acceptodds
Under review as a conference paper at ICLR 2027

SPREAD: Spectral Response Graph Diffusion for Representation Learning

Abstract

Diffusion models have emerged as a promising paradigm for graph representation learning, which employ progressive perturbation and denoising to learn discriminative representations. However, existing methods rely on uniform or predefined diffusion mechanisms that overlook how heterogeneous spectral distributions influence the diffusion dynamics and recovery capabilities of different frequency components. To address this limitation, we propose SPREAD, a spectral response graph diffusion framework that jointly adapts forward perturbation and reverse reconstruction in accordance with graph spectral properties. In the forward process, we introduce a learnable spectral response function parameterized by Bernstein polynomials to regulate noise intensity across frequency components. In the reverse process, a frequency-aware spectral reconstruction objective weights spectral bands according to the learned spectral response profiles, guiding the recovery of informative frequency components. Theoretical analysis establishes that the proposed model can uniformly approximate arbitrary continuous spectral distribution as the polynomial degree increases. Experiments on node classification benchmarks demonstrate that SPREAD achieves competitive performance compared to baseline methods, with ablation studies further supporting the effectiveness of both proposed mechanisms in SPREAD.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Representation Learning, Diffusion Models, Learnable Spectral Response, Spectral Heterogeneity

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_BPoBOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a graph diffusion-based framework for representation learning that solve spectral heterogeneity in graph signals. To achieve that purpose, this work introduces a learnable spectral response strategy to adapt noise perturbation across different frequencies and a frequency-aware reconstruction objective to guide the denoising process. Experiments on node and graph classification tasks show competitive performance.

Critical strengths
  1. Using diffusion models for representation learning is an interesting research direction.

  2. The paper has a clear line of reasoning. This paper (Section 3 and Figure 1) provides a clear, quantitative analysis of an important and interesting motivation before introducing the proposed method.

  3. The ablation studies (Tables 2 and 6) provide consistent evidence that removing the proposed key components reduces performance on the corresponding tasks.

Critical weaknesses
  1. The current visualizations mainly show that the model improves the spectral reconstruction objective it explicitly optimizes. However, they do not clearly demonstrate whether the frequency-adaptive mechanism preserves more useful information for downstream tasks.

  2. Theorem 4.1 mainly relies on the classical approximation properties of Bernstein polynomials to establish their greater expressiveness compared to constant functions. But this does not necessarily imply that learning a noise response using this parameterization leads to better representations.

Questions for authors
  1. If high-frequency signals are already more susceptible to information loss than low-frequency signals, why does the learned spectral response further increase noise at higher frequencies? Could the authors explain the relationship between Figures 4(b) and 1(b)?

  2. I am curious why assigning reconstruction weights based solely on spectral response coefficients can reflect the reliability of different spectral bands. Authors should give more hints about the motivation behind this design.

  3. How are node features constructed for graph classification datasets that do not provide original node attributes?

2 comments
AuthorsOpenReview ↗

Response to Reviewer

We thank the reviewer for valuable comments.

W1: Figures 4(b) and 5 explain response allocation and band-wise recovery. Beyond these spectral analyses, downstream benefits are evaluated through classification: Table 2 shows that removing either the forward spectral response or adaptive reconstruction weights reduces accuracy. To further evaluate the frequency-adaptive mechanism, we added classification comparisons at matched noise power against tuned and learnable global scalar responses and fixed increasing and decreasing spectral profiles. The results below on Photo provide evidence independent of spectral reconstruction metrics, supporting the benefit of adaptive frequency allocation for learning task-relevant representations. Full results are provided in Appendix I.2 of the revised manuscript.

MethodTuned global scalarLearnable global scalarFixed increasing profileFixed decreasing profileSPREAD
Accuracy (%)95.06±0.2894.86±0.2395.16±0.3194.75±0.2995.69±0.16

W2: Section 3 shows that a uniform response only scales noise globally and cannot adjust relative SNRs across frequencies. In contrast, Bernstein parameterization can express nonmonotonic response profiles, allowing diffusion to adapt to the spectral heterogeneity of graph signals. Theorem 4.1 establishes the expressive capacity, while classification and ablation studies evaluate representation quality. Specifically, the matched-noise-power comparisons in W1 show that the learned response outperforms scalar and fixed profiles, thereby providing empirical evidence of improved graph representations.

