Decision Tree is an excellent method in classification,but bad at imbalanced datasets.Cost-sensitive decision tree can deal with the imbalanced datasets,which acts on the different attributes but the classes that the instances are belonged to.A novel method is proposed in this paper,named Adaptive Dynamic-cost Optimization Decision Tree(ADODT),which searches an optimal misclassification cost for every subset with optimization in a feasible misclassification cost interval,trains a prediction model.We empiri cally evaluate ADOCT with Decision Tree and Cost-sensitive Decision Tree over 8 realistic imbalanced bi-class datasets from UCI with AUC,Kappa,G-mean of F-measure,and G-mean of Precision and Recall.The experimental results show that ADOCT outperforms the other two methods over all the imbalanced datasets.