- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper addresses the limitations of existing graph neural networks in modeling edge-level interactions and adaptive spectral patterns in brain connectomes. To address these challenges, the authors propose a Mixture of Wavelets framework, incorporating a Spectral Simplicial Wavelet Transform (SSWT) to jointly model multi-scale node and edge features and a Scale-aware eXpert (ScaleX) module to adaptively select informative wavelet scales. Theoretical analyses establish node–edge spectral relationships and orientation equivariance. Experimental results demonstrate superior classification performance, improved generalization, and promising interpretability across multiple neuroimaging datasets.
Critical strengths
-
Meaningful research problem. The paper addresses the limitations of existing graph neural networks in jointly modeling node- and edge-level interactions and capturing heterogeneous spectral patterns.
-
Theoretically grounded framework. The proposed SSWT extends spectral graph wavelets using Hodge Laplacians, providing a mathematical foundation for joint node–edge spectral analysis.
-
Comprehensive experimental evaluation. Experiments across multiple neuroimaging datasets demonstrate competitive performance, supported by ablation studies and external-cohort evaluations.
Critical weaknesses
-
Both Hodge Laplacian-based spectral analysis and mixture-of-experts mechanisms have been explored previously. The authors should further clarify the methodological novelty of their integration.
-
Although ScaleX improves classification performance, further analysis is needed to demonstrate whether the selected wavelet scales correspond to meaningful spectral characteristics of brain networks.
-
Spectral decomposition and adaptive expert routing may introduce additional computational costs. Further analysis of scalability on larger graphs would strengthen the practical value of the proposed method.
Questions for authors
-
Although important brain regions and connections are identified, their relationships with established neurological findings require more systematic validation.
-
The authors should include more recent state-of-the-art brain network learning methods to strengthen the experimental evaluation, such as:
[1] Interpretable Dynamic Brain Network Analysis with Functional and Structural Priors, IEEE TMI, 2025.
[2] NeuroH-TGL: Neuro-Heterogeneity Guided Temporal Graph Learning Strategy for Brain Disease Diagnosis, NeurIPS, 2026.
- The spectral representations depend heavily on the underlying graph topology. The authors should evaluate the robustness of the proposed method under different connectivity thresholds or graph sparsity levels.
7 comments
Response to Reviewer MSn4 (2/5)
W2. Spectral Relevance of Learned Wavelet Scales.
We thank the reviewer for this suggestion. We examined whether the wavelet scales selected by ScaleX align with disease-discriminative spectral bands identified from the data independently of the model.
Setup. We divided the non-zero graph spectrum into 10 quantile bands (, from lowest to highest frequency), with boundaries computed from the pooled eigenvalues of the training graphs in each fold, as in the scale initialization (App. E.2). We then computed two profiles for node and edge signals:
-
Discriminative profile : The Kruskal–Wallis effect size () across diagnostic groups of band energies (normalized by each subject’s total non-zero spectral energy), computed using training subjects only.
-
Selection profile : The squared wavelet responses of the selected top- experts, weighted by their routing weights and evaluated on held-out subjects.
We measured their alignment using Spearman correlation, reporting the mean ± std over 5 folds (averaged over 10 seeds per fold). Significance was assessed with a one-sided permutation test using 1,000 label permutations within the training subjects of each fold, with the run-averaged as the test statistic.
Table R10. Alignment between learned wavelet scales and disease-discriminative spectral bands.
(Spearman (mean ± std; higher is better). Permutation -values are reported for Learned.)