Q1: Figures 1(b) and 4(b) describe different settings. Figure 1(b) shows frequency-dependent sensitivity under uniform perturbation, revealing that high-frequency information is more susceptible to loss. In contrast, Figure 4(b) shows the spectral responses learned through joint training to adapt to these spectral characteristics. Our goal is to learn useful representations through denoising, rather than simply minimize perturbation or equalize SNRs across frequencies. Specifically, weak perturbation may make reconstruction too easy, while excessive perturbation can destroy recoverable information. Therefore, we learn the allocation of perturbations across frequencies jointly with the reconstruction objective. Under this objective, greater high-frequency sensitivity does not imply that the learned response must decrease with frequency. Moreover, on Computer, the response in Figure 4(b) declines at higher frequencies, showing that the model can learn nonmonotonic allocations from data rather than enforcing an increasing profile.

Q2: This design allocates reverse reconstruction supervision according to forward perturbation strength. Specifically, Section 3 shows that, for fixed signal and base noise power, a smaller spectral response yields lower injected noise power and a higher SNR. Therefore, inverse-response weighting prioritizes less perturbed spectral bands and prevents heavily perturbed bands from dominating the reconstruction objective. However, since each band’s actual SNR also depends on its signal and base noise power, these weights represent response-guided supervision allocation rather than a direct estimate of spectral reliability. Through this allocation, the shared response allows reconstruction supervision to adapt jointly with forward perturbation, which motivates our design.

Q3: For graph classification datasets without original node attributes, we use one-hot encodings of node degrees as node features.

Reviewer_BPoBOpenReview ↗

Response to authors

Thank you for your response. This addresses my concerns, and I will maintain my positive rating.

Reviewer_tzUdOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

Existing diffusion models have focused on uniform diffusion mechanisms that may overlook heterogeneous spectral features. To address this, the paper proposed SPREAD, a graph diffusion network that jointly adapts forward perturbation and reverse graph reconstruction considering graph spectral features. The forward process uses a learnable spectral response function that regulates noise intensity across frequency components, while the reverse process weighs spectral bands according to the learned features. The experiments were performed on multiple datasets for node and graph classification.

Critical strengths

• The idea is well-motivated and intuitive, and the paper is easy to follow. The idea that the forward and reverse processes are jointly adapted using the same learnable spectral response enhances the coherence of the method.

• The experiments were performed on both graph and node classifications on multiple datasets, and hyperparameter sensitivity analysis is also provided.

Critical weaknesses

• The theoretical contributions are relatively limited, as the main results largely rely on established properties of Bernstein polynomials and standard variance bounds. They do not directly explain why the proposed spectral adaptation leads to improved representation quality or why the frequency-adaptive diffusion yields better representation than a uniform diffusion.

• The reliance on graph Laplacian eigendecomposition raises significant scalability concerns. The paper lacks computational complexity, runtime, and memory comparisons, making it difficult to assess the practicality of SPREAD on large-scale graphs.

• The inverse-response weighting strategy lacks sufficient theoretical and empirical justification. Downweighting heavily perturbed frequency bands may discourage the recovery of challenging but informative components. It is not guaranteed that the learned response preserves the semantic information of the band or represents the recovery difficulty.

• Although the method is motivated by heterogeneous spectral characteristics, the evaluation lacks heterophilic graph benchmarks (e.g., Roman-empire dataset), where high-frequency information is particularly important. Such experiments would provide more direct evidence for the effectiveness of frequency-adaptive diffusion.

• The improvements over strong baselines are relatively modest on several datasets. Moreover, the graph classification experiments rely on a single pretraining seed and do not report standard deviations, limiting confidence in the statistical robustness of the results.

• The paper lacks comparisons with simpler alternatives, such as spectral filtering without diffusion and fixed low-pass/high-pass frequency weighting. Without these comparisons, it is difficult to determine whether incorporating learnable frequency modulation into the diffusion framework is truly necessary or provides meaningful benefits over simpler spectral methods.

Questions for authors

• Are all spectral-band cosine similarities guaranteed to be nonnegative?

• How does SPREAD scale with graph size?

• Does the method improve performance on heterophilic graphs?

• Why is inverse-response weighting preferable to direct-response weighting?

2 comments
AuthorsOpenReview ↗

Response to Reviewer

We thank the reviewer for valuable comments.

W1: Our theoretical analysis aims to explain how spectral adaptation affects information preservation and reconstruction learning. Specifically, Section 3 explains why uniform perturbation cannot adjust relative SNRs across frequencies. To address this limitation, our learnable response jointly controls frequency-dependent noise allocation and reconstruction supervision. To support this design, Theorem 4.1 establishes the expressive capacity needed to represent nonconstant spectral responses, while Theorem 4.2 supports band-wise aggregation and relates reconstruction weights to recovery requirements under the stated assumptions. Together, these results motivate the design. To complement this theoretical analysis, downstream classification and ablation studies provide empirical evidence of improved representation quality.

