Abstract: In view of the poor classification performance of graph neural network on large-scale graphs, it is unable to quickly form the embedding of unknown nodes and edges, and it is easy to lose the important features of the graph. A graph classification network model based on inductive learning and self-attention pooling is proposed. On the one hand, the inductive learning method after the improved aggregation function was used to quickly embed the node features of the graph. On the other hand, the self-attention pooling method was used to retain the important features of the graph. A hierarchical framework suitable for extracting large-scale graph information was used for downstream graph classification task. The experimental results show that the accuracy of this method is about 2%~10% higher than that of other graph classification models under the same common data set.