acceptodds
Under review as a conference paper at ICLR 2027

Dual Manifold Calibration for Graph Federated Learning

Abstract

Graph Federated Learning (GFL) enables collaborative representation learning across distributed graphs while preserving privacy. However, heterogeneity remains a critical challenge, as graphs across clients typically differ significantly in both semantics and structures. Existing methods address semantic heterogeneity and structural heterogeneity by enforcing the alignments of prototypes and spectral characteristics between clients and the server, respectively. However, these alignments implicitly enforce identical local representations across clients, which significantly compresses the personalized representation space of clients. To overcome these limitations, we propose erated raph anifold alibration (FedGMC), a novel paradigm that tackles semantic heterogeneity and structural heterogeneity from a unified manifold perspective. Instead of enforcing identical local representations, FedGMC introduces a dual manifold calibration mechanism that allows each client to establish client-specific correspondences with globally shared semantic and structural references. Specifically, for semantic heterogeneity, the server constructs a geometrically optimal semantic manifold via equidistant semantic anchors, so as to guide the calibration of local semantic manifolds. For structural heterogeneity, the server constructs a global structural manifold by building global structural templates, so as to guide the calibration of local structural manifolds. Finally, the server dynamically refines both global semantic manifolds and structural manifolds by aggregating local manifolds. Extensive experiments on eleven datasets demonstrate that FedGMC effectively balances global commonality and local personalization, thereby significantly outperforming eleven state-of-the-art methods. Our code is available at https://anonymous.4open.science/r/FedGMC.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Graph Federated Learning

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_EVs4OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies federated node classification under heterogeneous client subgraphs and proposes Federated Graph Manifold Calibration (FedGMC), which shares semantic and structural information through two calibration branches.

For semantic calibration, each client constructs class prototypes from structure-independent ego-embeddings. The server initializes global class anchors using a simplex equiangular tight frame (ETF), and clients compute orthogonal Procrustes transformations to establish correspondences between local prototypes and these anchors. For structural calibration, one-hop and two-hop neighborhood representations are treated as two-point empirical measures and matched to global structural templates using entropically regularized optimal transport. The server subsequently refines the semantic anchors through difficulty-weighted updates and updates the structural templates through a Gromov--Wasserstein (GW) barycenter-related procedure. Local training combines the classification loss with the two calibration losses.

The evaluation covers eleven homophilic and heterophilic graph datasets, non-overlapping and overlapping subgraph partitions, and multiple client counts. The paper also provides component ablations, quantitative geometry-preservation measurements, experiments under severe label heterogeneity, and computation and communication measurements. The reported results indicate that the complete method achieves strong overall predictive performance relative to the baselines included in the main result tables.

Critical strengths
  1. Broad empirical coverage, with meaningful improvements on heterophilic graphs.

The evaluation covers eleven datasets, two subgraph-partitioning settings, and multiple client counts, with means and standard deviations over ten independent runs reported for the main comparisons. Relative to FedIIH, the average homophilic result improves from 83.1083.10 to 83.9083.90 under non-overlapping partitions, while the average heterophilic result improves from 64.5264.52 to 68.9568.95. Under overlapping partitions, the corresponding gains are 0.230.23 and 2.422.42 percentage points. The heterophilic averages follow the paper's aggregation of Accuracy and AUC results and should be interpreted accordingly.

The improvements are therefore not confined to a single dataset or partitioning configuration. This is an important reason to regard the method as potentially useful in practice. See Tables 1, 2, 4, and 5.

  1. A reasonably motivated separation of semantic and structural knowledge, supported by substantial supplementary evaluation.

The use of class prototypes and multi-hop neighborhood summaries gives the two branches distinct roles in addressing heterogeneous client data. The comprehensive ablations in Table 6 show that the semantic branch, structural branch, and global refinement each contribute to the performance of the complete system.

The supplementary experiments also examine geometry preservation relative to locally trained reference models, missing classes and severe label heterogeneity, larger neighborhood depths, and efficiency. For example, on Cora, Table 10 reports PDC/GD values of 0.934/0.0820.934/0.082 for FedGMC, compared with 0.887/0.1430.887/0.143 for FedIIH. These measurements provide useful empirical support beyond the main predictive-performance tables.

Critical weaknesses
  1. The global refinement procedure does not establish the geometric properties attributed to the actual updates.

