EGNAN: An Enhanced Method for Interpretable Graph Neural Networks

Graph Neural Networks (GNNs) are increasingly used in various domains. However, their black-box nature limits interpretability. In this work, we propose EGNAN, a new interpretable method for GNNs. We provide a theoretical analysis of EGNAN, describing its key components. EGNAN is compared with the existing explainable Graph Neural Additive Network (GNAN), evaluating its performance and interpretability. Through experiments, we assess the models across different data distributions and graph characteristics, including variations in homophily and class imbalance. The results show that EGNAN is a promising classification method, combining high accuracy with reduced execution times compared to GNAN, while still ensuring interpretability. Additionally, graphical representations clearly depict the relationships between the target variable, features, and graph structure, enhancing the model’s interpretability. This suggests that EGNAN could provide a valuable contribution to more interpretable and efficient graph-based learning models.