Abstract: This paper investigates the interpretability of graph neural networks in recommendation systems by converting the recommendation problem into a graph classification problem and using interpretable graph neural networks to interpret the recommendation system. To enhance understanding of the graph patterns extracted from the interpreted models, this paper designs a visual analysis system, that explores three levels: single user, user group, and multiple user groups. The system also provides a visual verification of the interpretation's reliability and helps analysts explore the recommendation patterns of graph neural networks. Finally, the effectiveness of the system was verified with real Douban movie data.