GraphEx: A User-Centric Model-Level Explainer for Graph Neural Networks

Sayan Saha, MONIDIPA DAS, Sanghamitra Bandyopadhyay

Abstract: With the increasing application of Graph Neural Networks (GNNs) in real-world domains, there is a growing need to understand the decision-making process of these models. To address this, we propose GraphEx, a model-level explainer that learns a graph generative model to approximate the distribution of graphs classified into a target class by the GNN model. Unlike existing methods, GraphEx does not require another black box deep model to explain the GNN and can generate a diverse set of explanation graphs with different node and edge features in one shot. Moreover, GraphEx does not need white box access to the GNN model, making it more accessible to end-users. Experiments on both synthetic and real datasets demonstrate that GraphEx can consistently produce explanations aligned with the class identity and can also identify potential limitations of the GNN model.