Since data mining problems contain a large amount of data, sampling is a necessity for the success of the task. Decision trees have been developed for prediction, and finding decision trees with smaller error rates has been a major task for their success. This paper suggests a structural sampling technique that is based on a generated decision tree, where the tree is generated based on fast and dirty tree generation algorithm. Experiments with several sample sizes and representative decision tree algorithms showed that the method is more effective with respect to decision tree size and error rate than conventional random sampling method especially for small sample size.