acceptodds
Withdrawn from ICLR 2027

One-for-All Graph Foundation Model for Federated Graph-Level Clustering

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Federated graph-level clustering aims to discover meaningful pattern groups from decentralized graph data. However, most existing methods require training from scratch for each new non-IID setting and cannot rapidly generalize to unseen domains, resulting in substantial computational and communication overhead. To this end, we propose GFM-FGC, a One-for-All Graph Foundation Model for Federated Graph-Level Clustering, following a **pre-train once, adapt locally to each domain, and share knowledge once** paradigm. is first pre-trained to acquire generalizable graph prior knowledge from public graph data. When deployed in a new non-IID setting, the pre-trained model is distributed to clients and efficiently adapted to their private local graphs through lightweight fine-tuning. The locally adapted knowledge is then shared through a single aggregation round to produce a globally coordinated clustering model. In this way, GFM-FGC can rapidly generalize to unseen non-IID settings without repeated end-to-end training or multiple-round client-server communication. Extensive experiments across multiple non-IID settings demonstrate that GFM-FGC achieves competitive clustering performance while providing improved reusability and efficiency over methods trained separately for each scenario.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Clustering, Unsupervised Graph Learning, Federated Graph-level Clustering

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_scmJOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents a novel graph foundation model for federated graph-level clustering, which aims to enable efficient knowledge transfer across diverse non-IID settings. The proposed framework learns transferable structural knowledge through adversarial pretraining, supports efficient local adaptation, and enables global knowledge sharing with only one round of communication. Extensive experiments demonstrate promising clustering performance and efficiency across different graph domains.

Critical strengths

The paper addresses an important problem in federated graph clustering by exploring how pretrained graph knowledge can be reused across different non-IID settings.

The method integrates structural information into graph pretraining and employs prototype alignment to facilitate knowledge sharing across heterogeneous clients.

The experiments cover 16 graph datasets and six non-IID settings, which demonstrates promising clustering performance. The ablation studies and generalization experiments further support the effectiveness of the proposed framework.

Critical weaknesses

The model is pretrained only on IMDB-BINARY, a relatively small social network dataset, yet it claims to generalize across diverse graph domains. Comparisons with standard graph pretraining methods and experiments using more diverse pretraining datasets would help substantiate this claim.

The proposed framework integrates several established techniques, including subgraph separation, adversarial graph generation, adapter tuning, and optimal transport-based prototype alignment. Although these components are organized into a unified framework, most of them follow existing methodological principles, and their combination does not clearly introduce a fundamentally new learning mechanism. The main contribution therefore appears to lie more in the integration of existing techniques than in substantial algorithmic innovation.

It remains unclear whether clusters from unrelated graph domains share meaningful semantics and how their alignment benefits local clustering.

The current objectives do not establish causal relationships, and the separation loss may allow trivial cluster assignments. The authors should clarify how such solutions are prevented.

Several supervised methods are adapted to unsupervised clustering, which may affect their original effectiveness. More controlled comparisons under comparable training and communication budgets would strengthen the evaluation.

Questions for authors

The model is pretrained only on IMDB-BINARY, a relatively small social network dataset, yet it claims to generalize across diverse graph domains. Comparisons with standard graph pretraining methods and experiments using more diverse pretraining datasets would help substantiate this claim.

The proposed framework integrates several established techniques, including subgraph separation, adversarial graph generation, adapter tuning, and optimal transport-based prototype alignment. Although these components are organized into a unified framework, most of them follow existing methodological principles, and their combination does not clearly introduce a fundamentally new learning mechanism. The main contribution therefore appears to lie more in the integration of existing techniques than in substantial algorithmic innovation.

It remains unclear whether clusters from unrelated graph domains share meaningful semantics and how their alignment benefits local clustering.

The current objectives do not establish causal relationships, and the separation loss may allow trivial cluster assignments. The authors should clarify how such solutions are prevented.

Several supervised methods are adapted to unsupervised clustering, which may affect their original effectiveness. More controlled comparisons under comparable training and communication budgets would strengthen the evaluation.

4 comments
AuthorsOpenReview ↗

Response to reviewer scmJ (1/3)

We sincerely thank the reviewer for the questions raised, which have helped us improve the quality of our paper. We appreciate your recognition of the value contributed by this article. We also kindly request that the reviewer focus on our clarifications regarding these key issues. We have addressed the weaknesses and questions by labeling them W1–5 and Q1–5, respectively, and responding in that order. Please feel free to reach out if you have any further questions; we would be more than happy to answer them. We hope the response can address your concerns and lead to a more favorable assessment. The revised parts have been marked with blue underlines in the revised version.