The most consequential issue concerns the structural branch. Equation (9) defines the update of the vector-valued structural template as a GW barycenter problem. However, Appendix D.5 first optimizes an intrinsic metric matrix through Eqs. (102)--(104), and then reconstructs support vectors through coordinate averaging and normalization in Eqs. (105)--(106). The latter operations do not generally ensure that the reconstructed template induces the optimized metric.

A simple example illustrates the problem. Let e1\mathbf{e}_1 and e2\mathbf{e}_2 be orthogonal unit vectors, and consider two equally weighted local descriptors:

\begin{bmatrix} \mathbf{e}_1^\top\\ \mathbf{e}_2^\top \end{bmatrix}, \qquad \mathbf{R}_2= \begin{bmatrix} \mathbf{e}_2^\top\\ \mathbf{e}_1^\top \end{bmatrix}.$$ Both have the intrinsic metric $$\mathbf{A}_1=\mathbf{A}_2= \begin{bmatrix} 0&2\\ 2&0 \end{bmatrix}.$$ Their GW subproblems are identical and can return the same coupling. For a coupling of the form $$\mathbf{P}= \begin{bmatrix} p&\frac12-p\\ \frac12-p&p \end{bmatrix}, \qquad 0\leq p\leq\frac12,$$ Eq. (104) produces a positive off-diagonal metric value in $[1,2]$. Nevertheless, Eqs. (105)--(106) reconstruct both support vectors as $(\mathbf{e}_1+\mathbf{e}_2)/\sqrt{2}$, yielding zero distance between them. This example uses nonnegative, normalized representations in a common coordinate space and does not depend on assuming hard or unregularized transport. Consequently, the final broadcast template cannot generally be identified with the solution of Eq. (9), and the isometry invariance of the GW discrepancy does not automatically extend to the complete template update. There is a related issue in the semantic branch: Eq. (8) enforces unit norms but does not enforce the ETF cross-inner-product constraints in Eq. (1). Moreover, Theorems A.8--A.9 do not establish that the actual difficulty-weighted, clipped, OT/GW-based updates and template reconstruction satisfy their required assumptions. These gaps directly affect the paper's central interpretation of geometry-aware calibration and preservation. 2. The empirical comparisons do not sufficiently isolate the claimed advantage of calibration over direct alignment. Table 6 provides extensive component-removal ablations. However, removing an entire branch simultaneously removes its supervision and shared information, so these experiments do not isolate the contribution of the specific geometric mechanisms. It remains unclear how much of the improvement comes from orthogonal calibration, ETF supervision, generic auxiliary regularization, or the particular structural-template update. More direct controls would retain the same backbone, anchors, loss settings, and training budget while comparing $\mathbf{R}_m=\mathbf{I}$ against Procrustes calibration, and a simple template update against the proposed GW-related refinement. Predictive performance and geometry-preservation metrics should be examined together. The closest related comparison is also incompletely reported. FedSSA appears in the legends of Figs. 3, 6, and 7, but the main result tables and Appendix C.5 do not provide a consistent account of its final means, standard deviations, and implementation settings. Furthermore, FedSSA uses cluster-level references, so the characterization of prior approaches as forcing identical representations across all clients requires qualification. Together, these omissions limit the evidence for the claimed advantage over the most directly relevant alignment-based alternatives.
Questions for authors
  1. Relationship between the GW barycenter and the broadcast structural templates.

Equation (9) defines the structural-template update as a GW barycenter problem. In Appendix D.5, however, Eqs. (102)--(104) optimize an intrinsic distance matrix, whereas Eqs. (105)--(106) produce the broadcast template through coordinate averaging and normalization. Could the authors clarify the mathematical relationship between these two stages? In particular, is the reconstructed template intended to realize the optimized distance matrix, or is reconstruction an additional approximation? If it is an approximation, which geometric properties of the GW solution remain valid for the template actually broadcast to clients? This clarification is important for interpreting Eq. (9) and the isometry-invariance claim in Appendix D.6.

  1. Scope of the ETF geometry claim after semantic refinement.

Could the authors clarify whether the ETF optimality claim applies only to the initial anchors or also to the anchors after refinement? Section 4.4.1 describes the repulsive update as approximating maximal equidistance, while the surrounding text also states that this property is maintained. Equation (8) enforces unit norms, but the ETF condition in Eq. (1) additionally requires δi⊤δj=−1/(C−1)\boldsymbol{\delta}_i^\top\boldsymbol{\delta}_j=-1/(C-1) for i≠ji\neq j. If the refined anchors are intended to preserve this geometry only approximately, what analytical or empirical evidence supports that characterization? Please distinguish the properties of the initial ETF from those of the data-adapted reference used during training.

