- 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
- 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 to under non-overlapping partitions, while the average heterophilic result improves from to . Under overlapping partitions, the corresponding gains are and 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.
- 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 for FedGMC, compared with for FedIIH. These measurements provide useful empirical support beyond the main predictive-performance tables.
Critical weaknesses
- 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 and 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
- 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.
- 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 for . 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.
- 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.
- 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 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
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 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.
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
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.
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.. 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 C−1. 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 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.