- 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.