Robust explanations for graph neural network with neuron explanation component

Abstract

Graph neural networks (GNNs) have been successfully applied to a variety of graph-structure analysis tasks. Besides their outstanding performance, the explanation for GNNs' predictions is also an inextricable problem, which hinders trust in GNNs under practical scenarios. Consequently, great efforts have been made for interpreters of GNNs to understand their behavior. However, the existing works are still suffering two main problems: (i) explanation-shifting in normal explanation - the explanations provided by the interpreters are insufficient to precisely explain the behavior of the GNNs; (ii) gullibility failure in adversarial detection - the interpreters are easily bypassed by well-designed adversarial perturbations, resulting in the omission of anomalies. To address these issues, we propose a robust interpreter for GNN, named Neuron Explanation Component (NEC), from the perspective of the model neuron activation pattern. It measures the difference in GNNs' neuron path distribution between subgraphs and the original graph to generate explanations for the model's prediction. NEC outperforms previous works in explanation accuracy, robustness against adversarial attacks and transferability among different GNN's interpreters. Extensive evaluations are conducted on 4 benchmarks, 6 interpreters and 2 scenarios (i.e., normal explanation and adversarial detection). Significant improvements in explanation ability and adversarial detection performance demonstrate NEC's superior performance.