W2&Q2: We will add complexity analysis and inference-time comparisons to clarify how SPREAD’s computational costs scale with graph size. Specifically, for a graph with N nodes and feature dimension d, full eigendecomposition takes O(N^3) time, and storing the eigenbasis requires O(N^2) memory. The decomposition is performed during preprocessing and cached for reuse in training and inference. Each spectral transform or inverse transform costs O(N^2d) and evaluating the degree-M Bernstein response costs O(NM). Thus, the spectral module incurs an additional computational cost of O(N^2d+NM) per training step. Reconstruction losses are not computed during inference.

To complement this complexity analysis, we add an “Inference Time” subsection to Appendix G of the revised manuscript, comparing SPREAD with the diffusion baselines DDM and SDMG. The results below show additional inference overhead, mainly from transformations between the node and spectral domains. However, these computations enable perturbations to adapt to graph spectral properties, improving downstream classification and allowing direct analysis of perturbation allocation and reconstruction quality across spectral bands. Considering these benefits, this overhead represents a reasonable trade-off for the performance and interpretability benefits at the evaluated graph sizes.

MethodCora (ms)CiteSeer (ms)PubMed (ms)Computer (ms)Photo (ms)
DDM37.9670.17328.34254.97171.76
SDMG43.7673.84385.10295.74186.66
SPREAD49.3174.33454.98349.24203.02

Regarding scalability, with fixed feature dimension, full eigendecomposition scales cubically with node count, while eigenbasis storage and spectral transforms scale quadratically. For graph classification, decomposition is performed independently for each graph, with total cost O(\sum_g N_g^3). For larger individual graphs, partial eigendecomposition or approximate spectral computation could be explored.

W3&Q4: Inverse-response weighting allocates reconstruction supervision according to forward perturbation, prioritizing frequency components that retain more information and preventing heavily perturbed bands from dominating training. Specifically, Section 3 shows that, for fixed signal and base noise power, a larger response increases noise power through its square and reduces the SNR, which motivates our design. Here, the response describes noise modulation, rather than directly measuring semantic importance or recovery difficulty. Theorem 4.2 further shows that, given an upper bound on the overall reconstruction error, larger band weights impose stricter recovery requirements. Thus, inverse-response weighting assigns stricter supervision to bands with smaller responses, whereas direct-response weighting emphasizes recovery of more strongly perturbed bands.

To test this empirically, we directly compared inverse- and direct-response weighting. On Photo, inverse-response weighting achieves higher classification accuracy and lower unweighted errors in bands 1 and 2, thereby supporting its benefit for downstream representation learning. Full results are provided in Appendix I.4 of the revised manuscript.

Weighting strategyBand 1 error↓Band 2 error↓Band 3 error↓Accuracy (%)↑
Direct-response weighting0.27590.34080.321294.93 ± 0.19
Inverse-response weighting0.19400.32940.341295.69 ± 0.16
AuthorsOpenReview ↗

Response to Reviewer

W4&Q3: We added experiments on the heterophilic graph dataset Roman-empire to evaluate frequency-adaptive diffusion under heterophily. Selected results are shown below, with full results provided in Section 5 of the revised manuscript. SPREAD achieves the highest accuracy among the compared methods, supporting the effectiveness of frequency-adaptive diffusion under heterophily.

MethodGCNGATDGIMVGRLGraphMAEDDMSDMGSPREAD
Accuracy (%)61.0562.4366.1866.9569.4867.8570.9671.30

W5: We add graph classification results across multiple pretraining seeds, reporting mean accuracy and standard deviation in Appendix F of the revised manuscript to assess the robustness of the results.

W6: To assess the benefit of learnable frequency modulation within diffusion, we compare SPREAD with Spectral-only (spectral filtering without diffusion), Fixed-LP, and Fixed-HP (diffusion with fixed low-pass and high-pass frequency weighting, respectively). The results below on Photo show that SPREAD achieves higher classification accuracy than these simpler alternatives, thereby supporting the joint design of diffusion and learnable frequency modulation. Full results are provided in Appendix I.1 of the revised manuscript.

MethodSpectral-onlyFixed-LPFixed-HPSPREAD
Accuracy (%)93.12±0.4294.34±0.2595.16±0.1495.69±0.16

Q1: Our implementation uses S(x,y)=(1+cos⁡2(x,y))/2S(x,y)=(1+\cos^2(x,y))/2, ensuring nonnegative spectral-band similarities and therefore well-defined weighted powers in Eq. (12). We clarify this similarity definition in the revised manuscript.

