Abstract: To solve the problems of the current graph pooling method based on the node sampling, such as the simplistic strategy of node importance evaluation and the massive loss of sub-structure feature information of graph, a sub-structure representative hierarchical pooling model based on node sampling dubbed as SsrPool was proposed. This method mainly includes a sub-structure representative node selection module and a sub-structure representative node feature generation module. First, considering both the structure and feature information of the graph, different methods were used to evaluate the node importance, and the collaboration of varying importance scores generated a robust node ranking to guide node selection in the sub-structure representative node selection module. Second, sub-structure features information was retained through feature fusion in the sub-structure representative features generation module. By combining the SsrPool model with existing graph neural networks, experimental results of graph classification on different public datasets demonstrate the effectiveness of SsrPool.