acceptodds
Under review as a conference paper at ICLR 2027

SpecValley: Aggregation as Spectral Valley in Dense Heterophilic Graphs

Abstract

Node classification in heterophilic graphs remains an open problem: aggregating over neighbourhoods can damp precisely the part of the signal that carries the labels. The field has answered with ever more expressive propagation operators, leaving the aggregation step that precedes them fixed by degree normalisation. However, degree normalisation acts as a spectral valley in its own right, attenuating a fixed middle band whose position at is set by the normalisation rather than by the data, exactly where the eigenvalues of dense graphs accumulate. We therefore begin with the representations rather than the propagation, in two stages: a learnable aggregation followed by an ordinary spectral filter. The first, SpecValley, learns the position, sharpness and flank gains of the valley for every graph. We show formally how the aggregation operator and downstream filter interact. We establish that the response of degree normalisation vanishes exactly at that accumulation point. A spectral energy budget shows that what aggregation removes from a band, no linear spectral filter can recover. Crowding alone says nothing about the benefit. That is decided by how much label energy the valley removes, which we measure for every graph. SpecValley achieves strong results across heterophilic benchmarks, and a controlled ablation over aggregation operators and downstream filters attributes the gain to aggregation rather than the filter.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Machine Learning Heterophilic Graphs

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_x2HWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper observes that the degree normalization commonly used in graph neural networks induces a spectral “valley” effect, attenuating a fixed mid-frequency band centered at (\lambda = 1), where the eigenvalues of dense graphs tend to concentrate. Based on this observation, the authors propose learning the location, sharpness, and flank gains of this valley for each graph, and provide theoretical analyses and empirical evaluations of the proposed method.

Critical strengths

1.The authors propose a learnable aggregation stage that generalizes degree normalization and determines the available spectral energy budget for each downstream filter.

2.The paper provides theoretical analyses to support its claims and explains the effectiveness of the proposed method from the perspectives of eigenvalue distributions, label energy, and filtering capacity.

3.The proposed SpecValley model achieves strong performance in several experimental settings.

Critical weaknesses

1.The theoretical analysis in Section 4.3 is difficult to follow.

2.The paper motivates SpecValley by observing that the conventional spectral response vanishes at (\lambda = 1), and proposes making the spectral response of the aggregation stage learnable. However, the paper does not provide a sufficiently detailed theoretical and empirical analysis to establish whether the zero response at (\lambda = 1) actually leads to degraded classification performance.

3.If the authors believe that the spectral response harms classification performance, why not design a filter such as (1-\lambda/2), whose response is nonzero at (\lambda = 1), to preserve model performance, rather than introducing a computationally expensive aggregation stage?

4.The proposed method requires eigendecomposition, which may be prohibitively expensive for large graphs, such as OGBN-Arxiv.

5.Section 5.1 includes experimental analysis, which seems unusual, as this section should focus on describing the datasets and experimental settings.

6.The relationship between label energy and downstream classification performance requires further investigation to explain the performance implications of a zero spectral response at (\lambda = 1).

7.Table captions should be placed above the tables.

8.Figure 3 is too tall, and Table 3 should be placed to the right of the text.

Questions for authors

1.Could the authors design a filter such as (1-\lambda/2), whose response is nonzero at (\lambda = 1), to preserve model performance?

2.Does a zero spectral response at (\lambda = 1) lead to degraded classification performance? Could the authors provide more detailed theoretical and empirical evidence to support their claims?

3.What is the computational cost of performing eigendecomposition on a large graph such as OGBN-Arxiv?

4.Could the theoretical analysis in Section 4.3 be rewritten in a more accessible and easier-to-understand manner?

1 comment
AuthorsOpenReview ↗

Thank you for your review

We would like to thank you very much for the detailed reviews and the points you raised. Our current inclination is to take up your feedback and to develop the paper further from a method paper into an analytical paper, placing the focus more on the supporting property of improving the input representations with 1 − λ and less on a flexible, learned form of the spectral valley. We would try to explain in more depth why the mechanism works and to considerably extend the dataset basis.