Reviewer_qxFjOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SPREAD learns a spectral response function to control frequency-dependent noise injection and spectral reconstruction weights for graph representation learning. Experiments show competitive classification performance, but the formulation leaves concerns about zero-noise degeneration and whether gains persist under noise-strength-matched comparisons.

Critical strengths
  1. Clear motivation. The spectral SNR analysis provides a concrete explanation of why a global noise schedule cannot adjust the relative corruption levels across frequencies. This motivates investigating frequency-dependent perturbations.

  2. Compact and interpretable design. The low-degree response function introduces few additional parameters and allows the learned perturbation profile to be inspected directly. Connecting forward perturbation with reconstruction weighting is a potentially useful design choice.

  3. Encouraging empirical results. SPREAD performs well across several benchmarks, with improvements exceeding one percentage point over the strongest reported baselines on Computer and Photo. The ablation results provide initial evidence that both proposed components contribute to performance.

Critical weaknesses
  1. Potential zero-noise degeneration: The objective appears to permit vanishing noise responses, making reconstruction trivial; the manuscript does not explain how this is prevented.
  2. Missing noise-strength-matched controls: Current ablations cannot separate the benefit of frequency adaptation from simply reducing overall corruption.
  3. Limited novelty: The contribution mainly combines existing ideas in directional diffusion, spectral reconstruction, and Bernstein filtering.
  4. Unclear weighting behavior: Jointly learned reconstruction weights may downweight difficult frequency bands rather than improve their recovery.
Questions for authors
  1. What prevents the response from collapsing toward zero? Please specify coefficient constraints, initialization, gradient flow through the perturbation process, and any regularization or normalization omitted from the formulation. Can you report response magnitudes and actual injected noise energy throughout training?

  2. Does frequency-dependent allocation outperform simpler perturbations at matched noise power? Please compare against a tuned or learnable global scalar response and fixed spectral profiles, controlling for expected injected noise energy.

  3. How does the coupled weighting mechanism compare with fixed weights, independently learned weights, or weights whose gradient contribution to the response parameters is blocked? Do improvements persist when evaluating all methods with the same unweighted band-wise metric?

  4. Are responses constrained to be nonnegative? If not, how is weighting by signed response reconciled with perturbation power depending on its square? What evidence connects the learned weights to actual spectral reliability?

  5. How are repeated eigenvalues handled in the spectral loss? Please also provide computational costs and variability across multiple pretraining seeds for graph classification.

2 comments
AuthorsOpenReview ↗

Response to Reviewer

We thank the reviewer for valuable comments.

W1&Q1: We thank the reviewer for raising this concern. In our implementation, the Bernstein coefficients are parameterized as cm=σ(am)∈(0,1)c_m=\sigma(a_m)\in(0,1). Since the Bernstein basis functions satisfy Sm,M(λ~)≥0S_{m,M}(\tilde{\lambda})\geq 0 and ∑m=0MSm,M(λ~)=1\sum_{m=0}^{M}S_{m,M}(\tilde{\lambda})=1, the spectral response is a convex combination of these coefficients, satisfying 0<min⁡mcm  ≤  ri  ≤  max⁡mcm  <  10<\min_{m} c_m \;\leq\; r_i \;\leq\; \max_{m} c_m \;<\; 1. This ensures positivity and boundedness. Empirically, Figure 4(b) shows that the learned responses on all five datasets remain clearly above zero and vary across frequencies, indicating that the reported models do not exhibit near-zero response degeneration. Regarding initialization, the degree-two implementation sets the raw parameters to (0,0.5,1), yielding coefficients of approximately (0.5,0.62,0.73). Therefore, training starts with moderate and nonzero perturbations. During optimization, gradients reach the response parameters jointly through both (i) the forward perturbation path, via the noisy input and denoising network, and (ii) the spectral reconstruction path, via the spectral-band weights. As for normalization, the base directional noise undergoes layer normalization before spectral modulation, while the band weights are constructed through standardization and softmax. No additional noise-energy penalty or fixed-power normalization is used. To further examine training behavior, we report response magnitudes and actual injected noise energy at selected training epochs on Photo below. Both remain nonzero and increase slightly, showing no observed collapse toward zero during the reported training period.