  1. Applicability of the convergence results to the implemented algorithm.

Theorem A.8 analyzes the Riemannian stochastic-gradient update in Eq. (44), and Theorem A.9 assumes the sufficient-decrease and relative-error conditions in Eqs. (57)--(58). Which update sequences generated by FedGMC are covered by these results? For A.8, please explain how the difficulty-weighted and clipped update in Eq. (8) corresponds to the analyzed gradient step. For A.9, please clarify whether the client-side OT calibration, server-side GW refinement, and subsequent template reconstruction are covered by a common objective satisfying the stated conditions. If the results concern idealized or fixed-representation subproblems, please distinguish that scope from convergence of the complete training procedure.

  1. Evidence isolating calibration from direct alignment.

Table 6 already establishes the contribution of the main components, and Table 10 evaluates geometry preservation after training. To assess the more specific claim that orthogonal calibration improves on direct alignment, could the authors provide a controlled comparison with Rm=I\mathbf{R}_m=\mathbf{I} on the existing Cora/Tolokers setup in Appendix F, retaining the same backbone, anchor initialization, remaining losses, refinement procedures, and training budget? Reporting both task performance and PDC/GD would help determine whether the gains arise from the calibration mechanism rather than the additional supervision alone.

4 comments
AuthorsOpenReview ↗

To Reviewer EVs4

Thanks for your comments!


Re: “Q1”: Thanks! We clarify that the mathematical relationship between these two stages is min⁡∥tq(k)∥2=1∑m,bλm,b(q)∑i,k(Pm,b,q)i,k∥rm,b(i)−tq(k)∥22.\min_{\|\mathbf t_q^{(k)}\|_2=1} \sum_{m,b}\lambda_{m,b}^{(q)} \sum_{i,k}(\mathbf P_{m,b,q})_{i,k} \|\mathbf r_{m,b}^{(i)}-\mathbf t_q^{(k)}\|_2^2.. In particular, the reconstruction is an additional approximation to the GW barycenter objective in Eq. (9). The geometric properties of the GW solution, including prescribed weights, unit norms, and a valid matrix of squared Euclidean distances, remain valid for the template actually broadcast to clients.


Re: “Q2”: Thanks! We clarify that the ETF optimality claim applies only to the initial anchors. The refined anchors are intended to approximately preserve the geometry, supported by the analysis of data attraction and logarithmic repulsion, as revealed in Appendix A.2.1 of our revised paper. The properties of the initial ETF from those of the data-adapted reference used during training can be distinguished by examining their Gram matrices. All anchors of the initial ETF have unit norm, and every off-diagonal inner product equals −1C\frac{-1}{C}. In contrast, the off-diagonal inner products of data-adapted reference can deviate from the ETF value.


Re: “Q3”: Thanks! The convergence results cover two types of update sequences. Theorem A.8 covers the idealized Riemannian stochastic-gradient iterates as in Eq. (44). Theorem A.9 covers structural template and assignment iterates with fixed client representations as in Eqs. (57)–(58). For A.8, the difficulty-weighted and clipped update in Eq. (8) corresponds to the analyzed gradient step via the same spherical retraction. For A.9, we clarify that the client-side OT calibration, server-side GW refinement, and subsequent template reconstruction are covered by a common objective satisfying the stated conditions. The convergence results concern idealized and fixed-representation subproblems. Consequently, we must distinguish this scope from the convergence of the complete training procedure. A separate analysis accomplishes this distinction. This analysis accounts for the interaction among local encoder training, server-side refinement, and template reconstruction across communication rounds. Appendix A.3.2 of our revised paper clarifies this distinction in detail.


Re: “Q4”: Thanks! To assess the more specific claim that orthogonal calibration improves on direct alignment, we provide a controlled comparison with Rm=I\mathbf{R}_m = \mathbf{I} on the existing Cora/Tolokers setup, retaining the same backbone, anchor initialization, remaining losses, refinement procedures, and training budget. The detailed experiments are reported in Appendix F.2.2 of our revised paper. Here we report the task performance, PDC, and GD of our FedGMC and FedGMC--Identity on two datasets (i.e., Cora and Tolokers). The experimental results below demonstrate that the gains of our FedGMC arise from the calibration mechanism rather than the additional supervision, supporting the claim that orthogonal calibration improves on direct alignment.

