acceptodds
Withdrawn from ICLR 2027

ASPECT: Node-Level Adaptive Spectral Fusion for Graph Contrastive Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Spectral graph contrastive learning constructs complementary low- and high-frequency views, but existing methods typically combine them using graph-level or node-agnostic fusion rules. We show that such global fusion can incur irreducible regret when nodes have heterogeneous spectral preferences, motivating node-wise spectral fusion. We propose ASPECT, which learns a node-wise policy to fuse low- and high-frequency views and regularizes this policy using channel-wise contrastive evidence. We further introduce ASPECT-S, a stability-aware extension that incorporates channel-wise sensitivity estimated from generated graph-structure and feature perturbations, with a Rayleigh-based bias for perturbation search. Experiments on homophilic and heterophilic graphs show that ASPECT consistently improves clean representation quality over competitive spectral and graph contrastive baselines, while ASPECT-S further improves performance under joint structural and feature perturbations.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
graph contrastive learning, spectral graph learning, graph representation learning, graph neural networks, heterophily, robustness

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

The paper proposes ASPECT, a graph contrastive learning framework that adaptively combines low- and high-frequency representations using node-wise gating guided by channel-wise contrastive losses. It further introduces ASPECT-S, which incorporates perturbation sensitivity to improve robustness.

Critical strengths
  1. The paper presents extensive experiments on both homophilic and heterophilic graphs, with ablation studies
Critical weaknesses
  1. The research gap claim is overstated. The research motivation overlooks a substantial body of work on node-adaptive spectral filtering. FAGCN (AAAI 2021), ACM-GNN (NeurIPS 2022), and many more. This already investigates adaptive frequency selection or node-specific filter mixing.

  2. Weak theoretical motivation and insufficient connection to the proposed methodology.

Section 2 attempts to establish the limitations of existing methods and motivate the architecture introduced in Section 3. However, the theoretical analysis does not adequately achieve this objective.

Theorem 2.1 provides a limited theoretical contribution. The nonnegative regret of global fusion relative to a node-wise oracle follows directly from the inclusion of their optimization domains. The positive lower bound additionally assumes separated node-wise preferences and quadratic risk growth, which already impose a penalty for using a shared coefficient. But the theorem analyzes an unrestricted oracle rather than the proposed learned gating function and provides no guarantee that ASPECT approaches this oracle. The authors should justify these assumptions empirically

  1. Theorem 2.2 is a direct consequence of Lipschitz continuity and does not exploit any graph-specific or spectral properties. More importantly, the bound controls a contrastive surrogate rather than downstream classification performance. The analysis also simplifies the dependence of InfoNCE on other node representations that may change under graph perturbations. Therefore, it is unclear what specific theoretical insight this result provides for the proposed stability-aware spectral fusion mechanism.

The author should consider rewriting section 2 and its theme to avoid restating general limitations of globally shared fusion and perturbation-sensitive representations. Also, provide direct theoretical evidence or motivation for why their proposed model is better suited to solve the limitation.

  1. ASPECT/prposed model does not directly learn which spectral mixture is best for a node. Instead, it constructs a pseudo-target from the individual contrastive losses of two channels and trains the gate to imitate that target. A channel with a higher individual contrastive loss may still provide important information that improves the fused representation. The authors do not establish that independently measured channel losses are reliable indicators of the optimal fusion coefficient. Also, there is a ton of paper that uses combines both this frequency. This leads to oversmooting and oversquashing, which also has been studied extensively. For example, the paper "Multi-Track Message Passing: Tackling Oversmoothing and Oversquashing in Graph Learning via Preventing Heterophily Mixing" studies this and propose to use multi-track for keeping the frequency unmixed. Based on the previous research, I assume the Proposed method will be susceptible to oversmoothing and oversquashing.

  2. The method implicitly treats lower channel-wise contrastive loss as evidence of higher channel utility. This assumption is not sufficiently justified, particularly for heterophilic graphs where high-frequency representations may contain important discriminative information while being less consistent under graph augmentations. The authors should investigate whether the learned channel preferences correlate with downstream discriminative utility rather than only with the contrastive objective used to construct the pseudo-target.

  3. The policy target is generated from the same evolving channel representations that are jointly optimized with the gate. Although the target is detached, it is not independent of the model. This introduces a potential self-reinforcing feedback mechanism in which initially favored channels receive larger fusion weights, potentially reinforcing early training biases. This can also lead to representational collapse, as the model can try to cheat the loss by making all the representation space similar. For better understanding of this phenomenon, look at models such as the Burlow Twin or the Jepa model. This is not a graph model, but it talks about how representation collapse happens if the target is coming from the same evolving channel representations that are jointly optimized with the gate.

  4. The Rayleigh-based perturbation objective is poorly motivated. It encourages perturbations that reduce or reverse the expected spectral ordering of the low- and high-frequency channels, but the authors do not explain why this should improve robustness. Moreover, changes in Rayleigh energy may result from graph topology changes rather than spectral-filter instability. Stronger justification and comparisons with alternative objectives are needed.

Questions for authors

Why should the individual contrastive losses of the low- and high-frequency channels reliably determine their optimal fusion weights, especially when the channels contain complementary information?

