acceptodds
Under review as a conference paper at ICLR 2027

Interweaving Topology and Spectrum: Federated Spectral-Guided Riemannian Graph Learning

Abstract

Federated Graph Learning (FGL) enables collaborative model training on distributed graph data while preserving privacy. However, structural and frequency heterogeneity across clients induces two critical challenges: local embedding distortion in flat Euclidean spaces, and global knowledge collapse caused by entangling disparate spatial topologies with spectral features. To address these challenges, we propose Federated Spectral-Guided Riemannian Graph Learning (FedSGR), a novel framework that systematically interweaves spatial topological structures with frequency domain signals. Specifically, FedSGR eliminates embedding distortion by establishing personalized product manifolds with mixed curvature tailored to local topologies. Concurrently, it circumvents knowledge collapse by distilling domain invariant frequency patterns globally, which are adaptively fused with local curvature features to steer personalized message passing directly on mixed manifolds. Extensive experiments on various graph benchmarks demonstrate that FedSGR consistently surpasses state-of-the-art FGL methods, exhibiting superior predictive accuracy and robust personalized adaptation.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Federated Graph Learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_NKFkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FedSGR, a federated graph learning framework that integrates spatial topological information with frequency-domain signals to address non-IID graph data. To achieve this, the authors first introduce local geometric enhancement, allowing each client to construct personalized product manifolds combining Euclidean, hyperbolic, and spherical spaces. They then share global spectral knowledge by computing local graph Laplacian spectra and obtaining frequency weights through a spectral encoder and a shared filter generator. Finally, they propose spectral-guided dual-gated Riemannian convolution, which combines local Forman–Ricci curvature signals with global spectral guidance to determine neighbor-aggregation weights and perform message passing on the product manifolds. The method outperforms the compared FGL methods across its experimental settings, but it has some flaws in novelty, experiments, and the theory part.

Critical strengths

The method outperforms the compared FGL methods across its experimental settings

Critical weaknesses

I appreciate the work and FedSGR's performance, so with clarifications on these aspects in the authors’ response, I would be willing to increase the score.

Regarding novelty, this paper’s main contribution seems to be combining existing techniques. FedRGL has already discussed personalized mixed-curvature spaces and Ricci-gated convolution, while FedSSP has studied globally shared spectral knowledge with client-specific personalization. Although combining these techniques may be useful and achieve SOTA, the experiments do not sufficiently establish the proposed integration's independent contribution.

In particular, FedRGL is not included as a baseline. The w/o S ablation cannot replace this comparison because it may not faithfully implement FedRGL. Similarly, the w/o G ablation replaces the entire message-passing stack, so it cannot independently demonstrate the effectiveness of spectral–Ricci fusion over existing Ricci gating. This makes it difficult to determine how much of the performance improvement comes from the proposed fusion rather than the existing components.

The theoretical analysis contains an issue. Proposition 1 claims that global spectral guidance reduces cross-client gradient divergence. However, the derivation in Equation (22) omits the cross term when bounding the squared norm of the combined spatial and spectral gradient deviations. The final cross term cannot be omitted without additional assumptions. Also, the Lipschitz continuity of the spectral filter generator does not by itself establish that spectral gradient divergence is significantly smaller than spatial divergence. Therefore, the theoretical guarantee in Theorem 1 is not established by the current proof.

Questions for authors

Some details in the experiment need clarification:

  1. Section 4.5 describes the fusion weights (w_s) and (w_g) as client-specific learnable scalars. However, Appendix E.2 states that the default setting is (w_s=w_g=0.5), while Appendix E.3 introduces learnable weights separately as FedSGR-LS. This should be clarified: whether the main results in Table 1 use fixed or learnable weights.
  2. The reported results differ between Table 1 and Appendix Table 9 under the 10-client setting. One mismatch is that Table 1 reports PubMed accuracy of 72.30%, while Table 9 reports 85.21%.
Reviewer_7BJKOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies federated graph learning, or FGL, under non-IID client data, arguing that heterogeneity in graph topology and in spectral frequency distributions creates two problems: local embedding distortion when client graphs are forced into flat Euclidean space, and global knowledge collapse when spatial topology and spectral features are entangled during federated sharing. The proposed framework, FedSGR, has three parts. First, each client builds a personalized product manifold, Euclidean times hyperbolic times spherical, with client-specific learnable curvatures, and a geometric contrastive InfoNCE loss aligns Euclidean and curved views of node features. Second, a shared spectral filter generator is trained on sinusoidally encoded Laplacian eigenvalues; only these spectral filter parameters are aggregated at the server, and an edge-level spectral gate is reconstructed through the inverse graph Fourier transform. Third, a dual-gated Riemannian convolution fuses a Forman-Ricci curvature gate with the spectral gate and performs message passing directly on the product manifold with Möbius operations. A convergence analysis covering client drift contraction in Proposition 1 and a SCAFFOLD-style rate in Theorem 1 is given in the appendix. Experiments on six datasets with eight baselines, IID and Dirichlet non-IID partitions, ablations, efficiency measurements, and a learnable gate variant, FedSGR-LS, report consistent improvements of roughly 0.7% to 4.6% over the strongest baseline.