DatasetVariantTask performance (%) ↑\uparrowPDC ↑\uparrowGD ↓\downarrow
CoraFedGMC--Identity80.97$\pm$0.870.193$\pm$0.0410.705$\pm$0.083
FedGMC83.65$\pm$0.160.934$\pm$0.0420.082$\pm$0.030
TolokersFedGMC--Identity71.72$\pm$0.660.512$\pm$0.0100.508$\pm$0.031
FedGMC74.03$\pm$0.090.891$\pm$0.0270.119$\pm$0.035
Reviewer_EVs4OpenReview ↗

Response to rebuttal

Many thanks for the detailed and responsive revisions. My main concerns have been effectively resolved:

  • GW template reconstruction: The explicit clarification in Appendix E.5 makes the interpretation of Eq. (9) and the coverage of the theoretical guarantee much clearer.
  • Direct alignment comparison: Table 7, with the added Rm=IR_m=I control, convincingly demonstrates the benefits of Procrustes calibration (+2.68 Cora accuracy, +2.31 Tolokers AUC), alongside better geometry preservation.
  • Remaining question: The remaining question about the fidelity of refined anchors to the ETF geometry is a minor point I would be interested to see in a future version.

Overall, these revisions address my core feedback. I am happy to raise my score from 2 to 3.

AuthorsOpenReview ↗

Thank you for your review and feedback!

Dear Reviewer EVs4,

Thank you for taking the time to review our paper.

We sincerely appreciate your thoughtful feedback, which has significantly strengthened our work.

Best regards,

Authors of Paper ID 12771

AuthorsOpenReview ↗

Further Clarification Regarding Q2

Re: “Q2”: Thanks for your valuable comment. The GW coupling establishes correspondences between support points. However, it does not align their feature axes. Therefore, we have corrected this explanation in both the main paper and the supplementary material.

The common-reference mechanism in our method is the client-side Euclidean OT loss rather than a transformation supplied by GW. All clients evaluate their structural supports against the same broadcast template vectors. Consider a fixed inner OT plan. Completing the square shows that the Euclidean transport term penalizes each local support's deviation from a plan-weighted target expressed in template coordinates. We add this identity to clarify the existing learning signal. This mechanism provides soft alignment through training. It does not guarantee exact feature-axis alignment or identify a unique coordinate transformation.

The server-side GW subproblem serves a different purpose. Specifically, it aggregates within-manifold squared-distance matrices. With uploaded matching weights fixed, this relational objective is invariant to independent client rotations. Eq. (79) then performs a separate and coordinate-dependent reconstruction by using the uploaded support vectors. This step relies on the compatibility encouraged by local calibration. Consequently, the reconstruction is not guaranteed to remain invariant to independent client rotations. It is also not guaranteed to exactly realize the optimized relational matrix.

Therefore, we have restricted the invariance claim to the GW relational subproblem, as revealed in our revised paper. In addition, we explicitly state the limitation of support reconstruction.

Reviewer_suU6OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies subgraph-level graph federated learning, GFL, for node classification under non-IID client data, where heterogeneity is decomposed into semantic heterogeneity of node feature and label distributions and structural heterogeneity of topology. The authors argue that existing methods, which align prototypes, parameters, or spectral characteristics toward a global consensus, implicitly force identical local representations and compress each client's personalized representation space. They propose FedGMC, a dual manifold calibration framework. For semantics, the server maintains a simplex ETF of equidistant class anchors; each client builds a local semantic manifold from class means of structure-free ego-embeddings and calibrates it via an orthogonal Procrustes rotation, which provably preserves pairwise distances. For structure, each client samples B ego nodes, forms 2-by-d radial sequence matrices from normalized 1-hop and 2-hop neighborhood embeddings, and softly matches them to Q global structural templates through entropic OT and Sinkhorn normalization; the server refines templates via a Gromov-Wasserstein barycenter. Global semantic anchors are refined with difficulty-weighted, constraint-regularized updates projected onto the hypersphere. The paper reports experiments on 11 datasets against 11 baselines under two partitioning schemes and varying client counts, 66 scenarios in total, plus ablations, convergence, sensitivity, efficiency, communication cost, geometry preservation, label-skew robustness studies, and convergence analyses in the appendix. FedGMC attains the best average accuracy in most settings.