Response to W1&Q1:

Thank you very much for the question. We have added comparisons with encoders pretrained using GraphCL [1], DGMAE [2], and CORE [3] on the same public data, using the same backbone and pretraining budget while keeping the local adaptation and sharing procedure fixed (Appendix C, Table 7). Section 3.6 and Fig. 5 further evaluate four alternative pretraining sources—uracil, Cuneiform, DBLP_v1, and TRIANGLES—across five downstream federated settings. The competitive performance across these sources supports the transfer capability of GFM-FGC beyond IMDB-BINARY pretraining.

PretrainingAlignmentSM ACCSM NMISM ARISM-BIO-SY ACCSM-BIO-SY NMISM-BIO-SY ARICV ACCCV NMICV ARI
GraphCLRelational Transport62.31.82.665.47.56.835.922.118.7
DGMAERelational Transport64.83.34.566.310.411.234.628.721.5
CORERelational Transport66.25.36.164.47.14.935.030.819.7
OursRelational Transport72.012.413.570.514.816.941.447.530.7

Our pretraining outperforms GraphCL, DGMAE, and CORE, classic and advanced pretraining methods within the same adaptation and sharing pipeline. We attribute this advantage to its clustering-oriented design: subgraph separation encourages the encoder to preserve information supporting stable cluster assignments, while adversarial graph generation exposes it to additional structural variations. Together, these objectives encourage representations that retain clustering-relevant distinctions under structural variation, providing a more suitable initialization for local adaptation and subsequent prototype alignment.

In the revised version, we have included the above content in Appendix C.

Response to W2&Q2:

Thank you very much for your question; however, we respectfully disagree with the characterization of our contribution as merely combining existing techniques. We have incorporated them into an entirely new learning paradigm. Methodological innovation lies in formulating a new learning problem and developing the mechanisms needed to solve it. GFM-FGC addresses reusable, unsupervised federated graph-level clustering: knowledge is learned before target clients are observed, reused across different federated settings, and coordinated through a single communication round. This provides concrete advantages over the task-specific approaches considered in our comparison. The pretrained backbone and structural vocabulary are reused across deployments, avoiding repeated full-backbone training. Local adaptation optimizes lightweight modules, and one-shot coordination removes the dependence on iterative federated co-training. Moreover, this process requires neither ground-truth class labels nor language-model supervision. These capabilities arise from a specific connection between pretraining and deployment. Separation provides the structural reference for generation; generation expands the structural variations encountered during pretraining; and the same vocabulary subsequently describes client prototypes. Relational transport then coordinates independently adapted prototypes within the one-shot constraint. The contribution therefore lies in both the reusable learning paradigm and the seamlessly coupled mechanisms that make its deployment possible. Assessing novelty solely by whether each constituent technique already exists overlooks this methodological distinction.

AuthorsOpenReview ↗

Response to reviewer scmJ (2/3)

Response to W3&Q3:

Alignment is based on three quantities: the distribution of graph embeddings, the distribution of structural tokens, and the relationships among local prototypes (Eqs. 12–13). These descriptors are constructed using a shared pretrained backbone and structural vocabulary. Global prototypes consequently represent reference distributions in this common space, rather than predefined categories whose meanings must be identical across domains. Relational alignment supplements individual prototype similarity with information about how prototypes relate to one another. This shared information benefits local clustering through the feedback stage: clients first adapt to their own graphs, then use the aligned global references to refine their predictions. Unbalanced transport accommodates differences in prototype mass, while confidence-weighted feedback modulates the influence of the inferred correspondences.

The w/o OSC ablation in Fig. 3 shows that removing coordination reduces clustering performance in the evaluated settings. Table 7 further shows that nearest-neighbor matching and transport without the relational term (λR=0\lambda_R=0) both underperform relational alignment, with the pretrained encoder and feedback procedure fixed. These results support the benefits of both cross-client sharing and relational information beyond individual prototype similarity. Inter-cluster relationships provide additional context for comparing heterogeneous prototypes and constructing global references for local prediction refinement. This transfer relies on partial compatibility of clustering-relevant distributional and structural patterns, without requiring identical class semantics across domains.

Response to W4&Q4:

We clarify that “causal subgraph” refers to a causally inspired design: the objective encourages the retained subgraph to preserve clustering information across structural perturbations while reducing its statistical dependence on the complementary component. These objectives do not establish causal identification. We have revised the terminology in the main text to emphasize the causally inspired nature of the separation. Regarding trivial solutions, the separation loss is optimized jointly with the contrastive and reconstruction objectives in Eq. (8), rather than in isolation. The edge-ratio penalty discourages all-edge and no-edge masks, while contrastive learning encourages discriminative graph representations and reconstruction preserves structural information. During local adaptation, the entropy terms in Eq. (11) further discourage uninformative cluster assignments. These mechanisms address different forms of degeneracy, although they do not formally exclude every constant-assignment solution during pretraining.