Critical strengths
  1. Motivation: The problem is well motivated and clearly positioned. Structural and frequency heterogeneity across clients is a genuine and timely pain point in FGL, and the decomposition into local embedding distortion plus global knowledge collapse, shown in Figure 1, gives the paper a clean narrative. The gap is articulated concretely: spectral FGL methods such as FedSSP stay in flat Euclidean geometry, while Riemannian GNNs assume centralized data or a uniform geometry. The idea of sharing only frequency-domain knowledge while keeping geometry client-local is an attractive privacy and personalization story.

  2. Method design: The overall design is coherent and each module maps to one of the two stated challenges. Restricting server aggregation to the spectral filter parameters while keeping curvatures, projection weights, and gate scalars strictly local as described in Sections 4.1 and 4.4 is a sensible decoupling, and the Ricci gate plus spectral gate fusion in Eqs. 10-13 is a reasonable mechanism to reconcile global and local signals. The framework is also supported by a convergence analysis, an algorithm box, a notation table, complexity discussion, and a limitations section, which makes the paper comparatively complete and easy to follow at the system level.

  3. Experiments: The empirical coverage is broad: six datasets from three domains, eight baselines spanning generic FL, namely FedAvg and FedProx, and recent FGL methods, namely GCFL, FED-PUB, FedSSP, FedIIH, FedSPA, and FedAUX, client counts from 5 to 50, Dirichlet alpha sweeps, component ablations in Table 3 that support the claimed contribution of each module, communication and efficiency analysis in Tables 6-7, gate-weight sensitivity analysis in Table 8, and a learnable-scalar variant in Tables 9-10. Reporting rounds-to-target-accuracy is a useful and uncommon addition.

Critical weaknesses
  1. The key theoretical step is asserted rather than proven. Proposition 1 hinges on delta^2_spec << sigma^2_spat, but the "proof" just states that operating on 1D frequency responses makes the spectral divergence strictly smaller. Moreover, the move from Eq. 21 to Eq. 22 drops the cross-covariance term between spatial and spectral gradient deviations without stating an uncorrelatedness assumption; a plain Cauchy-Schwarz step would introduce a factor of 2 and cross terms. Theorem 1 then inherits its entire improvement over standard analysis from this asserted contraction, so the theory currently does not deliver the "rigorous mathematical justification" the remark after Theorem 1 claims.

  2. Scalability of the spectral module. The method requires exact eigendecomposition per client at order N^3 cost, acknowledged in Appendix F, plus storing the full eigenvector matrix at order N^2 memory, and Eq. 8 reconstructs the spectral gate per edge by summing over all N eigenpairs, which is order N per edge per round. The claim that this is fine because it is a one-time pre-processing step conflicts with Algorithm 1 line 22, where spectral gates are regenerated each round from the updated global filter. Realistic federated graphs, e.g., ogbn-products-scale clients and beyond, make this a first-order concern, not a footnote.

Questions for authors

Please refer to the weaknesses.

Reviewer_i7JVOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FedSGR, a personalized federated graph learning framework combining client-specific mixed-curvature representations with globally shared spectral filters. Spectral guidance is fused with local Ricci-curvature signals through a dual-gated mechanism for Riemannian message passing. The approach aims to mitigate local embedding distortion and cross-client knowledge conflict. Experiments cover six graph datasets under different client counts and data heterogeneity settings.

Critical strengths
  1. The paper addresses a relevant problem: accommodating heterogeneous graph structures while enabling useful knowledge sharing across clients.
  2. Keeping geometric parameters local while sharing spectral parameters provides an intuitive division between personalization and collaboration.
  3. The evaluation includes several competitive baselines, component ablations, convergence curves, and computational and communication cost analyses.
Critical weaknesses
  1. The connection between the motivation and the proposed mechanisms is insufficiently established. The paper attributes its improvements to eliminating embedding distortion and preventing knowledge collapse, but neither phenomenon is clearly quantified. The geometric contrastive objective encourages agreement across views without explicitly ensuring preservation of graph distances. Similarly, sharing spectral filter parameters does not by itself establish domain invariance across graphs with different spectra, features, and labels. The current evidence therefore supports the usefulness of the overall architecture more strongly than the proposed explanation for its effectiveness.

  2. Label-based Dirichlet partitioning does not independently establish robustness to topological or spectral heterogeneity. Moreover, removing the spectral module also removes the framework’s cross-client sharing mechanism, while removing the local convolution eliminates neighborhood message passing altogether. These ablations cannot clearly distinguish the benefits of spectral sharing and dual gating from those of collaboration and message passing more generally.

Questions for authors
  1. The main text defines the fusion weights as learnable, whereas Appendix E.3 introduces learnable weights as a separate FedSGR-LS variant. On PubMed, Table 1 reports FedSGR accuracy of approximately 71–72%, while Table 6 reports reaching 80% within three rounds and Tables 8–9 report approximately 85–86%. Different experimental settings could explain these discrepancies, but those settings need to be explicitly distinguished.

  2. Fewer communication rounds do not consistently translate into lower total runtime in Table 6, please justify a little bit more about this.

Comments on the paper

1 comment

Dear reviewer i7JV and 7BJK,

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.