Critical strengths
  1. Motivation: The problem is well motivated and clearly positioned. Heterogeneity in GFL is a real and active concern, and the critique that prototype or spectral alignment methods collapse client-specific geometry into a single consensus is a useful reframing. The distinction between "alignment", meaning forcing identical representations, and "calibration", meaning isometric matching to a shared reference, is conceptually appealing and the paper is explicit that the design goal is preserving local personalized geometry, which is then even measured quantitatively via PDC/GD in Appendix F.

  2. Method design: The overall design is coherent and technically clean in several places. The use of a simplex ETF as the global semantic reference connects to neural collapse theory; the orthogonal Procrustes solution is closed-form, exact, and comes with a genuine, if simple, distance-preservation guarantee in Proposition 1; and the division of labor between client-side OT matching in the shared embedding space and server-side GW barycenter refinement on intrinsic 2-by-2 metric matrices that are invariant to isometry is a thoughtful combination. Missing-class and imbalanced-label handling in Appendix G and the K-hop generalization in Appendix H show the framework was thought through beyond the default configuration.

  3. Experiments: The empirical effort is substantial: 11 datasets covering homophilic and heterophilic graphs, 11 baselines plus a local-only reference, non-overlapping and overlapping METIS partitioning, 5/10/20 and 10/30/50 clients, 10 independent runs with standard deviations, full ablations on all 11 datasets, convergence curves, hyperparameter sensitivity, per-round time cost, cumulative communication cost, and a severe label-skew stress test. An anonymous code repository is provided, and Appendix C gives reasonable implementation detail. The efficiency and communication analyses in Tables 7-9 address a dimension many GFL papers ignore.

Critical weaknesses
  1. The theoretical analysis is mostly decoupled from the actual algorithm. Theorem A.7 assumes deterministic gradient descent, yet Assumption A.3 on bounded stochastic gradient variance is never used in its proof; the proof is a standard descent-lemma argument. Theorem A.8 assumes unbiased Riemannian gradients and retraction smoothness for the anchor update, but the actual update rule in Eq. 8, which combines a deviation vector, a difficulty weight, a constraint vector, adaptive scaling, and renormalization, is never shown to satisfy these assumptions. Theorem A.9 invokes the Attouch-Bolte-Svaiter KL framework by assuming the sufficient decrease and relative error conditions rather than verifying them for the alternating Sinkhorn/GW procedure. None of the results model federated specifics such as client drift, partial participation, or cross-round coupling between local training and server refinement, so the theory adds limited explanatory power beyond what standard optimization texts already give.

  2. The privacy narrative is asserted, not analyzed. FedGMC uploads class-wise mean embeddings, per-class losses, and aggregated neighborhood embeddings every round, as detailed in Sections B.3 and C.6. Prototype leakage and property inference from exactly such statistics are known attack surfaces, yet the paper offers neither a formal privacy model nor an empirical leakage study; the claim in Appendix C.6 that the exchanged manifolds "reduce the exposure of private information" is qualitative. Relatedly, when a client observes very few classes, the orthogonal Procrustes problem over a small subset of anchors in Eq. 122 is severely underdetermined in d dimensions, and the paper does not discuss how the resulting rotational ambiguity affects calibration stability.

Questions for authors

Please refer to the weaknesses.

1 comment
AuthorsOpenReview ↗

To Reviewer suU6

Thanks for your positive comments!


1. Re: “W1”: Thanks! We revise our theoretical analysis by coupling it with the actual algorithm in four respects. First, we revise the proof of Theorem A.7 by using bounded stochastic gradient variance. In addition, client drift, partial participation, gradient bias, and changing references appear in the resulting bound, as revealed in Appendix A.1.3 of our revised paper. Second, we revise the proof of Theorem A.8 by considering the deviation vector, difficulty weight, constraint vector, adaptive scaling, and normalization, as revealed in Appendix A.2.1-A.2.2 of our revised paper. Third, we revise the proof of Theorem A.9 by distinguishing the alternating Sinkhorn/GW procedure descent/stationarity from a separately specified proximal variant, as revealed in Appendix A.3.1 of our revised paper. In addition, OT matching and GW refinement are analyzed as distinct subproblems. Fourth, we revise the theoretical results in Appendix A.2.2 and A.3.2 by modeling client drift, partial participation, heterogeneity, reconstruction error, and changes between rounds.


