acceptodds
Under review as a conference paper at ICLR 2027

Mixture of Wavelets on Simplicial Complex for Analysis of Brain Connectome with Neurodegeneration

Abstract

Comprehensive analysis of brain networks is essential for understanding neurodegenerative disease progression and their early diagnosis. While various Graph Neural Networks (GNNs) have shown promise in identifying disease-related biomarkers, most existing approaches focus primarily on node-level features, overlooking crucial edge-level interactions. Furthermore, many spectral GNNs rely on fixed-bandwidth filtering, limiting their ability to capture diverse frequency components in graph-structured data. To address these limitations, we propose a novel framework that integrates two key components: 1) a Spectral Simplicial Wavelet Transform (SSWT) for jointly analyzing node and edge features, and 2) a Scale-aware eXpert on simplicial complex (ScaleX) for adaptive multi-scale filtering. Our approach utilizes richer spectral decomposition and dynamically selects the most relevant wavelet scales for each graph. Extensive experiments on benchmark brain network datasets demonstrate its effectiveness in improving classification and interpretability, highlighting its potential for facilitating the analysis of neurodegenerative diseases. The code will be released upon publication.

Primary area:
applications to neuroscience & cognitive science
Keywords:
neurodegenerative disease, spectral graph analysis

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_MSn4OpenReview ↗
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
  1. 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.

  2. Theoretically grounded framework. The proposed SSWT extends spectral graph wavelets using Hodge Laplacians, providing a mathematical foundation for joint node–edge spectral analysis.

  3. Comprehensive experimental evaluation. Experiments across multiple neuroimaging datasets demonstrate competitive performance, supported by ablation studies and external-cohort evaluations.

Critical weaknesses
  1. 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.

  2. Although ScaleX improves classification performance, further analysis is needed to demonstrate whether the selected wavelet scales correspond to meaningful spectral characteristics of brain networks.

  3. 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
  1. Although important brain regions and connections are identified, their relationships with established neurological findings require more systematic validation.

  2. 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.

  1. 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
AuthorsOpenReview ↗

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.

W1. Methodological novelty of the integration

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\bar{\omega}, orientation-equivariant spectral filtering (Prop. 2) with an odd activation (tanh⁡\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.

AuthorsOpenReview ↗

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 (B1,…,B10\mathcal{B}_1,\ldots,\mathcal{B}_{10}, 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 D(b)\mathcal{D}(b): The Kruskal–Wallis effect size (ϵ2\epsilon^2) across diagnostic groups of band energies (normalized by each subject’s total non-zero spectral energy), computed using training subjects only.

  • Selection profile S(b)\mathcal{S}(b): The squared wavelet responses Ks(λ)2\mathcal{K}_s(\lambda)^2 of the selected top-kk 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 ρ\rho as the test statistic.

Table R10. Alignment between learned wavelet scales and disease-discriminative spectral bands.

(Spearman ρ\rho (mean ± std; higher is better). Permutation pp-values are reported for Learned.)

DatasetSignalInit. (post hoc)Fixed at Init.Uniform top-kkLearnedPerm. pp
ADNI-CTNode0.31 ± 0.170.44 ± 0.140.40 ± 0.160.76 ± 0.100.012
Edge0.29 ± 0.150.37 ± 0.180.51 ± 0.130.69 ± 0.120.031
ADNI-FDGNode0.42 ± 0.130.35 ± 0.190.57 ± 0.110.81 ± 0.080.006
Edge0.25 ± 0.180.46 ± 0.150.39 ± 0.170.72 ± 0.110.022
PPMINode0.33 ± 0.160.40 ± 0.150.48 ± 0.140.74 ± 0.110.014
Edge0.38 ± 0.170.31 ± 0.200.43 ± 0.160.71 ± 0.130.025

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-kk 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 (ρ=0.69\rho=0.69–0.810.81, p≤0.031p\leq0.031), consistently exceeding Init. (post hoc) (0.250.25–0.420.42), Fixed at Init. (0.310.31–0.460.46), and Uniform top-kk (0.390.39–0.570.57). 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 (B4\mathcal{B}_4–B9\mathcal{B}_9), peaking around B6\mathcal{B}_6–B8\mathcal{B}_8. 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 sˉ=0.27\bar{s} = 0.27 across nodes, corresponding to a kernel peak frequency of λ∗=1/sˉ≈3.70\lambda^* = 1/\bar{s} \approx 3.70 (within B8\mathcal{B}_8, among the most discriminative bands), whereas the expert predominantly selected for CN has a mean scale of sˉ=1.41\bar{s} = 1.41 (λ∗≈0.71\lambda^* \approx 0.71, within B3\mathcal{B}_3). Together, these findings suggest that ScaleX learns class-dependent spectral preferences rather than arbitrary scale assignments.