AuthorsOpenReview ↗

Response to reviewer scmJ (3/3)

Response to W5&Q5:

Thank you for the questions. To address the concern that removing supervision may disadvantage these methods, we additionally evaluated FedStar, CEFGL, FedGFM+, FedBook, and FedGALA using 2%, 5%, and 10% of the training labels with their original supervised objectives on four non-IID settings. These baselines use 10 communication rounds, while GFM-FGC uses no labels and one coordination round. As shown in Section 3.7 and Fig. 6, GFM-FGC achieves the best NMI and ARI across all four settings and label ratios. Its advantage therefore persists when the baselines receive training labels and additional communication opportunities. We also specify the training settings and cumulative costs in Appendices B.2 and D. The following are the evaluated model performance results.

SM-BIO non-IID setting

MethodNMI:2% labelNMI:5% labelNMI:10% labelARI:2% labelARI:5% labelARI:10% label
FedStar8.16.68.77.46.27.1
CEFGL3.04.84.72.33.43.7
FedGFM+3.87.68.04.25.85.5
FedBook5.87.36.83.55.26.0
FedGALA8.78.59.26.47.66.8

SM-BIO-SY non-IID setting

MethodNMI:2% labelNMI:5% labelNMI:10% labelARI:2% labelARI:5% labelARI:10% label
FedStar6.59.810.27.110.811.8
CEFGL3.43.73.63.43.93.9
FedGFM+8.37.88.98.610.39.3
FedBook5.55.99.25.85.68.3
FedGALA10.59.811.311.512.712.3

SN-SY non-IID setting

MethodNMI:2% labelNMI:5% labelNMI:10% labelARI:2% labelARI:5% labelARI:10% label
FedStar14.213.711.48.77.57.0
CEFGL9.710.15.85.64.62.9
FedGFM+6.47.16.53.23.53.2
FedBook9.08.810.34.45.55.1
FedGALA12.08.013.86.64.27.9

CV non-IID setting

MethodNMI:2% labelNMI:5% labelNMI:10% labelARI:2% labelARI:5% labelARI:10% label
FedStar12.421.429.06.812.017.1
CEFGL6.27.16.72.53.73.4
FedGFM+10.09.514.75.24.67.3
FedBook9.018.525.74.39.814.7
FedGALA31.821.934.217.89.220.5

Ours(0% label)

DomainNMIARI
SM-BIO11.810.0
SM-BIO-SY14.816.9
SN-SY18.711.9
CV47.530.7

[1] You Y, Chen T, Sui Y, et al. Graph contrastive learning with augmentations[J]. Advances in neural information processing systems, 2020, 33: 5812-5823.
[2] Zheng Z, Yang Y, Guan Z, et al. Discrepancy-aware graph mask auto-encoder[C]//Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 2025: 4038-4049.
[3] Bo J, Fang Y. CORE: Contrastive Masked Feature Reconstruction on Graphs[J]. arXiv preprint arXiv:2512.13235, 2025.

Thanks for the insightful comments again. We have provided a detailed response. We understand that Reviewer may have a busy schedule. Since only two days are left in the rebuttal period, we look forward to receiving your feedback and communication.

Reviewer_1SH5OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies federated graph-level clustering, hereafter FGC, and argues that existing FGC methods must be retrained from scratch, with many communication rounds, for every new non-IID client setting. To amortize this cost, the authors propose GFM-FGC, a graph foundation model that is pre-trained once on public graphs and then deployed to new federations through lightweight local adaptation plus a single round of knowledge aggregation. The framework has three stages: first, causal-structural adversarial pretraining, where each graph is decomposed by a learned edge mask into a cluster-defining, or causal, subgraph and a biased subgraph, hierarchical structural token vocabularies are built from the biased parts by vector quantization, and a conditional generator plus WGAN-style discriminator synthesize new subgraphs to train the foundation encoder; second, local adaptation on each client with a frozen backbone, residual adapters, and a cluster-level alignment loss against a pretrained Gaussian mixture memory; and third, one-shot knowledge sharing, where clients upload per-cluster Gaussian and structural-token statistics and the server solves a fused Gromov-Wasserstein barycenter problem to align local clusters to global slots, after which clients refine predictions with confidence-gated feedback. Experiments on 16 datasets from 5 domains under 6 non-IID settings report consistent improvements over centralized DGC, adapted supervised FGL, unsupervised FGC, and federated GFM baselines, together with ablations, a communication cost analysis, a generalization study over pretraining sources, and an error bound for the one-shot alignment in Theorem 1.