We are very grateful for your reviews and your feedback, which have been helpful for the further development of this work.

Reviewer_Dt8MOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SpecValley, a learnable spectral aggregation operator for heterophilic graph learning. The method adjusts the position, sharpness, and gains of a spectral valley before applying a downstream graph filter. The authors provide theoretical analyses of spectral concentration and energy attenuation, with experiments on six heterophilic benchmarks.

Critical strengths

Motivation: I think the overall idea of examining aggregation separately from downstream filtering interesting. In particular, the observation that spectral concentration alone does not determine whether aggregation helps is meaningful.

Evaluation: The controlled comparisons across aggregation operators and downstream filters are useful. The additional analysis of label-energy distributions and sensitivity to filter order is interesting.

Critical weaknesses

Potentially unfair comparison due to additional supervision. My concern is the training protocol described in Appendix C.1. The spectral parameters of SpecValley are optimized using half of the validation set, whereas the comparison methods appear to use validation labels only for model selection. This means SpecValley may effectively receive additional supervised information during parameter optimization. Although the remaining validation half is reserved for epoch selection, this does not resolve the difference in effective training-label budgets. Since the main empirical contribution is the improvement over alternative aggregation operators, I find this issue consequential. The authors should clarify whether the baselines receive equivalent supervision and whether the improvements remain under a matched-label protocol.

The contribution of the learned spectral valley is not sufficiently isolated. Equation (3) combines ego features, neighborhood aggregation, and a learnable spectral response with separate feature projections. Since all evaluated graphs are treated as undirected, the two neighborhood operators become identical. Therefore, the proposed architecture introduces multiple representation paths beyond simply learning the valley. I am not convinced that the current ablations establish whether the improvements come from the valley itself, preserving ego information, or additional model capacity. An ego-plus-neighborhood baseline without the learned valley, and a valley-only variant with comparable parameters, would help clarify this.

Unclear distinction from existing adaptive spectral filtering. The proposed response is itself a learnable function of the graph Laplacian, which makes it conceptually close to an adaptive spectral filter. Moreover, the composition of spectral aggregation and downstream spectral filtering can be viewed as another spectral transformation in the linear setting. I would like to understand what fundamentally distinguishes the proposed aggregation mechanism from placing an additional learnable spectral filter before ChebNet. Existing approaches such as GPR-GNN, BernNet, and HeroFilter already investigate adaptive frequency responses. The paper discusses these methods, but I do not think the current analysis fully establishes the distinct representational or optimization advantage of SpecValley.

The theoretical results do not fully explain the observed improvements. Proposition 4.1 characterizes a spectral property of degree normalization, while Corollary 4.2 establishes an energy-preservation constraint for linear spectral filters. These results provide useful intuition, but they do not directly justify the particular valley parameterization or guarantee improved classification. The downstream network also includes nonlinear operations that can redistribute spectral energy, and the ego branch can bypass the attenuated frequencies. Furthermore, the exact spectral implementation requires eigendecomposition and quadratic memory, while the largest graph uses a polynomial approximation. I would appreciate stronger evidence connecting the learned valley parameters to label-energy preservation, together with clarification of the approximation's effect on the reported gains.

Questions for authors

See weakness

1 comment
AuthorsOpenReview ↗

Thank you for your review

We would like to thank you very much for the detailed reviews and the points you raised. Our current inclination is to take up your feedback and to develop the paper further from a method paper into an analytical paper, placing the focus more on the supporting property of improving the input representations with 1 − λ and less on a flexible, learned form of the spectral valley. We would try to explain in more depth why the mechanism works and to considerably extend the dataset basis.

We are very grateful for your reviews and your feedback, which have been helpful for the further development of this work.

Reviewer_4GaCOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper revisits neighborhood aggregation in heterophilic graph learning from a spectral perspective. Its central observation is that symmetric degree normalization has a frequency response (1-\lambda), which vanishes at (\lambda=1), where the normalized Laplacian eigenvalues of dense graphs tend to concentrate. Whether this attenuation is helpful depends on how much label-relevant information occupies that spectral region.