AuthorsOpenReview ↗

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 O(N3+M3)O(N^3+M^3) operations, where NN and MM 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 B1B_1 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-kk 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 O(M2)O(M^2) 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 L0=UΛU⊤L_0=U\Lambda U^\top and recover the edge eigenvectors associated with its rr nonzero eigenvalues as Vr=B1⊤UrΣr−1V_r=B_1^\top U_r\Sigma_r^{-1}, where Σr\Sigma_r contains the corresponding positive singular values. The edge null-space projection is computed directly as Πker⁡L1ω=ω−VrVr⊤ω\Pi_{\ker L_1}\omega=\omega-V_rV_r^\top\omega, without constructing a separate null-space basis. This reduces the preprocessing complexity to O(N3+N2M)O(N^3+N^2M) and spectral basis storage from O(M2)O(M^2) to O(NM)O(NM), 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 3×10−73\times10^{-7}.

(c) Scalability evaluation.

To evaluate scalability, we generated Erdős–Rényi random graphs (Erdős & Rényi, 1959) with 5% edge density and N∈{100,200,400,800,1600}N\in\lbrace 100,200,400,800,1600\rbrace. 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.

NNMMFull SVD prep. (s)Efficient prep. (s)Full SVD train (s/step)Efficient train (s/step)Full SVD memory (GB)Efficient memory (GB)
1002600.0250.0040.00790.00800.040.03
2001,0270.0640.0050.00790.00700.290.10
4003,8990.1860.0120.0320.0073.540.50
80016,1161.1150.057OOM0.032OOM3.40
1,60064,108Infeasible0.544OOM0.217OOM24.89

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 N=800N=800, and its decomposition becomes infeasible at N=1,600N=1{,}600. In contrast, the efficient implementation scales to N=1,600N=1{,}600, requiring 0.217 s/step and 24.89 GB of GPU memory.

Since the computational and memory requirements increase with both NN and MM, the largest setting in Tab.R11 (N=1,600N=1{,}600, M≈64,000M\approx64{,}000) 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.

Reviewer_P58vOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a framework for brain connectome classification that processes signals defined on both nodes and edges. It combines spectral simplicial wavelet filtering with ScaleX, a sparse routing mechanism that selects learnable wavelet scales for individual nodes and edges. Experiments cover Alzheimer’s and Parkinson’s disease classification, with additional evaluation on ABIDE and two external Parkinson’s disease cohorts.

Critical strengths

1.The most interesting contribution is the signal–scale routing mechanism. Rather than applying one spectral scale throughout a graph, ScaleX selects different combinations of scales for individual nodes and edges. This is a useful extension of adaptive spectral filtering, particularly for connectomes where regional measurements and connection strengths carry complementary information. 2.The empirical evaluation is reasonably thorough. The paper includes repeated cross-validation, comparisons with several brain-specific and simplicial models, external-cohort testing, and ablations of signals, filters, routing strategies, and routing regularization.

Critical weaknesses

The experiments support the benefit of learned routing, but do not separately establish the contribution of learning the wavelet scales, which is presented as a key part of the method.

Questions for authors

Can the authors compare fixed and learnable wavelet scales while keeping the network architecture, number of experts, and routing mechanism unchanged?

3 comments
AuthorsOpenReview ↗

Response to Reviewer P58v

We sincerely thank the reviewer for the positive assessment and for recognizing the signal–scale routing mechanism as a useful extension of adaptive spectral filtering. The suggested comparison helped us clarify the independent contribution of learnable wavelet scales, and we have incorporated this additional ablation into the revised manuscript (App. F.4).

W1/Q1. Contribution of learnable wavelet scales.

To isolate the contribution of scale learning, we compared fixed and learnable wavelet scales while keeping the architecture, number of experts, top-kk routing, and training objective unchanged. Specifically, variant (a) freezes the element-wise band-pass scales at their initialization, sp,i=λ~p−1exp⁡(0.1ξp,i)s_{p,i}=\tilde{\lambda}_p^{-1}\exp(0.1\xi_{p,i}), where λ~p\tilde{\lambda}_p denotes the pp-th equally spaced quantile between the 15th and 85th percentiles of the nonzero eigenvalues of the training graphs, and ξp,i∼N(0,1)\xi_{p,i}\sim\mathcal{N}(0,1). This initialization is identical to that of our full model. In variant (b), each expert uses a single fixed scale sp,i=λ~p−1s_{p,i}=\tilde{\lambda}_p^{-1} shared across nodes (and analogously for edges).