Critical strengths
  1. Motivation: The problem is well motivated and clearly positioned. Retraining a federated clustering model for every new non-IID setting is genuinely wasteful, and a "pre-train once, adapt locally, share once" paradigm for unsupervised federated graph clustering is a useful and, to my knowledge, new formulation. Table 1 gives a clean comparison against existing federated GFMs, namely FedGFM+, FedBook, FedGALA, and FedGAMMA, along the axes of label dependence, clustering orientation, and PLM/VLM dependence, which makes the gap concrete.
  2. Method design: The design is technically substantial. The cluster-sufficiency and intervention-consistency terms in the separation loss in Eq. 2, the multi-scale structural token vocabularies in Eq. 4, the cluster-conditioned adversarial generation in Eqs. 5-8, and especially the fused Gromov-Wasserstein barycenter for permutation-invariant one-shot alignment of local clusters in Eq. 13 are each reasonable and together form a coherent pipeline. Theorem 1, which decomposes the final clustering error into local assignment error, exponentially small alignment error under a separation margin, and feedback fitting error, is a genuine attempt to justify why a single aggregation round can suffice. A complexity analysis for pretraining and adaptation is also provided.
  3. Experiments: The empirical study is broad: 16 datasets spanning molecules, bioinformatics, social networks, synthetic graphs, and letter recognition, organized into 6 non-IID settings with 2 to 10 clients, four metric families, namely ACC, NMI, ARI, and macro-F1, four baseline categories, module ablations in Fig. 3, a sharing-strategy ablation against FedProx/FedPer/FedNova in Table 3, a communication overhead study in Table 4, hyperparameter sensitivity for R and M in Fig. 4, and a pretraining-source generalization test in Fig. 5. The margins over the best baselines are often large, e.g., 72.0 vs 67.5 ACC on SM2 with 7 clients and 47.5 vs 42.8 NMI on CV15 with 3 clients.
Critical weaknesses

1.Experiments: fairness of adapted baselines. A large fraction of Table 2 compares against methods the authors themselves converted to unsupervised clustering, namely supervised FGL marked *, anomaly detection marked †, centralized DGC marked ‡, and the supervised federated GFMs. Several of these adaptations, e.g., FedGFM+ and FedBook on SM2 with 7 clients and SN3 with 2 clients, collapse to near-zero NMI/ARI, which suggests the adaptation protocol may be weak rather than the underlying methods. Section 3.1 and Appendix B.2 describe the adaptation only at a high level: "replace label-dependent training objectives with an unsupervised clustering objective"; the exact objective, tuning budget, and validation protocol per baseline are not given, making it hard to judge whether the comparison is fair. 2. Experiments: single pretraining source and possible leakage. Appendix B.2 states that all main results pretrain only on IMDB-BINARY, a small binary social-network dataset, while the SN and SN-SY evaluation settings include IMDB-MULTI. The paper never discusses whether IMDB-BINARY and IMDB-MULTI graphs overlap or come from the same underlying source, which is a potential train/test contamination issue for precisely the settings where the model is evaluated. The "foundation model" claim also rests on pretraining with roughly one thousand graphs; Fig. 5 varies the pretraining dataset but only over four choices and without leakage analysis. 3. Method design/theory: limited practical bite of Theorem 1. The bound assumes a stationary transport plan with a positive separation margin kappa on the "effective alignment cost" in Eq. 35, which itself depends on the unknown optimal plan, and the bound contains the average local assignment error eta-bar that the theory cannot control. It is best read as a consistency argument rather than a guarantee, and this should be stated more honestly in the main text.

Questions for authors

Please refer to weakness.

4 comments
AuthorsOpenReview ↗

Response to reviewer 1SH5 (1/2)

We sincerely appreciate your recognition of our approach and kindly ask for your continued support, especially regarding the innovation issue that other reviewers raised. Below is a point-by-point response to the questions you raised. We hope this helps address your concerns and leads to a better score. We have addressed the weaknesses and questions by labeling them W1–3, respectively, and responding in that order. Please feel free to reach out if you have any further questions; we would be more than happy to answer them. The revised parts have been marked with blue underlines in the revised version.

Response to W1:

We thank the reviewer for raising this important concern regarding the fairness of the adapted baselines. In the original manuscript, we replaced the label-dependent supervised loss with unsupervised node attribute reconstruction and a KL loss between soft cluster assignments and sharpened target distributions to adapt these supervised methods. Reconstruction and KL self-supervision are widely used and standard self-supervised objectives in the clustering literature. Regarding the near-zero NMI/ARI results, we believe there are two contributing factors. First, adapting these models to the unsupervised setting can substantially degrade their performance. Second, there is a dataset-level issue: even models designed specifically for graph-level clustering [1,2] do not achieve strong NMI/ARI on these datasets. Together, these factors explain the overall low performance. To further address your concerns and assess whether removing supervision accounts for the performance degradation, we have additionally evaluated FedStar, CEFGL, FedGFM+, FedBook, and FedGALA using 2%, 5%, and 10% of the training labels with their original supervised objectives on four non-IID settings. These baselines use 10 communication rounds, while GFM-FGC uses no labels and one coordination round. As shown below, GFM-FGC achieves the best NMI and ARI across all four settings and label ratios. Its advantage therefore persists when the baselines receive training labels and additional communication opportunities. We also specify the training settings and cumulative costs in Appendices B.2 and D. The following are the evaluated model performance results.

