KnowGNN: a knowledge-aware and structure-sensitive model-level explainer for graph neural networks

Abstract

Model-level Graph Neural Network (GNN) explanation methods have become essential for understanding the decision-making processes of GNN models on a global scale. Many existing model-level GNN explanation methods often fail to incorporate prior knowledge of the original dataset into the initial explanation state, potentially leading to suboptimal explanation results that diverge from the real distribution of the original data. Moreover, these explainers often treat the nodes and edges within the explanation as independent elements, ignoring the structural relationships between them. This is particularly problematic in graph-based explanation tasks that are highly sensitive to structural information, which may unconsciously make the explanations miss key patterns important for the GNNs’ prediction. In this paper, we introduce KnowGNN, a knowledge-aware and structure-sensitive model-level GNN explanation framework, to explain GNN models in a global view. KnowGNN starts with a seed graph that incorporates prior knowledge of the dataset, ensuring that the final explanations accurately reflect the real data distribution. Furthermore, we construct a structure-sensitive edge mask learning method to refine the explanation process, enhancing the explanations’ ability to capture key features. Finally, we employ a simulated annealing (SA)-based strategy to control the explanation errors efficiently and thus find better explanations. We conduct extensive experiments on four public benchmark datasets. The results show that our method outperforms state-of-the-art explanation approaches by focusing explanations more closely on the actual characteristics of the data.