Abstract: A study on the interpretability of graph neural networks in recommendation systems. Starting from the interpretable model, we convert the recommendation problem into the graph classification problem, and use interpretable graph neural networks to explain the recommendation system, breaking through the previous situation in which explanations in recommendations were mostly at the instance level. The study explores multi-granularity explanations in the recommendation scenario from the instance level to the group level. In addition, to enhance the understanding of the graph patterns extracted by the interpretable model, a visual analysis system is designed to better understand the graph patterns and the model interpretation process. The visual system is explored from three levels: single user, user group and multiple user groups, which facilitates users to explore the recommended pattern of graph neural networks, thereby verifying the reliability of the explanation. Finally, the study designs and implements comparative experiments on two real data sets, Douban movie and Last-FM, in which the extracted graph pattern is used to improve and adjust the training sets. The experimental results show that the recommendation evaluation indicators are improved, which further proves the reliability of the explanation from a quantitative perspective.