 40.32 ± 2.9355.60 ± 1.22 / 42.07 ± 1.1974.44 ± 0.10 / 65.24 ± 0.05
BIO-SN-CV24.14 ± 0.00 / 13.81 ± 0.0023.20 ± 9.05 / 12.12 ± 5.0349.97 ± 0.35 / 42.88 ± 0.35
BIO-CHEM54.62 ± 0.22 / 42.12 ± 0.5951.92 ± 4.08 / 38.17 ± 4.9666.11 ± 0.11 / 59.12 ± 0.22
BIO-CHEM-SN-CV34.81 ± 0.00 / 21.50 ± 0.0041.15 ± 0.48 / 25.84 ± 1.8456.24 ± 2.02 / 50.13 ± 0.77

Results and Scope: OSFGC achieves higher mean ACC and Macro-F1 than both DENSE and Co-Boost in all eight settings.

AuthorsOpenReview ↗

Response to Reviewer 5Hh7 (4/4)

In the following, we address weaknesses 4 and 5 and the corresponding questions concerning component contributions and cross-dataset label semantics.


4. The contributions of GDS and CR-KMM. (W4, Q4)

Our ablation experiments show the contributions of GDS and CR-KMM at different stages. The five removals in Figure 5 of the main paper distinguish GDS-guided generation, candidate selection, distribution reweighting, and inference calibration. Mean ACC / Macro-F1 (%) over three repeated runs on PROTEINS and BIO-SN-CV are reported below:

AblationInterventionPROTEINS (ACC / Macro-F1)BIO-SN-CV (ACC / Macro-F1)
Full OSFGCRetain all components.76.68 / 73.3649.97 / 42.88
w/o DGReplace learned diffusion with GDS-based random graph sampling.62.75 / 39.4130.63 / 17.36
w/o D-GDS-GDisable GDS structural guidance during diffusion sampling.72.19 / 65.1133.63 / 18.21
w/o CSRemove candidate screening and its associated quality reward.73.78 / 70.6136.61 / 27.13
w/o CR-KMMReplace CR-KMM with uniform allocation within each client-class group.73.30 / 67.8534.22 / 21.64
w/o GDS-PRemove GDS-based predictive calibration.72.41 / 66.0033.06 / 19.23

Additional Comparisons: We also report GDS-only prediction, which classifies directly from the uploaded GDS without graph generation or classifier training, alongside an experiment jointly removing CR-KMM and GDS predictive calibration. Mean ACC / Macro-F1 (%) over three repeated runs are:

VariantPROTEINS (ACC / Macro-F1)BIO-SN-CV (ACC / Macro-F1)
Full OSFGC (reference)76.68 / 73.3649.97 / 42.88
GDS-only64.19 / 48.9331.75 / 21.95
w/o CR-KMM and GDS predictive calibration71.88 / 65.7632.87 / 18.90

GDS-only prediction falls below full OSFGC in both metrics on both datasets, showing that direct classification from GDS alone does not reproduce the full method's performance. Jointly removing CR-KMM and GDS predictive calibration also reduces both metrics on both datasets, supporting the contribution of server-side calibration and fusion. Together with the individual ablations, these comparisons provide complementary evidence for the roles of GDS and CR-KMM in the full pipeline.


5. Label Semantics in Cross-dataset Evaluation. (W5, Q5)

Thank you for the valuable comment. Yes, in our cross-dataset and cross-domain settings, each dataset retains its original class IDs while sharing a common output space. Identical IDs across datasets do not imply identical semantics.

Task Interpretation: With a shared head, the model predicts a graph's class ID within its source dataset. This requires the classifier to implicitly distinguish the relevant domain from graph characteristics and then discriminate among classes within that domain. Differences in graph structure and attributes, such as graph size and connectivity patterns in molecular versus social-network graphs, can provide cues for this implicit domain distinction. Multiple classification tasks are therefore learned jointly through one shared output layer. At test time, each client evaluates only on its own test graphs using its own labels. The results assess how well one global model serves the participating heterogeneous tasks.

Our data partition follows the GCFL/FedStar setup, with each client holding one dataset. Following the strict one-shot protocol, the server learns one global model without further downloads during training or local fine-tuning. We implement this protocol using a shared classification head that serves all clients. OSFGC and the evaluated native one-shot baselines train and evaluate in the same shared label space using a single shared classification head. Meanwhile, local-head methods originally designed for multi-round learning retain their official local classification heads, avoiding cross-dataset label conflicts at the output layer. Thus, the shared-label protocol is consistent across one-shot comparisons and does not give OSFGC an advantage over these local-head baselines: OSFGC additionally has to serve heterogeneous label semantics through one shared head.


Finally, thanks again for your time and helpful comments. We hope your concerns are well addressed. For any other questions, we are happy to discuss them and make further clarifications.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response by Authors

We appreciate the time and effort the reviewers devoted to evaluating our work. Their thoughtful suggestions have guided revisions to both the methodological explanations and the experimental evaluation. We are grateful for the reviewers' recognition of the following strengths:

  • Addressing the meaningful and practically significant problem of one-shot federated graph classification (Reviewers 5Hh7 and ppZK);
  • Proposing a novel combination of graph generation, structure-guided verification, and constrained distribution matching (Reviewer wSg7);
  • Providing reasonable mechanisms for heterogeneous client integration (Reviewer 5Hh7) and synthetic-graph weighting to better match the original distribution (Reviewer ppZK);
  • Improving strict one-shot accuracy (Reviewer wSg7), with promising results across eight heterogeneous settings, particularly in Macro-F1 (Reviewer ppZK);
  • Presenting a well-organized paper with a clear problem formulation and detailed descriptions of the main components (Reviewer 5Hh7).

We have carefully addressed the reviewers' concerns and provide detailed responses to the individual reviews below. Changes in the revised manuscript are highlighted in blue. Key revisions include:

  • Relate the population and synthetic-training objectives through the weight decomposition and distinguish equal within-class client weighting (ρ=0\rho=0) from optional loss-based weighting (Sections 2.1, 2.4–2.5 and Appendices A.3–A.4).
  • Clarify the calibration variants and component ablations (Section 2.5 and Appendix A.5). Compare each prediction branch with full OSFGC and report the joint removal of CR-KMM and GDS calibration (Appendix B.2.2, Table 2).
  • Add graph adaptations of DENSE and Co-Boost across all eight settings (Section 3.2, Table 1 and Appendix B.1). Compare one-shot and multi-round training budgets for four baselines on PROTEINS (Appendix B.2.6, Table 6).
  • Clarify cross-dataset label semantics, the shared-head protocol, and the adaptations of multi-round baselines (Appendix B.1).
  • Specify the fixed-length graph representation for MLP inputs and provide Parameter Settings details (Appendix B.1).
  • Report upload sizes for OSFGC, CeFGC, and OASIS across all eight settings, with a breakdown of OSFGC's payload (Section 3.4 and Appendix B.2.3, Table 3).
  • Add experiments under scarce local data and severe label skew with predictive calibration disabled (Appendix B.2.4, Table 4).
  • Add joint generator and GDS perturbation experiments and clarify the scope of generator-only privacy-budget estimates (Appendix B.2.5, Table 5).

We hope that these revisions and the responses below address the reviewers' concerns. We thank the reviewers again for their time and helpful comments, and welcome further questions and discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.