Explanation-Assisted Data Augmentation for Graph Learning

1. It appears the effectiveness of EA-DA is heavily dependent on the quality of the subgraph explanations. If the explanations are not accurate, the augmented samples may not be label-preserving.
2. The EA-DA approach may be more complicated to implement compared to traditional DA techniques. The justification of the practical usefulness of the proposed methods is needed.

1. While the paper includes thorough theoretical discussions, it lacks specific details on the practical implementation of key components:

• Classifier $f(\cdot)$: There is no description of the pre-training process.
• Explainer $\Psi(\cdot)$: The authors mention that real-world applications may not have ground-truth explanations available and the explainer is pre-trained (lines 391- 392) in this paper. However, the author does not provide clear details on how this pre-training addresses such limitations. Additionally, important aspects of the explainer’s architecture, including model type and learning approach, are omitted. These design details are crucial to understanding the overall framework, just as much as the theoretical foundations.

2. In experimental settings, it is stated that GNNExplainer [1] and PGExplainer [2] are adopted as the explainer $\Psi(\cdot)$ instead of designing a novel one, raising a few concerns:

• It is uncertain whether GNNExplainer and PGExplainer align with the assumptions and theoretical conclusions presented in earlier sections.
• Moreover, the potential increase in computational complexity, such as training time, due to EA-DA is unclear. An efficiency comparison would be helpful to determine whether EA-DA effectively balances performance and resource consumption.

1. One potential limitation of EA-DA is its reliance on the quality of the explanation subgraphs. The method assumes that the explanation subgraph effectively represents the most important parts of the graph for classification. However, existing graph explanation methods (e.g., GNNExplainer, PGExplainer) are not perfect, and their outputs may be noisy or incomplete. If the explanation subgraphs are inaccurate, the augmented data might lead to performance degradation. The paper could benefit from a sensitivity analysis showing how robust the method is to low-quality explanations.

2. While the method focuses on perturbing non-explanatory edges, the diversity of the augmented samples may be limited. The current perturbation strategy modifies the graph structure only by adding or removing edges, which may not fully explore the potential augmentation space.

3. The process of generating explanation subgraphs for each sample could be computationally expensive, especially for large graphs or large datasets. Although the authors do not explicitly discuss the computational cost of their method, this could be a practical limitation in real-world scenarios where efficiency is critical. A discussion of the trade-offs between the increased computational complexity of EA-DA and the performance gains it offers would be useful.

4. The paper demonstrates the effectiveness of EA-DA on several benchmark datasets, but it is unclear how well this method would generalize to very large-scale graphs, dynamic graphs, or graphs with different structural properties (e.g., dense graphs). A discussion or analysis of the method's scalability and applicability to a broader range of tasks would strengthen the paper's argument.

The experimental section of this paper does not compare with many existing articles on graph data augmentation, which I believe is a significant oversight and a key issue. The list of articles is as follows:

[1] LSPAN: Spectrally Localized Augmentation for Graph Consistency Learning. [2] Graph random neural networks for semi-supervised learning on graphs. [3] Spectral augmentation for self-supervised learning on graphs. [4] Dropedge: Towards deep graph convolutional networks on node classifcation. [5]Data augmentation for graph neural networks.

In terms of datasets, commonly used graph datasets such as CORA, CITESEER, PPI, BLOGC, FLICKR, and AIR-USA were also not considered, which I believe undermines the persuasiveness of the study.

• The method necessitates training the model twice, adding to the computational cost. Specifically, it requires an initial pretraining phase on the original dataset to obtain a functional model, followed by training an explainer on this pretrained model. After generating the augmented dataset based on this explainer, the model is then retrained on both the original and augmented datasets, effectively doubling the training effort.

• The augmentation technique is highly model-dependent, meaning that the generated augmentations may not be transferable to other models, or at least this option is not discussed in the paper.

Minor:

• Including references to recent relevant works, such as GeoMIX [1], and adding recent methods like [2] to Table 1 would enhance the paper by providing a more comprehensive comparison with existing approaches.

[1] Zhao, Wentao, et al. "GeoMix: Towards Geometry-Aware Data Augmentation." Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 2024. [2] Ling, Hongyi, et al. "Graph mixup with soft alignments." International Conference on Machine Learning. PMLR, 2023.