Table R9. Fixed vs. learnable wavelet scales (accuracy, %).

Scale settingElement-wiseLearnedADNI-CTADNI-FDGPPMITaoWu (zero-shot)Neurocon (zero-shot)
(b) Fixed, expert-wise constant––88.9 ± 2.390.7 ± 2.079.8 ± 3.568.2 ± 5.269.4 ± 4.8
(a) Fixed at initialization✓–90.8 ± 1.992.4 ± 1.682.6 ± 3.074.3 ± 4.175.1 ± 4.0
Learned (ours)✓✓94.9 ± 1.396.4 ± 0.987.4 ± 2.181.8 ± 2.682.3 ± 3.6
(ref.) SSWT w/o ScaleX✓ (1 expert)✓92.5 ± 1.192.8 ± 0.784.6 ± 2.773.1 ± 2.974.7 ± 1.3

All ScaleX variants share the same architecture, number of experts, top-kk routing, training objective, hyperparameters, and data splits; only the band-pass expert scales differ. SSWT w/o ScaleX is included for reference, as its architecture differs. External cohorts are evaluated zero-shot using PPMI-trained models, as in Tab. 3.

The learned scales consistently outperform both fixed-scale variants across all datasets. Compared with (a), scale learning improves accuracy by 4.1, 4.0, and 4.8 pp on ADNI-CT, ADNI-FDG, and PPMI, respectively. These gains extend to the zero-shot external cohorts, with improvements of 7.5 and 7.2 pp on TaoWu and Neurocon. Moreover, (a) outperforms (b), suggesting that element-wise scale initialization provides additional benefits even without scale optimization. Together with the routing ablation in Tab.7, these results support the contributions of learnable wavelet scales and adaptive expert routing. We clarify that ``Fixed Routing'' in Tab.7 fixes expert assignments rather than wavelet scales.

We have added the controlled scale ablation to App. F.4 and the scale initialization details to App. E.2.

Thank you very much for the author's detailed response. It has addressed my concerns. I will maintain my original positive score.

We appreciate your thoughtful feedback and are pleased that our clarifications have resolved your concerns. Thank you again for your time and consideration.

Reviewer_4U1xOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a spectral framework for brain connectome classification that combines a Spectral Simplicial Wavelet Transform (SSWT) with a Scale-aware eXpert on simplicial complex (ScaleX). The node and edge pathways each contain multiple learnable wavelet-scale sets, each set acts as an expert, and a gating network computes signal compatibility scores, selects experts with Top-K, and aggregates them with Softmax weights. SSWT applies low-pass and multi-scale band-pass filtering to node and edge signals using Laplacians induced by the incidence matrix. The authors conduct five-fold cross-validation on ADNI-CT, ADNI-FDG, and PPMI, report ablations of node and edge inputs, routing strategies, and low-pass/band-pass branches, and perform zero-shot external tests on TaoWu and Neurocon using a classifier trained on PPMI.

Critical strengths

1.The paper studies node and edge signals in brain connectomes and attempts to select different wavelet experts through scale routing. The problem is meaningful and the motivation is reasonably clear. 2.The paper conducts main and ablation experiments on ADNI and PPMI, and adds external-cohort tests on TaoWu and Neurocon. Overall, the experimental coverage is relatively broad.

Critical weaknesses

Major concern: The classification results are of some value, but the manuscript does not sufficiently demonstrate the independent gains from the wavelet mixture and scale routing. The complete-reconstruction statement in Eq. (6) raises questions in relation to the band-pass kernel, and Eqs. (7), (8), and (11) do not clearly explain how the scale experts are computed in practice. Meanwhile, the PPMI minority class and the external cohorts are small, so the current results do not support strong claims about clinical generalization, disease progression, or stable biomarkers. Minor concern 1: The experiments mainly use a one-dimensional simplicial complex, and the node and edge branches are combined mainly at the readout stage. The paper should therefore use more cautious language when describing higher-order simplex interactions, cross-order representations, and clinical mechanisms. Minor concern 2: The preprocessing, model configuration, class-wise PPMI results, and uncertainty of the external tests require further reporting. These details, together with fair baseline settings, would help assess stability and enable independent reproduction.

