GNNAnatomy: Systematic Generation and Evaluation of Multi-Level Explanations for Graph Neural Networks (2406.04548v2)

Abstract: Graph Neural Networks (GNNs) excel in machine learning tasks involving graphs, such as node classification, graph classification, and link prediction. However, explaining their decision-making process is challenging due to the complex transformations GNNs perform by aggregating relational information from graph topology. Existing methods for explaining GNNs face key limitations: (1) lack of flexibility in generating explanations at varying levels, (2) difficulty in identifying unique substructures relevant to class differentiation, and (3) little support to ensure the trustworthiness of explanations. To address these challenges, we introduce GNNAnatomy, a visual analytics system designed to generate and evaluate multi-level GNN explanations for graph classification tasks. GNNAnatomy uses graphlets, primitive graph substructures, to identify the most critical substructures in a graph class by analyzing the correlation between GNN predictions and graphlet frequencies. These correlations are presented interactively for user-selected group of graphs through our visual analytics system. To further validate top-ranked graphlets, we measure the change in classification confidence after removing each graphlet from the original graph. We demonstrate the effectiveness of GNNAnatomy through case studies on synthetic and real-world graph datasets from sociology and biology domains. Additionally, we compare GNNAnatomy with state-of-the-art explainable GNN methods to showcase its utility and versatility.