Improving SVM classification with imbalance data set

Abstract

In view of inconsistent problems caused by that Synthetic Minority Over-sampling Technique (SMOTE) and Support Vector Machine (SVM) work in different space, this paper presents a kernel-based SMOTE approach to solve classification with imbalance data set by SVM. The method first preprocesses the data by oversampling the minority instances in the feature space, then the pre-images of the synthetic samples are found based on a distance relation between feature space and input space. Finally, these pre-images are appended to the original dataset to train a SVM. Experiments on real data set indicate that compared with SMOTE approach, the samples constructed by the proposed method have the higher quality. As a result, the effectiveness of classification by SVM on imbalance data set is improved. © 2009 Springer-Verlag Berlin Heidelberg.