Questions for authors
  1. Please explain how Eq. (6) should be corrected, and provide a clear computational flow showing how the multiple wavelet experts are selected and mixed.
  2. Please add baselines with inputs and capacity comparable to the proposed method, and report the reliability of the PPMI minority-class results and the TaoWu and Neurocon external tests.
  3. Please clarify the data preprocessing and model configuration, explain the stability of the attribution results, and limit the clinical conclusions to what the current cross-sectional classification evidence supports.
9 comments
AuthorsOpenReview ↗

Summary of Responses to Reviewer 4U1x

We thank the reviewer for the constructive feedback. We have addressed the concerns through additional controlled experiments, methodological clarifications, uncertainty analyses, and revisions to the scope of our claims. The key responses and corresponding manuscript revisions are summarized below.

ConcernOur responseEvidenceManuscript revisions
W1-(a) Independent contributions of wavelet mixture and scale routingWavelet mixture without learned routing provides modest gains (+0.2–0.6 pp), while learned routing (+2.0–2.5 pp over uniform/random aggregation) and learnable scales (+4.1–4.8 pp over frozen scales) yield larger improvements.Table R1App. F.3 (Tab. 7); App. F.4 (Tab. 8, added)
W1-(b) / Q1-(a) Reconstruction correctnessCorrected Eq. (6) with explicit null-space projections (CK=1/4C_\mathcal{K}=1/4) and clarified null-space preservation by the low-pass branch. Implementation and results remain unchanged.—Sec. 4.1 (Eq. 6); Sec. 4.3
W1-(b) / Q1-(b) Scale expert computationClarified shared SVD preprocessing, signal–scale top-kk routing, element-wise wavelet filtering, and weighted expert aggregation.Algorithm 1Sec. 4.3; App. E.1 (Algorithm 1, added)
Minor 2 / Q2-(a) Comparable baselinesSCNN and SaNN use matched node/edge inputs and readout, with parameter counts within 1% of ours. Our model (SSWT w/ ScaleX) outperforms both across all datasets.Table R2—
W1-(c) / Q2-(b) PPMI minority-class reliabilityAdded class-wise precision, recall, F1, bootstrap CIs, and a confusion matrix. CN has the lowest recall (68.0%, CI [44.0, 88.0]).Tables R3–R4—
W1-(c) / Q2-(b) External evaluation reliabilityAdded bootstrap CIs for Acc., Bal. Acc., and AUROC, and paired-bootstrap CIs for accuracy gains over AGT (+11.2 pp [5.1, 17.3] and +9.1 pp [2.8, 15.4]). Acknowledged uncertainty from small cohort sizes.Table R5Sec. 5.2
Minor 2 / Q3-(a) Preprocessing and model configurationConsolidated preprocessing, model configurations, and training protocols to improve reproducibility.Tables R6–R7Sec. 5.1; Apps. D–E
Q3-(b) Attribution stabilityAcross 50 runs, Grad-CAM shows consistent top-10 overlap (Jaccard 0.53–0.72) and rank agreement (Spearman ρ\rho 0.68–0.84); 7–9 of the top-10 elements in Tab. 2 appear in ≥80% of runs.Table R8—
Minor 1 / Q3-(c) Scope of simplicial modelingClarified the 1D simplicial setting, separate node/edge processing, shared non-zero spectrum, and readout-stage fusion. Removed wording implying explicit cross-order interactions; higher-order extensions remain unevaluated.Tab. 4Abstract; Intro; Related Work; Sec. 4; Conclusion; App. B
Minor 1 / Q3-(c) Scope of clinical claimsRevised claims concerning disease progression, early diagnosis, clinical interpretability, and biomarkers. Restricted conclusions to cross-sectional classification and described attributions as model-discriminative rather than clinically validated.—Abstract; Intro; Secs. 5.2–5.3; Conclusion; Apps. D.2, G

All additional analyses reported in this response (Tables R1–R8) will be included in the appendix of the final version. We hope these responses and additional analyses address the reviewer's concerns. We welcome any remaining questions or concerns and would be happy to discuss them further during the discussion period.

AuthorsOpenReview ↗

Response to Reviewer 4U1x (1/5)

W1-(a). Independent contributions of the wavelet mixture and scale routing.

We consolidate the relevant ablations from Tab. 1, Tab. 7 (App. F.3), and Tab. 8 (App. F.4) in Table R1. All variants share the same SSWT backbone, readout, and training protocol.

Table R1. Contributions of multi-scale mixture, routing, and learnable scales. Classification accuracy (%) reported as mean ± std.

