Extracting Global Logical Rules for GNN Node Classification

Researchers have introduced a novel logic-based framework aimed at enhancing the interpretability of Graph Neural Networks (GNNs), specifically for node classification tasks within Simple Graph Convolution (SGC) networks. While GNNs excel in performance, understanding their decision-making process remains a challenge. This new approach addresses the limitation of previous methods that often produce node-specific explanations with redundant information. Instead, it computes minimal abductive explanations for each node, identifying the smallest set of node-feature pairs necessary to preserve the predicted class. These concise explanations are then utilized to train decision trees, from which global logical rules are extracted. Experiments on benchmark datasets confirm that this framework generates compact, globally applicable rules that maintain high fidelity to the original SGC model's predictions, offering a clearer understanding of how GNNs classify nodes.