SM-BIO non-IID setting

MethodNMI: 2% labelNMI: 5% labelNMI: 10% labelARI: 2% labelARI: 5% labelARI: 10% label
FedStar8.16.68.77.46.27.1
CEFGL3.04.84.72.33.43.7
FedGFM+3.87.68.04.25.85.5
FedBook5.87.36.83.55.26.0
FedGALA8.78.59.26.47.66.8

SM-BIO-SY non-IID setting

MethodNMI: 2% labelNMI: 5% labelNMI: 10% labelARI: 2% labelARI: 5% labelARI: 10% label
FedStar6.59.810.27.110.811.8
CEFGL3.43.73.63.43.93.9
FedGFM+8.37.88.98.610.39.3
FedBook5.55.99.25.85.68.3
FedGALA10.59.811.311.512.712.3

SN-SY non-IID setting

MethodNMI: 2% labelNMI: 5% labelNMI: 10% labelARI: 2% labelARI: 5% labelARI: 10% label
FedStar14.213.711.48.77.57.0
CEFGL9.710.15.85.64.62.9
FedGFM+6.47.16.53.23.53.2
FedBook9.08.810.34.45.55.1
FedGALA12.08.013.86.64.27.9

CV non-IID setting

MethodNMI: 2% labelNMI: 5% labelNMI: 10% labelARI: 2% labelARI: 5% labelARI: 10% label
FedStar12.421.429.06.812.017.1
CEFGL6.27.16.72.53.73.4
FedGFM+10.09.514.75.24.67.3
FedBook9.018.525.74.39.814.7
FedGALA31.821.934.217.89.220.5

Ours(0% label)

DomainNMIARI
SM-BIO11.810.0
SM-BIO-SY14.816.9
SN-SY18.711.9
CV47.530.7

We follow the original recommended hyperparameters of each baseline. For the unsupervised comparisons in Table 2, any hyperparameter tuning uses only unsupervised validation objectives. Our adaptation only replaces the label-dependent supervised loss with reconstruction and KL losses, while retaining the original architecture and aggregation mechanism. We have expanded Appendix B.2 to specify the training epochs and communication rounds for each baseline. Each baseline uses 10 communication rounds, with each client performing 10 local training epochs per round.

AuthorsOpenReview ↗

Response to reviewer 1SH5 (2/2)

Response to W2:

Thank you for your question. First, our evaluation is not restricted to a single pretraining source. IMDB-BINARY is the default source for the main comparisons. Section 3.6 and Figure 5 additionally evaluate pretraining on uracil, Cuneiform, DBLP_v1, and TRIANGLES across multiple downstream settings, including SN. These results provide evidence that the observed transfer behavior is not specific to IMDb-derived pretraining. The public dataset documentation does not explicitly specify whether IMDB-BINARY and IMDB-MULTI overlap. We therefore conducted a cross-dataset graph-isomorphism check and found no isomorphic graph pairs, indicating no exact structural duplication between the pretraining and evaluation graphs. We have included this verification in the appendix. Regarding the foundation-model terminology, we emphasize the reuse of a single pretrained backbone across domains and federated deployments. Our experiments demonstrate this transfer capability with a modest pretraining corpus, and we have clarified that the demonstrated scope is federated graph-level clustering.

Response to W3:

Thank you for your attention to this issue. Theorem 1 considers a positive stationary transport plan with the server variables held fixed, without requiring global optimality; its effective cost is evaluated at that plan. To clarify when the required margin arises, we have added a geometric sufficient condition in Appendix E. Under a shared reference-prototype model, let Δ\Delta denote the minimum reference separation, r<Δ/2r<\Delta/2 the combined local and global prototype deviation, and BB bound the adverse contribution of relational and marginal terms. We derive κeff≥Δ(Δ−2r)−B.\kappa_{\mathrm{eff}}\geq\Delta(\Delta-2r)-B. A positive lower bound supplies the separation margin required by Theorem 1. The term ηˉ\bar{\eta} represents the local assignment error entering coordination. Given this local quality, the theorem quantifies how alignment and feedback contribute to the final clustering error. We therefore characterize the result as a conditional error-propagation bound, with explicit geometric conditions supporting reliable alignment.

[1] Cai J, Zhang Y, Fan J, et al. Dual contrastive graph-level clustering with multiple cluster perspectives alignment[C]//Proceedings of the 33rd International Joint Conference on Artificial Intelligence, IJCAI 2024. 2024: 3770-3779.
[2] Chen M S, Lai P Y, Liao D Z, et al. Graph prompt clustering[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025, 47(7): 5794-5805.