VariantMulti-scale mixtureRoutingLearnable scalesADNI-CTPPMI
SSWT w/o ScaleX (Tab. 1)––✓92.5 ± 1.184.6 ± 2.7
Uniform Aggregation (Tab. 7)✓Uniform✓92.7 ± 1.185.2 ± 2.3
Random Routing (Tab. 7)✓Random✓92.9 ± 2.784.9 ± 3.3
Fixed at Init. (Tab. 8)✓Learned–90.8 ± 1.982.6 ± 3.0
ScaleX (full)✓Learned✓94.9 ± 1.387.4 ± 2.1

Wavelet mixture (Rows 1–3). Introducing multiple learnable wavelet experts without learned routing yields modest improvements over a single scale set (+0.2–0.4%p on ADNI-CT and +0.3–0.6%p on PPMI). This indicates that increasing the number of wavelet experts alone provides limited benefit.

Scale routing (Rows 2–3 vs. Row 5). With the same multi-scale mixture and learnable scales, learned routing improves accuracy over uniform aggregation by 2.2%p on both datasets, and over random routing by 2.0%p on ADNI-CT and 2.5%p on PPMI. This demonstrates the importance of adaptively selecting and weighting wavelet experts for individual nodes and edges.

Learnable scales (Rows 4–5). With the same learned-routing architecture, optimizing the wavelet scales improves accuracy by 4.1%p on ADNI-CT and 4.8%p on PPMI, supporting the benefit of optimizing the expert scales rather than freezing them at initialization. Notably, the fixed-scale variant (Row 4) underperforms SSWT w/o ScaleX (Row 1), which learns its scales with a single expert. This indicates that routing over fixed scales cannot compensate for the absence of scale learning, and that the two components are complementary rather than independently additive.

Overall, these results indicate that the wavelet mixture provides candidate spectral resolutions, while learnable scales and adaptive routing enable their effective utilization. The full ScaleX model achieves improvements of 2.4%p and 2.8%p over SSWT without ScaleX on ADNI-CT and PPMI, respectively.

W1-(b) / Q1-(a). Correction of the Reconstruction Formula.

We agree that the original Eq. (6) did not account for the null-space components of the Hodge Laplacians. Since the band-pass kernel satisfies K(0)=0\mathcal{K}(0)=0, the wavelet integral reconstructs only the non-zero spectral components. We have therefore corrected Eq. (6) by explicitly adding the orthogonal projections Πker⁡L0f\Pi_{\ker\mathcal{L}_0}f and Πker⁡L1ω\Pi_{\ker\mathcal{L}_1}\omega to the node and edge reconstruction formulas, respectively. The corrected formula reads

f=Πker⁡L0f+1CK∫0∞ψsWf(s)dss,ω=Πker⁡L1ω+1CK∫0∞ϕtWω(t)dtt.f=\Pi_{\ker\mathcal{L}_0}f+\frac{1}{C_\mathcal{K}}\int_0^\infty \psi_s W_f(s)\frac{ds}{s},\qquad \omega=\Pi_{\ker\mathcal{L}_1}\omega+\frac{1}{C_\mathcal{K}}\int_0^\infty \phi_t W_\omega(t)\frac{dt}{t}.

For the band-pass kernel K(x)=xe−x\mathcal{K}(x)=xe^{-x}, CK=1/4<∞C_\mathcal{K}=1/4<\infty, so the corrected formula is exact.

Importantly, our framework already incorporates a low-pass branch with Kslow(λ)=e−sλ\mathcal{K}^{\mathrm{low}}_s(\lambda)=e^{-s\lambda}, which preserves the null-space components since Kslow(0)=1\mathcal{K}^{\mathrm{low}}_s(0)=1. We have clarified this relationship in the revised manuscript. This correction does not affect the implementation or experimental results, as our model employs finite-scale spectral filtering rather than continuous inverse wavelet reconstruction.

AuthorsOpenReview ↗

Response to Reviewer 4U1x (2/5)

W1-(b) / Q1-(b). Clarification of Scale Expert Computation.

We have clarified the relationship among Eqs. (7)-(11) and added Algorithm 1 in Appendix E.1 (Computational Procedure of SSWT with ScaleX) to detail the complete forward computation. The key steps are summarized below.

1. Shared Spectral Preprocessing. For each subject, a single SVD of the incidence matrix B1=UΣV⊤B_1=U\Sigma V^\top provides the spectral bases of both node and edge Hodge Laplacians. This decomposition is performed once per subject, and the resulting spectral bases are shared across experts.

