In data mining, a basic assumption is that the data distribution of training set samples are consistent with that of test set samples. But as data volumes increase, how to find out representative data in huge amounts of data becomes particularly difficult. By studying existing data selection methods, we find that it is difficult to evaluate their sampling effect because they are not integrated with the data mining tool, such as simple random sampling and progressive sampling. Due to contingency factors and uncertainty, it is difficult to guarantee the basic assumptions of data mining, which also makes the generalization error of the model larger. In order to solve these problems, we put forward a structured data sampling method based on double decision tree. Firstly, we generate a decision tree with the C4.5 algorithm, which is used to select appropriate data and data collection points in the data source. Then, we generate another decision tree to evaluate the quality of the selected data set and achieve data sampling of high efficiency and high quality. Experiments show that compared with random sampling, the accuracy of the model based on our sampling is improved obviously.