Thanks for the insightful comments again. We have provided a detailed response. We understand that Reviewer may have a busy schedule. Since only two days are left in the rebuttal period, we look forward to receiving your feedback and communication.

Thank you again for your insightful comments. We have provided a detailed response. We understand that you may have a busy schedule. As the rebuttal period is about to end, we look forward to your feedback and hope to receive your support.

Reviewer_vuKnOpenReview ↗
Rating
2
Confidence
3
Informativeness
3(+1)
Presentation
2
Soundness
3(+1)
Read review
Paper summary

The paper presents GFM-FGC for unsupervised federated graph-level clustering. A graph encoder is pretrained on public data using subgraph separation, structural tokens, and adversarial graph generation. Clients then adapt a frozen backbone and exchange cluster statistics in one communication round. The server aligns these statistics through relational optimal transport and returns prototypes to refine local clustering. Experiments cover six federated settings built from sixteen datasets and report improvements over existing methods.

Critical strengths
  1. A useful deployment goal. The paper addresses the cost of repeatedly training federated clustering models for new settings. Reusing a pretrained backbone and coordinating clients through one exchange of cluster summaries is a practical direction worth studying.
  2. Consistent reported improvements. The method leads the listed baselines across six federated settings and four clustering metrics in Table 2. Comparisons with recent methods and module ablations provide evidence that the complete framework is effective on these benchmarks.
Critical weaknesses
  1. The contribution is incremental and the design needs stronger justification. Subgraph separation and prototype sharing already appear in FGCN-DKS. The main addition is a reusable pretraining and one-shot coordination pipeline. The paper explains the intended role of each module, but gives limited justification for combining causal separation, adversarial generation, and relational transport. Preserving a learned pseudo-partition does not establish that the retained structure is causal or transferable. The scope of the alignment is also unclear: clusters from different domains need not follow the same grouping criteria. Theorem 1 assumes that the correct match already has a lower effective cost, leaving the main question of when reliable transfer is possible unresolved.
  2. The experiments do not establish the source of the gains or the overall efficiency. The complete system performs well, but the experiments do not isolate the benefit of the proposed mechanisms. A standard self-supervised encoder trained on the same public data, followed by the same local adaptation and sharing procedure, is missing. Figure 3 tests whether components help within this implementation, but not whether they outperform simpler alternatives. Table 3 also lacks a comparison with simpler prototype matching under the same backbone. Finally, Table 4 reports communication per round without baseline round counts or total runtime. The claimed efficiency gains therefore remain unverified once pretraining, local adaptation, server optimization, and feedback are included.
Questions for authors
  1. Clarification of the setting. What shared structure makes clusters from different domains suitable for alignment to common global prototypes? Please also clarify the local and global cluster counts in SN-SY, where the client datasets have different numbers of classes.
  2. Evidence for the main contribution. How much of the improvement remains with standard self-supervised pretraining on the same public data, keeping the backbone, local adaptation, and sharing procedure fixed? A comparison with simpler prototype matching would also help establish the value of relational alignment.
  3. Clarification of the costs. What are the cumulative communication and total runtime of each method? Please state the baseline communication rounds and whether the reported costs include initial model distribution, local adaptation, server optimization, and feedback. The pretraining cost and its amortization across deployments should also be reported.
  4. Clarification of the evaluation. Table 2 reports 23.5% ACC for FGCN-DKS on SN-SY. Appendix B.3 specifies optimal Hungarian matching and equal numbers of predicted clusters and classes. With four clusters, this gives an ACC of at least 25%. Please clarify the actual cluster counts, matching procedure, and averaging scheme.
8 comments
AuthorsOpenReview ↗

Response to reviewer vuKn (weakness part) (1/5)

We sincerely thank the reviewer for the questions raised, which have helped us improve the quality of our paper. We also kindly request that the reviewer focus on our clarifications regarding these key issues. We have addressed the weaknesses and questions by labeling them W1–2 and Q1–4, respectively, and responding in that order. We hope this will address your concerns and lead to a more favorable assessment. Please feel free to reach out if you have any further questions; we would be more than happy to answer them. The revised parts have been marked with blue underlines in the revised version.

