Abstract: When using graph neural networks for graph or subgraph-level tasks, identifying important causal subgraph structures within the graph data can improve model generalization and interpretability. However, existing methods do not consider the difference between causal subgraphs and environmental subgraphs when disentangling them, which limits the performance of the model in excavating causal subgraphs. Based on this, a model that applies contrastive learning to enhance the distinction between causal and environmental subgraphs is proposed. Experiments results show that this model achieves higher prediction accuracy in graph classification tasks compared to traditional methods, excavates more significant causal subgraphs, and offers better generalization and explainability.