acceptodds
Withdrawn from ICLR 2027

CH-GCCIL: Clique-Hierarchical Spectral Intervention for Graph Causal Contrastive Invariant Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Graph contrastive learning has emerged as a dominant framework for graph representation learning by learning discriminative node representations via graph augmentation. Nevertheless, conventional stochastic augmentation schemes fail to differentiate causal factors from non-causal confounding factors, making the learned representations prone to capturing spurious correlational patterns rather than invariant causal structures. Current causal-invariant graph learning methods primarily rely on global spectral decomposition, which imposes a uniform frequency partition of causal and non-causal components on the entire graph. Such global paradigms overlook the inherent structural heterogeneity of graph cliques: nodes inside individual cliques exhibit consistent causal mechanisms, while cross-clique connections are dominated by non-causal noisy associations, thereby resulting in coarse and inaccurate causal separation. To address this limitation, this paper proposes a novel Graph Causal Contrastive Invariant Learning method with Clique-Hierarchical Spectral Intervention (CH-GCCIL) to achieve fine-grained causal disentanglement from a clique hierarchical perspective. Specifically, we first mine maximal cliques from the graph and categorize them into strongly causal cliques and noisy cliques based on a three-dimensional confidence metric. On this basis, we construct asymmetric dual augmented views and adopt Chebyshev approximation for efficient spectral filtering. Furthermore, we design a dual-scale contrastive loss function to optimize representation learning at both global graph and intra-clique levels. Extensive node classification experiments on five benchmark datasets validate the superiority of CH-GCCIL. It achieves Macro-F1 scores of 78.8%, 84.1%, and 44.1% on Wiki-CS, PubMed, and Flickr datasets, outperforming state-of-the-art baseline methods by 3.2%, 2.6%, and 4.1% respectively, which demonstrates the effectiveness and advancement of the proposed clique-hierarchical spectral intervention mechanism.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Contrastive Learning, ‌ Invariant Learning, Causal Inference,Spectral Intervention,Clique Mining

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_GcWfOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes CH-GCCIL which is a graph contrastive learning method where the construction of the contrastive samples are based on maximal cliques from the input graph. The idea behind the proposed method is that nodes within the same clique typically have similar causal mechanisms and their structure should be therefore preserved during the construction of contrastive samples. Based on this, the proposed method first finds all maximal cliques inside the input graph and gives each clique an aggregate score (label consistency, feature similarity, and topological openness). Based on these scores, the top half of the cliques are categorized as strong causal cliques and receive a small amount of noise during augmentation and the bottom half of the cliques are categorized as noise cliques and receive a large amount of noise during augmentation. The model is then trained with a three-component loss function (invariance loss, independence loss, and standardization loss). The proposed method is evaluated on five node classification datasets and achieves SOTA performance on many of them and close to SOTA on the rest.

Critical strengths
  • The idea that causal mechanisms can be linked to maximal cliques is very interesting. The contribution of this paper over GCIL which is the closest line of work is also significant.
  • The empirical results achieved by the method is promising as there are significant improvements over the evaluated baselines in several of the datasets.
Critical weaknesses
  • The paper provides no support for the claim that nodes within the same clique have similar causal mechanisms. Even if it were true, the proposed method separates maximal cliques into strong cliques and noise cliques based on the three metrics. If the paper claims that nodes within the same clique have similar causal mechanisms, why are large amounts of noise added to the low-scoring cliques? Essentially, the paper claims that their scoring mechanism can identify cliques that contain nodes with causal correlations and cliques that contain nodes with spurious correlations without any support.

  • The paper provides very few details on the design choices of the proposed method and almost no justification for these design choices. How were the default weights of the scoring mechanism chosen to be (0.6, 0.25, 0.15)? Why is the threshold for separating "strong causal cliques" and "noise cliques" chosen to be 0.5?

Some smaller typographical issues:

  • GCN was the best performing method on Flickr but is not bolded (Table 1).
  • The citation for GCIL seems to be wrong. Yixin Mo and Xiang Wang in the author list should be Yanhu Mo and Xiao Wang.
Questions for authors

See weaknesses.

Additionally:

  • Can the authors provided some clarification and additional analysis on the ablation experiments? For example, the results in Table 2 indicate that label + topology achieves 65.3 and 71.9 on Macro-F1 and Micro-F1 respectively. The results in Table 3 indicate that only topology achieves 75.0 and 78.2 on Macro-F1 and Micro-F1 respectively. This suggests that adding label to topology significantly decreases performance (despite label having the highest weight at 0.6). This is also true for the other two dimensions.

