Abstract: In view of the fact that the classification performance of current classification algorithms is not ideal for the unbalanced dataset,through combining supervised learning and unsupervised learning,this paper proposed a sub-sampling method based on centroid,namely ICIKMDS.In practical applications,some data are not easily to be obtained or different types of data are different in quantity,resulting in uneven distribution of data,such as the disproportion of the sufferer and the normal people in the detection of diseases,the disproportion of the fraud users and normal users in credit card fraud and so on.The new method solves the disproportion problem of dataset well.In this method,the initial centroid is obtained by solving the Euclidean distance between samples,and then the k-means algorithm is used to cluster the large-class sample sets to make the disproportionate dataset more balanced in distribution,effectively improving the effect of classifiers.The proposed method makes the classification accuracy of the classifier much better than that of random under-sampling and SMOTE algorithm on the subclass of test set,and its accuracy on the whole test set has little difference from other algorithms.

Key words: k-means, SMOTE algorithm, Unbalanced, Under-sampled