Response to W1.1:

Thank you for your question. However, we respectfully disagree that the contribution is adequately characterized by the presence of subgraph separation and prototype sharing. Compared with FGCN-DKS and existing FGC baselines, our model has a significant paradigm distinction. Our model instead learns reusable knowledge before deployment and quickly applies it to new non-IID settings through local adaptation and a single coordination round. The contribution is therefore a change in how clustering knowledge is learned, reused, and coordinated across deployments. This setting motivates the combination of the following three mechanisms. Because target clients are unavailable during pretraining, the model must acquire useful clustering structure without optimizing for a particular federation. Subgraph separation identifies the structural information that the pretraining objective seeks to preserve, while structure-conditioned generation introduces variation around that retained information. The separation thus supplies a reference for generation, rather than leaving augmentation unconstrained by the learned clustering structure. Importantly, the structural vocabulary connects pretraining and deployment: its tokens guide generation and subsequently provide structural descriptors for client prototypes. At deployment, independent local adaptation can produce different prototype arrangements, while a single coordination round offers no opportunity for repeated federated correction. Relational transport addresses this problem by using both individual prototype distributions and relationships among prototypes. Together, these design choices generate a seamless framework that enables rapid adaptation across different non-IID environments. This reduces repeated full-model training and avoids iterative client–server coordination, easing the computational burden of repeated deployments.

More importantly, our approach differs fundamentally from the components of FGCN-DKS. FGCN-DKS emphasizes cluster compactness, inter-cluster separation, and environmental invariance. Our objective explicitly preserves full-graph assignments, suppresses cluster predictability in the complementary branch, and reduces dependence between the two branches. We further convert complementary structures into tokens for generation and deployment-time prototype description, turning separation into a source of reusable pretraining knowledge. Prototype sharing. FGCN-DKS uses graph-kernel affinities and historical relationships to weight prototype and parameter aggregation. We jointly learn soft correspondences and global distributional prototypes through relational transport, incorporating covariance, structural-token distributions, and prototype relationships. This coordinates independently adapted clients in one round while keeping the pretrained backbone fixed, avoiding iterative encoder-parameter aggregation.

Furthermore, during this rebuttal process, we have conducted additional experiments that further highlight our advantages.

AuthorsOpenReview ↗

Response to reviewer vuKn (weakness part) (2/5)

Response to W1.2:

Thank you very much for the reviewer's questions; we address them point by point.

About "Preserving a learned pseudo-partition does not establish that the retained structure is causal or transferable."

  1. The purpose of pseudo-partitions is to extract cluster-oriented substructures from within the clusters; these substructures exhibit a degree of transferability across clusters in different domains, as noted in [1–3]. Eq. (2) combines several complementary constraints. The partition-preservation term encourages retained subgraphs to preserve the full-graph assignments, while intervention consistency favors structures supporting stable assignments across perturbations. The complementary-branch entropy term discourages confident cluster predictions from the remaining topology, and HSIC reduces dependence between the retained and complementary representations. The edge-ratio constraint additionally prevents retaining all edges. Together, these constraints encourage the separation of stable, partition-supporting structure from unstable correlations. Moreover, we also evaluate transferability experimentally. The w/o CS ablation examines the contribution of separation, while Section 3.6 evaluates transfer using different pretraining sources across downstream settings.

About "The scope of the alignment is also unclear: clusters from different domains need not follow the same grouping criteria."

  1. The alignment does not assume identical class semantics; it relies on compatible clustering-relevant structural patterns across domains. Eq. (13) operates on embedding distributions, structural-token distributions, and relationships among prototypes. The method also accommodates heterogeneous correspondences through soft, unbalanced transport instead of a hard one-to-one class assignment. During feedback, Eq. (16) weights the transferred targets by matching confidence and regularizes uncertain local assignments toward their pre-feedback values. These mechanisms explain how shared information is incorporated without treating every inferred correspondence equally. The w/o OSC ablation provides direct evidence that coordination improves clustering performance in the evaluated federations.

About "Theorem 1 assumes that the correct match already has a lower effective cost, leaving the main question of when reliable transfer is possible unresolved"

  1. To clarify when the margin assumed in Theorem 1 arises, we have added a sufficient condition based on prototype separation and distortion. Under a shared reference-prototype model, let Δ\Delta denote the minimum reference-prototype separation, rr the combined local and global prototype deviation, and BB bound the adverse contribution of the relational and marginal terms. We derive κeff≥Δ(Δ−2r)−B.\kappa_{\mathrm{eff}}\geq\Delta(\Delta-2r)-B. When this lower bound exceeds εlog⁡((J−1)(1−α)/α)\varepsilon\log((J-1)(1-\alpha)/\alpha), the transport mass assigned to incorrect correspondences is at most α\alpha. This makes the condition explicit: reliable alignment is supported when distinguishable prototype structure survives cross-domain distortion with sufficient separation relative to entropy regularization.

[1] Wang Z, Zhang Z, Ma T, et al. Towards graph foundation models: Learning generalities across graphs via task-trees[J]. arXiv preprint arXiv:2412.16441, 2024.
[2] Miao S, Liu M, Li P. Interpretable and generalizable graph learning via stochastic attention mechanism[C]//International conference on machine learning. PMLR, 2022: 15524-15543.
[3] Wu Y X, Wang X, Zhang A, et al. Discovering invariant rationales for graph neural networks[J]. arXiv preprint arXiv:2201.12872, 2022.

