CausGNN: A Causal-Based Explanation Framework for Graph Neural Networks

Abstract

Graph Neural Networks (GNNs) are currently used in many real-world applications. With this notable spread, the development of sophisticated techniques for explaining their decisions becomes highly necessary. Although many works have been proposed with the aim of explaining their predictions, most of them generate explanations as subgraphs. In this paper, we argue that relying only on explanatory subgraphs is not sufficient. In this regard, we propose CausGNN: a causal explanation framework based on the structural model of causality. By adapting the definition of actual cause, our framework provides comprehensive explanations that incorporate both nodes features and edges in a complementary manner. Furthermore, as the need for robust explanations grows, we address this issue and show that the explanations provided by CausGNN are very robust to perturbations. Finally, CausGNN does not intend to compete with existing explanation frameworks for GNNs, but rather acts as a complementary tool.