A New Cost-sensitive SVM Algorithm for Imbalanced Dataset

Abstract

Support Vector Machine(SVM) is a popular machine learning algorithm for its excellent generalization ability. However, similar to most of traditional algorithms, the proposal of SVM is based on an assumption that the dataset is nearly balanced, and when SVM is applied in imbalanced dataset, the result may be bias towards majority class which leads to poor performance. To solve this problem, a new cost-sensitive SVM algorithm based on samples density are proposed in this paper. In the proposed algorithm, samples’ weights are depended on sample density estimated from Kernel Density Estimation(KDE) method, and furthermore, the samples’ weights are modified to enlarge the weights of border samples and reduce the weights of noise samples based on Support Vector Data Description(SVDD) algorithm. The experiments result shows that the proposed algorithm could achieve satisfactory performance.