2. Signal-Scale Routing. Each node expert pp and edge expert qq learns node- and edge-specific scale vectors sp={sp,i}i=1N\mathbf{s}_p=\lbrace s_{p,i}\rbrace_{i=1}^{N} and tq={tq,e}e=1M\mathbf{t}_q=\lbrace t_{q,e}\rbrace_{e=1}^{M}, respectively. Eqs. (9)-(10) compute compatibility scores between each signal and its corresponding expert-specific scale, followed by Top-kk selection and softmax normalization. We have corrected the routing notation to explicitly use sp,is_{p,i} and tq,et_{q,e}.

3. Sparse Wavelet Filtering. Routing precedes wavelet filtering, so only the selected node-expert and edge-expert pairs are evaluated using the same scales employed during routing. The filtered features follow Eq. (7), applying K2\mathcal{K}^2 to the adaptive scales in Eq. (8), where the squared kernel arises from successive wavelet analysis and synthesis with a scale shared across nodes. Concretely, fsp,i=∑ℓUiℓK2(sp,iλ0,ℓ)f^ℓf_{\mathbf{s}_p,i}=\sum_{\ell}U_{i\ell}\mathcal{K}^2(s_{p,i}\lambda_{0,\ell})\hat{f}_\ell with f^=U⊤f\hat{f}=U^\top f, i.e., the ii-th entry of Eq. (7) evaluated with the shared scale set to the node-specific value sp,is_{p,i}; edge features are computed analogously with VV, Λ1\Lambda_1, and tq,et_{q,e}. Spectral coefficients are shared across experts, avoiding redundant spectral projections.

4. Weighted Expert Aggregation. The selected expert outputs are combined using their normalized routing weights, as defined in Eq. (11). Unselected experts have zero routing weights, so the aggregation effectively involves only the selected kk experts. The mixed features are subsequently combined with low-pass features for the final representation.

These clarifications are reflected in the revised ScaleX layer description, with the complete computational procedure provided in Algorithm 1 (Appendix E.1).

Minor 2 / Q2-(a). Baselines with comparable inputs and capacity.

The original baselines in Table 1 receive the same region-wise node features and connectivity matrix as our model, with differences in how connectivity is incorporated. The exceptions are SVM and MLP, which use node features without graph structure, and GAT, which uses the connectivity pattern but learns attention weights instead of directly using the original connectivity values.

SCNN and SaNN provide directly comparable baselines, as both process node and edge signals, including edge filtering through L1\mathcal{L}_1. In our experiments, we configured both methods with the same dataset-specific hidden dimensions and readout architecture as our model (as selected by the grid search in App. E).

Table R2. Comparison with simplicial baselines under matched inputs, readout, and capacity.

(Accuracy (%), mean ± std.)

ModelParameters: CT / FDG / PPMIADNI-CT Acc. (%)ADNI-FDG Acc. (%)PPMI Acc. (%)
SCNN480,999 / 320,865 / 579,10787.0 ± 1.192.3 ± 1.879.8 ± 0.7
SaNN480,837 / 320,805 / 577,71584.4 ± 2.192.1 ± 1.877.4 ± 2.9
Ours482,507 / 322,607 / 576,22194.9 ± 1.396.4 ± 0.987.4 ± 2.1

Both baselines are within 1% of our trainable parameter count on every dataset, yet our model consistently achieves higher accuracy under matched inputs and readout.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers for their constructive comments and thoughtful assessments of our work. We particularly appreciate the recognition of our signal–scale routing mechanism, which adaptively selects learnable wavelet scales for individual nodes and edges (P58v). We also thank the reviewers for recognizing the importance of modeling both node- and edge-level signals in brain connectomes (4U1x, P58v, MSn4), the theoretical grounding of SSWT through Hodge Laplacians (MSn4), and the broad experimental evaluation, including ablations and external-cohort tests (4U1x, P58v, MSn4).

In the following, we address all concerns and questions raised by the reviewers in detail. All corresponding revisions are highlighted in blue\color{blue}{\textbf{blue}} in the updated manuscript.

We hope that our responses and revisions resolve the remaining concerns. We welcome any further questions or feedback and would be happy to discuss them during the discussion period.

Thank you to the reviewer who has already followed up on the authors' responses. For reviewers who have not yet done so, please post a brief follow-up acknowledging the authors' responses within the next 24 hours, so the authors have time to respond before discussion closes. In your reply, please indicate which main concerns have been addressed, what remains, and whether your overall assessment changes or stays the same.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.