AuthorsOpenReview ↗

Response to reviewer vuKn (weakness part) (3/5)

Response to W2

Thank you very much for your question. To address your issue, we have conducted the following extended experiments; we provide a point-by-point response below.

About "A standard self-supervised encoder trained on the same public data, followed by the same local adaptation and sharing procedure, is missing."

We have compared our pretrained encoder with GraphCL [4], DGMAE [5], and CORE [6] encoders trained on the same public data, using the same backbone and pretraining budget. All variants use the same local adaptation, prototype sharing, and feedback procedure. The experimental results are as follows:

PretrainingAlignmentSM ACCSM NMISM ARISM-BIO-SY ACCSM-BIO-SY NMISM-BIO-SY ARICV ACCCV NMICV ARI
GraphCLRelational Transport62.31.82.665.47.56.835.922.118.7
DGMAERelational Transport64.83.34.566.310.411.234.628.721.5
CORERelational Transport66.25.36.164.47.14.935.030.819.7
OursRelational Transport72.012.413.570.514.816.941.447.530.7

Our pretraining outperforms GraphCL, DGMAE, and CORE, classic and advanced pretraining methods within the same adaptation and sharing pipeline. We attribute this advantage to its clustering-oriented design: subgraph separation encourages the encoder to preserve information supporting stable cluster assignments, while adversarial graph generation exposes it to additional structural variations. Together, these objectives encourage representations that retain clustering-relevant distinctions under structural variation, providing a more suitable initialization for local adaptation and subsequent prototype alignment.

About "Figure 3 tests whether components help within this implementation, but not whether they outperform simpler alternatives. Table 3 also lacks a comparison with simpler prototype matching under the same backbone."

We verify the effectiveness of the method by adding simpler alternatives. We compare nearest-prototype matching and transport without the relational term λR=0\lambda_R=0. The backbone and feedback procedure were kept fixed. These comparisons distinguish the benefit of relational information from that of prototype sharing and pointwise matching alone.

PretrainingAlignmentSM ACCSM NMISM ARISM-BIO-SY ACCSM-BIO-SY NMISM-BIO-SY ARICV ACCCV NMICV ARI
OursNearest Matching67.08.411.267.412.813.638.244.925.5
OursTransport (λR=0\lambda_R=0)70.110.612.068.912.714.539.543.127.4
OursRelational Transport72.012.413.570.514.816.941.447.530.7

Replacing relational alignment with nearest-neighbor matching or removing the relational term reduces clustering performance. Pointwise prototype similarities can be ambiguous across heterogeneous clients, while relationships among clusters provide additional context for resolving these ambiguities. This helps explain the more effective feedback and stronger local clustering performance of the full model.

[4] You Y, Chen T, Sui Y, et al. Graph contrastive learning with augmentations[J]. Advances in neural information processing systems, 2020, 33: 5812-5823.
[5] Zheng Z, Yang Y, Guan Z, et al. Discrepancy-aware graph mask auto-encoder[C]//Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 2025: 4038-4049.
[6] Bo J, Fang Y. CORE: Contrastive Masked Feature Reconstruction on Graphs[J]. arXiv preprint arXiv:2512.13235, 2025.

Comments on the paper

3 comments
AuthorsOpenReview ↗

We sincerely request the reviewer's reply.

Dear reviewers:
Thank you for your comments and suggestions. We have posted our response addressing the points raised. If any reviewer has a moment to take a look and share further thoughts, we would be very grateful. Thank you again for your time and consideration.

Submission5811 Authors

Dear Reviewers,

As the author–reviewer discussion period is approaching its end, we would like to express our deepest gratitude for your valuable time and efforts.

We sincerely and humbly implore you to kindly revisit our responses and reconsider your evaluations. We would be immensely grateful for any additional feedback or further comments you may be willing to provide.

Your reconsideration would mean a great deal to us, and we wholeheartedly appreciate your kindness, patience, and consideration.

AuthorsOpenReview ↗

Rebuttal Summary

Dear AC, SAC and Reviewers:
Thank you very much for the time spent by AC, SAC and reviewers for our manuscript review.

Our paper initially received scores of 3, 3, and 2, with two reviewers recommending acceptance and one expressing strong support for our approach. Following our rebuttal, the reviewer who initially assigned a score of 2 acknowledged that we had addressed most of their concerns and increased their Soundness and Informativeness scores. Their remaining concern focused on the motivation, which we further clarified in our follow-up response. The other reviewers raised no further questions following our rebuttal. Overall, we believe the discussion has substantially addressed the reviewers’ concerns and strengthened the case for acceptance.

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

Related papers

Discussion (0)

Sign in to comment.