2. Re: “W2”: Thanks! The formal privacy model of our FedGMC is an honest-but-curious server threat model, as revealed in Appendix C.6 of our revised paper. In addition, our FedGMC provides barriers against the attacks of prototype leakage and property inference. This is because both the local encoder Wmego\mathbf{W}_m^\mathrm{ego} and the orthogonal transformation Rm\mathbf{R}_m remain strictly private on client mm and are never uploaded. Since the attack of prototype leakage or property inference requires inverting a many-to-one average without knowledge of either Wmego\mathbf{W}_m^{ego} or Rm\mathbf{R}_m, it is an underdetermined blind inverse problem. Furthermore, uploaded radial matrices are computed over BB randomly resampled anonymous nodes without node IDs or edge indices, and undergo ℓ2\ell_2-normalization as in Eq. 6, which does not leak raw structural information.

We add Appendix H.2 to discuss how the resulting rotational ambiguity affects calibration stability. As revealed in Appendix H.2 of our revised paper, we specify when rotational ambiguity is harmless and when additional stability control is needed.

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

The paper proposes FedGMC for semantic and structural heterogeneity in graph federated learning. Semantic calibration uses ETF class anchors and client-specific orthogonal transformations. Structural calibration matches multi-hop neighborhood representations to global templates through optimal transport. The server further refines the semantic anchors and structural references, with the latter formulated through GW barycenter optimization. Experiments report improvements across eleven datasets.

Critical strengths
  1. The design of the methods is intuitive, yet retains theoretical elegance. The proposed methods, such as orthogonal transformations, are both reasonable and well-suited to the original intent of this paper.

  2. The experiments are extensive, and the results are state-of-the-art.

Critical weaknesses
  1. Orthogonal transformations preserve the distances between existing representations, but Equation (3) pulls similar nodes toward the same anchor through training. This personalized geometric preservation actually creates a contradiction between orthogonal transformations and anchor approximation.

  2. The client uses Euclidean distance for OT; the server uses GW to compare internal distances. What causes this discrepancy between the local and global metrics? More importantly, Euclidean distance is dependent on coordinate directions, whereas GW is not. Therefore, could this discrepancy in metrics lead to the coordinates being incomparable?

  3. Eq. (8) projects onto the unit sphere, ensuring only that the length of each anchor point is 1; it does not guarantee that any two anchor points will still form the same angle.

  4. When a local client samples B nodes, the sampling method should affect whether the local structural reference is representative. However, the authors do not discuss this.

Questions for authors
  1. How does minimizing Eq. (3), which pulls nodes of the same class toward a shared anchor, preserve personalized local geometry beyond the distance-preserving property of the orthogonal transformation itself?

  2. Why does client-side calibration use Euclidean OT while server-side refinement uses GW? How are structural embeddings from different clients made comparable in a common coordinate system, particularly when averaging their support vectors in Eq. (105)?

  3. What ensures that the anchor update in Eq. (8) preserves the ETF equiangular property beyond unit norms? Is maximal equidistance guaranteed, or merely encouraged by the repulsion term?

4.How are the BB local nodes selected, and are they resampled each round? What makes this sampling strategy appropriate for representing heterogeneous local structures, including less frequent patterns?

  1. Pseudocode for an algorithm would be extremely helpful in understanding the overall process of federated training on the client and server sides.

  2. Could the authors explicitly define the AGGR and COMBINE operations in Eqs. (4) and (5)? In particular, what does COMBINE do with the single input in Eq. (4), and how does it integrate the two neighborhood summaries in Eq. (5)?

3 comments
AuthorsOpenReview ↗

To Reviewer qabz

Thank you for your comments! Our explanation and clarification can be found as follows.


Re: “Q1”: Thanks! In Eq. (3), RmR_m dynamically aligns the orientation of the local prototype PmP_m with the global anchor Δ\Delta via a closed-form solution obtained through SVD. Therefore, RmR_m in Eq. (3) preserves the personalized local geometry beyond the distance-preserving property of the orthogonal transformation itself.


Re: “Q2”: Thanks! Client-side calibration and server-side refinement are two different types of alignment. Client-side calibration is performed against explicit template vectors broadcast by the server. Therefore, Euclidean OT pulls local representations toward the same template coordinates. In contrast, server-side refinement compares within-manifold distances, which capture the separation between the 1-hop and 2-hop embeddings. GW therefore aggregates radial geometry independently of the absolute orientation of each local embedding. Support vectors in Eq. (105) are explicitly transformed into an identical coordinate system before being used, so structural embeddings from different clients are directly comparable.