SpecValley introduces a learnable aggregation stage before a conventional spectral filter. The aggregation combines ego features, neighborhood information, and a parameterized spectral response whose center, sharpness, and flank gains can adapt to the graph.

The theoretical analysis characterizes spectral concentration and the energy remaining after aggregation. The experimental evaluation crosses several aggregation operators with four downstream filters on six heterophilic benchmarks, and includes label-spectrum diagnostics and robustness studies.

Critical strengths

S1: The observation that (P_{\mathrm{sym}}=I-L) has a zero response at (\lambda=1) is straightforward mathematically, but connecting this observation to spectral concentration and the distribution of label energy is informative. I particularly appreciated that the paper does not equate eigenvalue concentration with predictive benefit. The Crocodile and Tolokers comparison illustrates this distinction well. Although both graphs have substantial spectral mass near (\lambda=1), the central band contains markedly different fractions of label energy. This offers a plausible explanation for why the same normalization can help on one graph and hurt on another.

S2: Crossing five aggregation operators with four filters, while also including a no-aggregation reference, is a useful way to investigate whether the aggregation stage contributes something beyond the downstream filter. The additional sweep over Chebyshev polynomial order is valuable because it addresses the natural objection that a more expressive filter might absorb the aggregation stage.

Critical weaknesses

W1: The validation set is divided into two halves (according to Appendix C1): the spectral operator parameters (\theta=(c,p,a_L,a_R,\tau)) are optimized on half A, while half B is used for epoch selection. The remaining parameters follow the training loss. This means that SpecValley uses additional labeled observations for parameter optimization that are ordinarily reserved for validation, whereas the competing models appear to use their validation labels only for model selection. Although the paper is explicit about this protocol and does not report using test labels during training, the resulting comparison is not a standard matched-supervision evaluation. I consider this consequential because the parameters receiving the additional supervision are precisely those responsible for the proposed adaptive spectral response. Without a train-only variant, it is difficult to determine how much of the reported improvement should be attributed to the aggregation mechanism rather than to the additional labeled data.

W2: Equation (3) contains an ego branch, two neighborhood branches, and the parameterized spectral branch, each with learned projections. Since the experiments symmetrize the graphs, the two neighborhood operators become identical, and their contributions can be combined into a single learned linear transformation. The comparisons in Table 3 change the aggregation architecture, not just the valley parameters. What I miss is a matched-architecture experiment in which the ego and neighborhood branches, projections, training protocol, and parameter budget remain unchanged, while the spectral response is fixed, removed, or learned. This matters because the theoretical analysis itself recognizes that the ego branch preserves information that conventional aggregation may suppress. It is therefore plausible that some of the gains attributed to moving the spectral valley actually arise from preserving the original features or increasing the model's parameterization.

W3: There is also a nontrivial computational cost: the exact spectral branch requires an eigendecomposition and quadratic memory. A comparison against matched-cost alternatives would make the practical contribution easier to assess.

W4: As a side note, the benchamrk used are too small to have an idea of the real usefulness of the approach.

Questions for authors

Q1: Can the authors rerun the main comparisons with all SpecValley parameters, including the spectral operator parameters, trained exclusively on the training labels, reserving validation labels only for hyperparameter and epoch selection?

Q2: Could the authors provide an ablation retaining the complete architecture of Eq. (3) while comparing a learned valley against a fixed (1-\lambda) response, a fixed (|1-\lambda|) response, and no spectral branch?

Q3: How does the accuracy-compute trade-off compare with increasing the capacity or polynomial order of a conventional spectral model under matched training resources?

1 comment
AuthorsOpenReview ↗

Thank you for your review

We would like to thank you very much for the detailed reviews and the points you raised. Our current inclination is to take up your feedback and to develop the paper further from a method paper into an analytical paper, placing the focus more on the supporting property of improving the input representations with 1 − λ and less on a flexible, learned form of the spectral valley. We would try to explain in more depth why the mechanism works and to considerably extend the dataset basis.

We are very grateful for your reviews and your feedback, which have been helpful for the further development of this work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.