Abstract: A great number of machine learning methods have been successfully applied for customer credit risk assessment cases, and support vector machine (SVM) is considered as an "off-the-shelf" supervised learning algorithm to solve classification problem by many researchers. Unfortunately, SVM fails to provide excellent enough classification performance when the data set is imbalanced, i.e., the accuracy of the majority class is usually much higher than that of the minority class due to the shifting of the hyper-plane. In most cases, people pay more attention on the minority class such as fault diagnosis and credit default. Thus, a Synthetic Minority Over-sampling Technique (SMOTE) is presented to deal with the imbalanced classification by generating new samples in the whole minority class. However, in the process of solving SVM by Sequential Minimal Optimization (SMO) algorithm, only those support vector samples x_#em/em# with the corresponding α_#em/em# > 0 can affect the position of the hyper-plane while the samples far from the hyper-plane have no influence on the final result. It is obvious that the classic SMOTE algorithm can generate more redundant samples which are far from the hyper-plane. In this article, an improved method for classic SMOTE algorithm is proposed that SMOTE is looped and only misclassified samples in the previous loop are selected to be processed in the next loop until the minority class outnumbers the majority class or all minority class samples are correctly classified. In the empirical study, a data set granted by a micro-credit company in Jiangsu Province is studied. The data set originates from a company that provides loans to local individuals and enterprises for the house condition improving, farm production expanding, business operating and so on. The customers' information are analyzed according to the characteristics of micro-loan industry, and a credit risk assessment index system is suggested from four aspects with sixteen attributes in this paper. G-mean and F-measure score are used to evaluate the classification performance of the minority class, which is the accuracy of detecting default customers in this case. The results show high prediction accuracy of default customers, indicating the effectiveness of our method on credit risk assessment.