Towards Self-Interpretable Graph Neural Networks via Augmentation-Contrastive Learning

Graph Neural Networks have demonstrated superior performance in downstream tasks, yet the decision-making mechanism lacks interpretability. Self-interpretable methods have been proven to unveil the internal decision-making mechanism and obtain reliable explanations. However, the conflict between task and interpretability performance is a widely recognized challenge. Additionally, most methods lack a collaborative interpretation of the graph structure and node features. This paper proposes a self-interpretable ACIG model based on a dynamic masking-contrastive learning mechanism, which improves task accuracy and the credibility of explanations through joint structure-feature modeling. Attention mechanism is used to learn importance probabilities, and then dynamic masking is applied to generate augmented samples for contrastive learning. End-to-end optimization is adopted to explicitly associate the importance probabilities with model performance, and sparse regularization is introduced to enhance computational efficiency and the readability of interpretations. Experiments have verified that ACIG jointly improves task performance and interpretability accuracy, providing concise and precise decision-making bases for downstream tasks.