Reviewer_7wxbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes ASPECT, which is a graph contrastive learning method that learns node-wise weights to combine low- and high-frequency representations. The authors first derive a regret lower bound to show that node wise fusion is better than global fusion. Then ASPECT is proposed which uses channel-wise contrastive losses to guide the fusion policy through a self-supervised regularizer. The paper also introduces ASPECT-S, which adds channel-wise sensitivity estimates from generated graphs and feature perturbations. Experiments on homophilic and heterophilic graphs show improvements in clean node classification and robustness under the evaluated perturbations compared to multiple baselines. Implementation details are also included in the appendix of the paper.

Critical strengths
  1. The research problem is valid: a single graph level fusion cannot adapt to various local graph structures. This limitation is real when the graph is mixed with homophily and heterophily.
  2. The design of proposed model is simple and easy to follow.
  3. Extensive experiments on multiple datasets with a variety of baselines are conducted. Appendix also provides abundant details, implementation settings as well as more theoretical analysis which helps improve the reproducibility of the paper.
Critical weaknesses
  1. The introduction argues that nodes with different local structures and homophily levels benefit from different spectral components. However, Theorem 2.1 defines spectral preferences through the minimizers of a self-supervised contrastive surrogate and assumes that these preferences differ across node groups. The paper does not analyze how these preferences correspond to the local structural properties used in the motivation. Additional analysis of the relationship between optimal fusion coefficients, local homophily, and learned gate weights would strengthen the theoretical and empirical justification for the proposed method.
  2. The policy target is based on the assumption that a lower InfoNCE loss indicates that the channel is more useful for fusion and prediction. However, InfoNCE primarily measures contrastive consistency and instance discrimination. It does not measure how much each channel contributes to the representation, nor does it measure downstream label informativeness. Experiments that compare the proposed target against alternative utility estimates such as mutual information with the fused representation would strengthen the justification for this claim in section 3.1.
  3. The design of ASPECT-S lacks rigorous justification and complete component analysis. The stability loss is not ablated and there is no comparable group for global fusion trained with the same perturbations. The sensitivity estimate uses raw L2 distance while the loss operates on normalized embeddings, and the inner maximization procedure is not specified. Additional ablation experiments including alternative perturbation objectives, and some explicit algorithm descriptions would help validate the design.
  4. The robustness experiments only reports results that combines Metattack-based edge perturbations with random feature masking. Structure-only and feature-only results are not reported, so the contribution of each corruption type cannot be isolated. Feature masking also overlaps with the augmentations already used during self-supervised training, which may impact the results.The authors should clarify the evaluation setup on this in a more rigorous way, report separate perturbation results, and include results under more attack scenarios to further demonstrate the robustness.
Questions for authors
  1. How does ASPECT perform on larger graphs? Is current design easy to be appleid to much larger graph datasets? What could be the limitations here?
Reviewer_AYefOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes RD-GAE for anomaly detection on graph-linked compositional observations. Row-stochastic graph aggregation and category mappings, an escort-temperature gate, and a Dirichlet bottleneck keep the data-carrying representations on simplices. A Rényi reconstruction loss with order below one is used during fitting, while a separately selected order above one scores discrepancies after the reconstruction model is frozen. The paper derives the Dirichlet prior divergence, a conditional result for anomaly–normal score separation, and fixed-predictor contamination-propagation bounds. Experiments use synthetic perturbations of MNIST, Omniglot, microbiome, and spatial transcriptomics profiles, with comparisons against six autoencoder families and architectural ablations.

Critical strengths
  1. A coherent compositional architecture with explicitly scoped analysis (Equations (1)–(3), Equation (5), Theorems 3.1/3.3, and Appendix C). This gives the proposed architecture a reasonably transparent mathematical interpretation.
  2. Consistent empirical gains within the evaluated controlled-contamination settings. RD-GAE has the highest mean AUROC and AUPRC among the main baselines on all four datasets in Table 1. The same-order comparisons and the architectural ablations in Table 6 are useful evidence beyond a single aggregate benchmark result. The paper also appropriately distinguishes synthetic biological perturbations from independently validated natural anomalies (Appendix D).
Critical weaknesses
  1. The current evidence supports the usefulness of the combined detector, but does not establish which part of its central robust-fitting claim is necessary for the gains. In other words, the evidence does not separate the claimed robustness of subunit-order training from the benefit of a different detection score.
  2. The reported results cannot yet be tied to a fully specified, consistent evaluation protocol. Section 2.4 selects the detection order by validation AUROC, whereas Appendix E, page 17, selects by average precision and uses AUROC only to break ties. More seriously, the Table 2 note explicitly says that Omniglot's nominal contamination rate “requires confirmation from the saved payload” because the preprocessing call and injector signature differ. This is material for interpreting precision–recall results and for reproducing the experiment.
Questions for authors
  1. With the architecture, contamination, optimization/selection budget, and detection-order search held comparable, how does λtr=1\lambda_{\mathrm{tr}}=1 with validation-selected λdet>1\lambda_{\mathrm{det}}>1 compare against the proposed subunit-training model? Does applying the same score-order selection to the strongest simplex-output baselines preserve the reported advantage? These controls would materially affect whether I attribute the contribution to robust training, post-training scoring, or their combination.
  2. Which detection-order selection rule actually produced Table 1: validation AUROC or average precision? Please verify the realized Omniglot contamination counts from the saved experiment data and identify the graph condition and inductive/transductive protocol for each reported result. If the verified settings differ from the stated protocol, do the conclusions remain after the corresponding results are corrected?

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

Related papers

Discussion (0)

Sign in to comment.