| Dataset | Signal | Init. (post hoc) | Fixed at Init. | Uniform top- | Learned | Perm. |
|---|---|---|---|---|---|---|
| ADNI-CT | Node | 0.31 ± 0.17 | 0.44 ± 0.14 | 0.40 ± 0.16 | 0.76 ± 0.10 | 0.012 |
| Edge | 0.29 ± 0.15 | 0.37 ± 0.18 | 0.51 ± 0.13 | 0.69 ± 0.12 | 0.031 | |
| ADNI-FDG | Node | 0.42 ± 0.13 | 0.35 ± 0.19 | 0.57 ± 0.11 | 0.81 ± 0.08 | 0.006 |
| Edge | 0.25 ± 0.18 | 0.46 ± 0.15 | 0.39 ± 0.17 | 0.72 ± 0.11 | 0.022 | |
| PPMI | Node | 0.33 ± 0.16 | 0.40 ± 0.15 | 0.48 ± 0.14 | 0.74 ± 0.11 | 0.014 |
| Edge | 0.38 ± 0.17 | 0.31 ± 0.20 | 0.43 ± 0.16 | 0.71 ± 0.13 | 0.025 |
Init. (post hoc) replaces only the kernel scales with their initial values while keeping the routing weights computed by the learned model unchanged; Fixed at Init. uses the separately trained fixed-at-initialization model (App. F.4);
Uniform top- assigns equal weights to the selected experts; and Learned uses the full ScaleX model.
Findings. The learned scales show strong alignment with disease-discriminative spectral bands across all datasets and both signal types (–, ), consistently exceeding Init. (post hoc) (–), Fixed at Init. (–), and Uniform top- (–). These results suggest that both scale learning and adaptive routing contribute to emphasizing diagnostically informative spectral components. Across datasets and both signal types, the strongest differences in normalized spectral energy occur in intermediate-to-high frequency bands (–), peaking around –. These bands reflect relatively rapid signal variations across connected brain regions, complementing the global, low-frequency information preserved by the low-pass branch.
We further examined the learned scales of the experts shown in Fig. 4 (ADNI-CT). The expert predominantly selected for AD has a mean scale of across nodes, corresponding to a kernel peak frequency of (within , among the most discriminative bands), whereas the expert predominantly selected for CN has a mean scale of (, within ). Together, these findings suggest that ScaleX learns class-dependent spectral preferences rather than arbitrary scale assignments.
Response to Reviewer MSn4 (3/5)
W3. Computational cost and scalability.
We clarify the computational costs of spectral decomposition and expert routing, and further evaluate scalability using an efficient implementation that avoids explicitly constructing the full edge eigenbasis.
(a) Computational complexity and efficiency.
As discussed in App. F.7, separately decomposing the node and edge Hodge Laplacians requires operations, where and denote the numbers of nodes and edges. Since the two Laplacians share the same nonzero spectrum (Prop. 1), a single SVD of the incidence matrix provides both spectral bases, avoiding separate eigendecompositions. This decomposition is performed once per graph during preprocessing and reused throughout training and inference.
Expert routing operates in a low-dimensional space, with computational cost linear in the number of graph elements for a fixed number of experts. Moreover, top- routing evaluates only the selected experts, reducing training time by 49% and memory usage by 24% compared with dense expert evaluation (Tab. 11). On ADNI-CT, our model requires 0.10 s/epoch and 6.9 GB under full-batch training, comparable to GAT (0.09 s/epoch, 10.7 GB) and more efficient than BrainNetTF (0.16 s/epoch, 32.2 GB).
(b) Efficient spectral implementation.
Although the shared SVD avoids separate eigendecompositions, our original implementation explicitly constructs the full edge eigenbasis, including its null space. This requires memory and introduces substantial computational overhead as the number of edges increases.
To address this limitation, we implemented an equivalent formulation that retains only the nonzero spectral components. Specifically, we eigendecompose the node Laplacian and recover the edge eigenvectors associated with its nonzero eigenvalues as , where contains the corresponding positive singular values. The edge null-space projection is computed directly as , without constructing a separate null-space basis. This reduces the preprocessing complexity to and spectral basis storage from to , while preserving the exact spectral filtering formulation.
The two implementations are mathematically equivalent when all nonzero spectral components are retained. We further confirmed their numerical equivalence on our brain connectome datasets, with a maximum absolute output difference below .
(c) Scalability evaluation.
To evaluate scalability, we generated Erdős–Rényi random graphs (Erdős & Rényi, 1959) with 5% edge density and . We used the ADNI-CT model configuration with a batch size of four graphs on a single NVIDIA RTX 6000 Ada GPU (48 GB). Preprocessing time was measured per graph, while training time was measured per forward–backward step.
