At present, financial fraud detection is interested by many machine learning researchers. This is because of existing a big ratio between normal transactions and abnormal ones in data set. Therefore, a good result of prediction rate does not mean that there is a good detection result. This is explained that the experimental result might be effected by the imbalance in the dataset. Resampling a dataset before putting to classification process can be seen as the required task for researching in financial fraud detection area. An algorithm, so-called as MASI, is proposed in this paper in order to improve the classification results. This algorithm breaks the imbalance in the data set by re-labelling the major class samples (normal transactions) to the minor class ones basing the nearest neighbor’s samples. This algorithm has been validated with UCI machine learning repository data domain. Then, the algorithm is also used with data domain, which is taken from a Vietnamese financial company. The results show the better in sensitivity, specificity, and G-mean values compared to other publication control methods (Random Over-sampling, Random Under-sampling, SMOTE and Borderline SMOTE). The MASI also remains the training dataset whereas other methods do not. Moreover, the classifiers using MASI resampling training dataset have detected better number of abnormal transactions compared to the one using no resampling algorithm (normal training data).