  • How were the baselines selected for comparative analysis? It is unclear what the criteria were for the baselines in Table 1. For example, LIRS (Yao et al, ICLR 2025) was included in the related works section but omitted in the experiments. Furthermore, the authors specified self-supervised methods as their baselines but the proposed method uses training labels in the scoring mechanism.

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

This paper proposes CH-GCCIL, a graph causal contrastive invariant learning framework that addresses the limitations of global spectral interventions in capturing local structural heterogeneity. The method first identifies maximal cliques and categorizes them into strongly causal and noisy cliques using a three-dimensional confidence score based on label consistency, feature similarity, and topological openness. It then constructs asymmetric augmented views through clique-specific spectral interventions, employing polynomial approximation for efficient spectral filtering. Additionally, a dual-scale contrastive objective combines global and intra-clique invariance and independence constraints to learn stable node representations. Experiments on five node classification benchmarks demonstrate competitive performance.

Critical strengths
  1. The paper identifies the limitation of existing graph spectral intervention methods in overlooking local structural heterogeneity and explores the use of maximal cliques for more fine-grained graph augmentation, addressing a meaningful research problem.

  2. The proposed method demonstrates promising performance, outperforming the most closely related baseline, GCIL, on the PubMed, Wiki-CS, and Flickr datasets.

Critical weaknesses
  1. The paper assumes that high-confidence cliques correspond to causal structures, while low-confidence cliques and cross-clique connections mainly capture non-causal information. However, whether label consistency, feature similarity, and topological closure can reliably indicate causality remains questionable. The authors provide neither rigorous theoretical justification nor empirical evidence to support this key assumption.

  2. The paper does not establish whether the assumption that low-frequency components capture causal information while high-frequency components represent non-causal information holds across different local cliques. In particular, applying strong full-spectrum perturbations to noisy cliques may also destroy useful causal information, yet no theoretical or empirical evidence is provided to rule out this possibility.

  3. Although the paper emphasizes causal invariant representation learning, experiments are limited to standard node classification benchmarks without explicitly considering distribution shifts or evaluating cross-environment invariance. Therefore, improvements in classification performance alone are insufficient to substantiate the claimed ability to learn causally invariant representations.

  4. Most experimental baselines were proposed in or before 2024, with insufficient comparisons against more recent competitive methods. Moreover, CH-GCCIL performs noticeably worse than several these methods on CiteSeer, raising concerns about its robustness and general applicability.

  5. Figures 1 and 2 appear to use low-resolution, non-vector graphics that become blurry when enlarged. Moreover, Figure 1 is visually cluttered yet provides limited methodological insight, resembling a simple arrangement of text boxes and arrows rather than an informative framework illustration.

Questions for authors

See Weaknesses.

Reviewer_NT7jOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a graph causal contrastive invariant learning framework that aims to improve the robustness of graph representations by addressing the limitations of global spectral interventions in distinguishing causal and non-causal information. The proposed approach consists of three main components, including clique-hierarchical construction, asymmetric hierarchical spectral intervention, and dual scale causal constraints.

Critical strengths
  • The paper clearly identifies the limitations of existing global spectral intervention methods and motivates the need for fine-grained causal representation learning based on graph clique structures.

  • The proposed framework integrates clique hierarchical construction, asymmetric spectral intervention, and dual-scale causal constraints into a unified graph representation learning approach.

Critical weaknesses
  • The proposed method fails to consistently outperform competitive baselines across datasets. However, some claims of superior performance are inconsistent with the reported results. For example, in Table 1 of the Flickr dataset, the GCN baseline clearly outperforms the proposed method by 4%. But in the paper, the results of the proposed method are mistakenly marked as the best performance. Besides, in the same table, the performance on the CiteSeer and Cora datasets does not outperform the baselines either.

  • The novelty of this paper is incremental.

  • The paper lacks rigorous theoretical analysis to justify the assumption that clique structures correspond to causal mechanisms and that the proposed interventions effectively preserve causal information.

  • Hallucinated citations are included in the references part.

Questions for authors
  • Could the authors explain this discrepancy and clarify the claims of superior performance in Table 1? Could the authors explain why the proposed method fails to consistently outperform competitive baselines across different datasets?

  • Could the authors clarify the fundamental methodological novelty and explain how the proposed approach differs substantially from existing methods?

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

Related papers

Discussion (0)

Sign in to comment.