Re: “Q3”: Thanks! We clarify that the anchor update in Eq. (8) does not preserve the ETF equiangular property. In contrast, by combining the deviation update vector, the difficulty weight, and the manifold constraint vector, the property of maximal equidistance is guaranteed, as revealed in Section 4.4.1 of our paper.


Re: “Q4”: Thanks! In each round, each client uniformly samples BB local nodes without replacement from the entire local node set Vm\mathcal{V}_m. They are resampled in each round. This sampling strategy effectively captures heterogeneous and less-frequent local structural patterns through two mechanisms. First, BB local nodes are resampled in each round, and the server initializes the GW barycenter optimization from the previous round's global templates TqT_q. Structural patterns are accumulated over RR rounds, therefore covering up to R×BR \times B ego-subgraphs per client. Therefore, heterogeneous local structure patterns are progressively integrated and retained in the global structural templates. Second, the column constraint F⊤1B=βF^\top \mathbf{1}_B = \beta prevents majority structural patterns from dominating all templates, ensuring that less-frequent local structures are preserved by dedicated templates during the weighted GW barycenter update.


Re: “Q5”: Thanks! We provide the pseudocode for FedGMC on the client side and server side in Appendix D of our revised paper, respectively.


Re: “Q6”: Thanks! In Eqs. (4) and (5), AGGR is defined as AGGR({hm,uego:u∈N1(b)})=∑u∈N1(b)hm,uego∣N1(b)∣∣N1(u)∣\mathrm{AGGR}(\{\mathbf{h}_{m,u}^{\mathrm{ego}}:u \in \mathcal{N}_1(b)\})=\sum_{u \in \mathcal{N}_1(b)}{\frac{\mathbf{h}_{m,u}^{\mathrm{ego}}}{\sqrt{|\mathcal{N}_1(b) | |\mathcal{N}_1(u) |}}}, which performs degree-normalized neighborhood aggregation over the structure-free ego-embeddings. For ease of representation, we denote AGGR({hm,uego:u∈N1(b)})\mathrm{AGGR}(\{\mathbf{h}_{m,u}^{\mathrm{ego}}:u \in \mathcal{N}_1(b)\}) as am,b(l)\mathbf{a}_{m,b}^{(l)}. Therefore, COMBINE is defined as COMBINE({am,b(l)}l=1k)=1k∑l=1kam,b(l)\mathrm{COMBINE}(\{ \mathbf{a}_{m,b}^{(l)} \}_{l=1}^{k})=\frac{1}{k}\sum_{l=1}^{k}\mathbf{a}_{m,b}^{(l)}, which integrates the hop-wise summaries via element-wise mean pooling. Since there is a single 1-hop summary in Eq. (4), COMBINE naturally reduces to the identity mapping. We retain the COMBINE notation in Eq. (4) for notational consistency with Eq. (5). In Eq. (5), COMBINE integrates the two neighborhood summaries via element-wise mean pooling.

Reviewer_qabzOpenReview ↗

Reply

Thank you for your detailed response. I have reviewed the revised main text section by section, and most of my questions have been addressed.

At the same time, I used an LLM to review the supplementary materials section. Regarding Q2, the LLM pointed out that Eq. (79) averages the client’s support vectors, and therefore cannot identify the explicit coordinate transformation described in the response. GW coupling establishes a correspondence between support points, but their feature axes may still differ. After checking, I agree with the LLM’s assessment. Therefore, I have raised my rating to 3. I hope the author will further address my concerns.

AuthorsOpenReview ↗

Thank you for your review and feedback!

Dear Reviewer qabz,

Thank you very much for your careful re-evaluation and for increasing your rating. We sincerely appreciate your constructive feedback and valuable suggestions regarding Q2. We will further examine the concern about Eq. (79) and clarify the distinction between GW-based correspondence and explicit coordinate alignment in the revised manuscript. Thank you again for your time and thoughtful assessment.

Best regards,

Authors of Paper ID 12771

Comments on the paper

1 comment

Dear Reviewer qabz and suU6,

The authors have responded to your concerns. Please review their responses as soon as possible and provide any necessary feedback. Thank you.

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.