To solve the problem of data drift and asymmetric misclassification costs in customer segmentation, a cost sensitive learning method integrated with two-step cluster is proposed. This method firstly applied kmeans cluster by the posterior probability distribution of give region to group similar regions together,and then used cost-sensitive support vector machine to find customer segmentation for each region-group. The results show that the cluster based on similarity of customer segmentation structure can improve the total accuracy and the proposed cost-sensitive support vector machine is an effective method to distinguish high value customers compared to the original support vector machine.