Graph-CoRe: Graph Representation Learning with Contrastive Subgraph Replacement

The interpretability of Graph Neural Networks (GNNs) is crucial for enhancing trustworthiness and practical utility in various domains like finance, healthcare, and transportation. Given the unique challenges posed by graph data, traditional interpretability methods fall short in deciphering the intricate workings of GNNs. In this study, we propose a subgraph replacement strategy combined with supervised contrastive learning to improve the interpretability and predictive capabilities of GNNs when processing graph data. By introducing perturbations during training and addressing data biases, our approach enhances model robustness, stability, and key feature identification. Moreover, by mitigating label imbalance issues through tailored data augmentation techniques, we ensure a more balanced and accurate prediction model. Experimental results demonstrate the effectiveness of our proposed method in outperforming other subgraph representation techniques across diverse datasets, showcasing its utility in enhancing model interpretability and performance.