Enhance Graph Contrastive Learning with Perturbation Discrimination

1. The proposed method introduces discriminative components to Graph Contrastive Learning (GraphCL), which seems to represent a contribution of the components rather than the framework itself. Therefore, it is necessary to apply these components to other Graph Contrastive Learning algorithms. However, the evaluation has only been conducted on GraphCL. I recommend applying the Perturbation Discrimination components to other algorithms, such as AD-GCL, AutoGCL, and GPA, to assess their effectiveness.

2. The baselines differ between the settings of Unsupervised, Semi-supervised, and Transfer Learning, but the authors do not explain why certain algorithms present in the Unsupervised setting are absent in the Semi-supervised and Transfer Learning settings. Please provide an explanation for these differences.

3. The paper lacks hyper-parameter sensitivity experiments, particularly for parameters such as $\lambda_1, \lambda_2, \lambda_3$, Even though the authors mention “optimal hyper-parameters in the pilot studies”, there are no details provided about these pilot studies. Please include the details about pilot studies to enhance the understanding of the hyper-parameter selection process.

4. The authors provide components to control the augmentations; however, many graph contrastive learning models also focus on controlling augmentations, such as spectral-based methods [4, 5, 6], community-based methods [1, 4], and saliency-related methods [7]. These should be discussed in the literature review for a more comprehensive understanding of the field.

5. The section on related work regarding Graph Contrastive Learning (GCL) is insufficient, as it primarily discusses methods developed before 2023. Additionally, there is no comparison between your work and other recent methods. Please include a broader range of recent graph contrastive learning works [1-7] and provide detailed comparisons between your approach and these related works.

6. The source code is lacking.

[1] Clusterscl: Cluster-aware supervised contrastive learning on graphs, WWW, 2022.

[2] Graph Self-supervised Learning with Augmentation-aware Contrastive Learning. WWW 2023.

[3] Simple and Asymmetric Graph Contrastive Learning without Augmentations. NeurIPS 2023

[4] Community-Invariant Graph Contrastive Learning, ICML, 2024.

[5] Spectral Feature Augmentation for Graph Contrastive Learning and Beyond, AAAI, 2023.

[6] Spectral augmentation for self-supervised learning on graphs, ICLR, 2023.

[7] Boosting graph contrastive learning via graph contrastive saliency, ICML, 2023.

• The paper lacks evidence on to what extent do perturbations enhance GCL performance and to what extend do perturbations hurt GCL performance. Providing insights into this quantitatively would improve the paper's clarity and practical value.
• The model involves numerous hyperparameters to tune, including key parameters $\gamma_n$, $\gamma_e$, $\lambda_1$, $\lambda_2$, and $\lambda_3$, which could limit its generalizability and increase tuning complexity.
• An in-depth analysis of the discriminators' performance would be helpful. Specifically, how effective are they at discriminating perturbed nodes/edges, and how does this capability relate to overall GCL performance?
• Constructing the ground truth to train the node-oriented and edge-oriented discriminators may be computationally costly. Each graph is augmented in each training epoch, requiring ground truth calculations for the discriminators in every instance. The authors are encouraged to provide a complexity analysis and plot the training time overhead to clarify this aspect.
• The motivation behind certain design choices in the model is not entirely clear. For example, in Equation 5, it is unclear why the model would benefit from a higher proportion of perturbed nodes or edges. The authors should offer more insights into this decision.
• The performance improvements reported are relatively modest. Expanding the experimental evaluation to additional datasets would provide a clearer picture of the model's robustness and general applicability.

1. At line 228, the authors mention that the discriminator is trained to predict whether nodes are dropped or perturbed. However, if a node is dropped, there would be no corresponding node representation. In this case, why is it necessary for the discriminator to make this prediction, and how can it accomplish this task?

2. Although the authors introduce two discriminators to assess whether the graph augmentation is excessively aggressive, the proposed design does not inherently guides the strength of the graph augmentation. This raises the possibility that the discriminator could make correct predictions for all nodes and edges, causing the framework to degrade into a standard Graph Contrastive Learning (GCL) method. The authors should analyze the likelihood of such degradation and clarify the implications of their method if this degradation occurs.

3. The authors are encouraged to conduct experiments demonstrating the correlation between the discrimination accuracy and the final performance outcomes.