CStaG: A Causality-Inspired Stable Generative Model for the Interpretation of Graph Neural Networks

Graph neural networks (GNN) are receiving increasing attention as a mainstream method for processing graph tasks. Meanwhile, there is a lack of understanding of their operational mechanisms when people use it. Some of the current explanation methods, which are also mainly based on statistical correlation for corresponding explanations, find it difficult to explore the deeper operation mechanism of graph neural networks. In this study, we attempt to understand the reasons for its decisions from a causal perspective, thus providing a deeper understanding. Inspired by this point, we propose a new casual-based explanation framework called CStaG for generating explanations for a pre-trained GNNs model based on learned latent factors. Using the Structural Causal Model (SCM), we eliminate the confounding effect via backdoor adjustment formula. In addition, it does not rely on statistical relationships, including both linear and nonlinear relationships, of the features. In this study, a proof-of-concept for CStaG is presented in the context of canonical graph classification problems. Empirical evidence indicates that CStaG is proficient in identifying caudal semantics for the purpose of generating causal explanations, demonstrating a significant performance advantage over alternative methods.