Table R11. Scalability comparison between full-SVD and efficient implementations.
| Full SVD prep. (s) | Efficient prep. (s) | Full SVD train (s/step) | Efficient train (s/step) | Full SVD memory (GB) | Efficient memory (GB) | ||
|---|---|---|---|---|---|---|---|
| 100 | 260 | 0.025 | 0.004 | 0.0079 | 0.0080 | 0.04 | 0.03 |
| 200 | 1,027 | 0.064 | 0.005 | 0.0079 | 0.0070 | 0.29 | 0.10 |
| 400 | 3,899 | 0.186 | 0.012 | 0.032 | 0.007 | 3.54 | 0.50 |
| 800 | 16,116 | 1.115 | 0.057 | OOM | 0.032 | OOM | 3.40 |
| 1,600 | 64,108 | Infeasible | 0.544 | OOM | 0.217 | OOM | 24.89 |
The efficient implementation reduces preprocessing time by 6–20$\times$ and achieves up to 4.6$\times$ faster training with 7.1$\times$ lower peak GPU memory where both implementations are feasible. The full-SVD implementation runs out of memory during training at , and its decomposition becomes infeasible at . In contrast, the efficient implementation scales to , requiring 0.217 s/step and 24.89 GB of GPU memory.
Since the computational and memory requirements increase with both and , the largest setting in Tab.R11 (, ) provides a reference for graphs with fewer nodes and edges under the same model configuration. This covers commonly used brain parcellations at full density, including Desikan–Killiany (68 regions), AAL (116), Destrieux (160), and Brainnetome (246). Glasser (360), Schaefer-400, and Schaefer-1000 are covered up to approximately 99%, 80%, and 12.8% edge density, respectively.
Reference: Erdős, Paul, and Alfréd Rényi. "On random graphs I." Publicationes Mathematicae Debrecen 6 (1959): 290–297.
Response to Reviewer MSn4 (1/5)
We thank the reviewer for the positive assessment and constructive suggestions. In response, we (i) clarify the methodological novelty of integrating Hodge spectral analysis with sparse expert routing (W1), (ii) demonstrate that learned wavelet scales align with disease-discriminative spectral bands (W2), (iii) introduce an efficient spectral implementation that scales to graphs with approximately 64K edges (W3), (iv) evaluate anatomical concordance and attribution stability (Q1), (v) compare against two recent baselines, ST2GCN and NeuroH-TGL (Q2), and (vi) assess robustness to graph sparsity and connectivity thresholds (Q3).
We hope these responses address the reviewer's concerns. We welcome any further questions or discussion.
We acknowledge that Hodge Laplacian spectral analysis and sparse MoE are established techniques; our contribution lies in how they are integrated.
A key technical challenge in this integration is preserving orientation equivariance when routing and mixing edge representations. Since edge orientations on undirected graphs are arbitrary, routing directly on oriented edge signals or using non-odd activations can violate orientation equivariance. ScaleX addresses this through orientation-invariant routing on ωˉ, orientation-equivariant spectral filtering (Prop. 2) with an odd activation (tanh), and alignment of edge embeddings to a reference orientation before readout (Sec. 4.4).
Beyond orientation consistency, ScaleX defines experts as learnable sets of element-wise wavelet scales rather than independent feature transformations. Routing thus determines the spectral resolutions emphasized at individual nodes and edges, coupling adaptive expert selection with multi-scale Hodge spectral analysis.
This distinguishes ScaleX from prior work on both sides of the integration. Existing graph MoE methods define experts as separate modules operating on node representations, such as message-passing layers with different hop sizes in GMoE (Wang et al., 2023) or embeddings in Riemannian spaces of different curvatures in GraphMoRE (Guo et al., 2025). In contrast, ScaleX experts are spectral resolutions parameterized by element-wise wavelet scales and applied to both node and edge signals through a shared incidence-based spectrum (Prop. 1). Conversely, Hodge-based spectral methods such as Hodgelets (Roddenberry et al., 2022) construct wavelet dictionaries with fixed scales, without learnable or element-wise scale selection.
These design choices are supported by controlled ablations. Tab. 7 (App. F.3) shows the benefit of learned routing over fixed routing, random routing, and uniform aggregation, while Tab. 8 (App. F.4) shows consistent gains from learning wavelet scales under identical architecture and routing settings.