Epoch1100200300400500
Response magnitude (%)63.5864.1764.8765.3565.5865.84
Injected noise energy (%)36.3137.1237.4338.5938.9139.18

W2&Q2: To separate the effects of frequency allocation from overall perturbation strength, we conducted noise-power-matched comparisons against a tuned global scalar response, a learnable global scalar response, and fixed increasing and decreasing spectral profiles. Specifically, all variants use the same expected injected noise energy and otherwise identical settings to isolate the effect of forward frequency allocation. The results below on Photo show that SPREAD outperforms scalar and fixed spectral responses at matched noise power, thereby supporting the benefit of adaptive frequency allocation. Full results are provided in Appendix I.2 of the revised manuscript.

MethodTuned global scalarLearnable global scalarFixed increasing profileFixed decreasing profileSPREAD
Accuracy (%)95.06±0.2894.86±0.2395.16±0.3194.75±0.2995.69±0.16

W3: The key idea is to explicitly model the spectral response of graph signals so that diffusion can adapt to their intrinsic spectral properties. To this end, our contribution is the joint design of forward perturbation and reverse reconstruction supervision through a shared learnable spectral response. As shown in Section 3, a uniform diffusion schedule cannot adjust relative SNRs across frequencies. To address this, we learn frequency-dependent noise intensities and derive spectral-band reconstruction weights from the same response. In this framework, directional noise supplies the base perturbation, while Bernstein polynomials parameterize the response. Therefore, the novelty lies in coupling noise allocation with reconstruction supervision to address spectral heterogeneity. Ablation studies further support the contribution of each component.

AuthorsOpenReview ↗

Response to Reviewer

W4&Q3: Our reconstruction weights are derived from the forward spectral response to adjust each band’s contribution to supervision according to perturbation strength, rather than directly downweighting difficult bands based on reconstruction errors. To assess the effect of this weighting, existing experiments provide two forms of evidence: (i) w/o Weight in Section 5.3 replaces adaptive weights with fixed uniform weights, supporting the role of response-guided weighting; (ii) Figure 5 reports unweighted band-wise errors, showing improvements beyond the lowest spectral band. To directly address the requested comparisons, we further compare SPREAD with fixed uniform weights, independently learned weights and weights whose gradient contribution to the response parameters is blocked. All variants use identical training settings and are evaluated with the same unweighted band-wise reconstruction metric and classification accuracy. On Photo, SPREAD achieves the highest classification accuracy and the lowest unweighted errors in all bands, thus indicating recovery gains beyond merely reducing the weighted loss. Full results are provided in Appendix I.3 of the revised manuscript.

MethodBand 1 error↓Band 2 error↓Band 3 error↓Accuracy (%)↑
Fixed uniform weights0.26480.33150.356395.06±0.21
Independently learned weights0.20850.34180.355795.28±0.17
Blocked weight-to-response gradients0.19750.33850.364195.35±0.24
SPREAD0.19400.32940.341295.69±0.16

Q4: Our spectral responses are constrained to be nonnegative (see W1 & Q1). Therefore, their ordering is consistent with that of their squares, so inverse-response weighting introduces no sign ambiguity. Regarding spectral reliability, Section 3 shows that it depends jointly on the original signal power, base noise power, and spectral response. Our weights thus provide response-guided relative allocation of reconstruction supervision, rather than directly estimating the actual SNR. To examine this connection empirically, we additionally report normalized learned weights, mean information-preservation scores, and unweighted reconstruction errors on Photo. Across the three bands, higher weights correspond to greater information preservation and lower errors, thereby empirically supporting the reliability interpretation of response-guided weighting in this setting.

Spectral bandNormalized weightMean information-preservation score ↑Unweighted error (1 − s) ↓
B₁0.91930.43290.1940
B₂0.06220.41200.3294
B₃0.01850.39210.3412

Q5: For spectral reconstruction, the loss compares original and reconstructed signals in the same orthonormal basis and is invariant to eigenvector sign flips. However, index-based band partitioning may split a repeated-eigenvalue eigenspace. Moreover, Eq. (11) averages component-wise cosine similarities, so the loss may still depend on the orthonormal basis chosen within that eigenspace. Regarding computational costs, for a graph with N nodes and feature dimension d, full eigendecomposition takes O(N^3) time, and storing the eigenbasis requires O(N^2) memory. The decomposition is performed during preprocessing and cached for reuse in training and inference. Each spectral transform or inverse transform costs O(N^2d) and evaluating the degree-M Bernstein response costs O(NM). Thus, the spectral module incurs an additional computational cost of O(N^2d+NM) per training step. Reconstruction losses are not computed during inference. Moreover, to assess the robustness of the results, we add graph classification results across multiple pretraining seeds, reporting mean accuracy and standard deviation in Appendix F of the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.