In order to improve the classification accuracy of random forests algorithm, the decision trees in the random forest model are first sorted according to the AUC value of the classification performance evaluation index. And then the trees with high AUC value is selected to calculate the similarity matrix. Finally the decision tree is clustered according to the similarity matrix. So a new random forest model is generated by selecting the tree with the highest AUC value from each category and to achieve the goal of improving the accuracy of random forests algorithm. Experiments on UCI datasets show that the improved random forest algorithm has improved the highest classification accuracy of 2.91%.