Virtual Sample Generation Approach for Imbalanced Classification

Imbalanced classification problem is a hot topic in machine learning and data mining. The traditional classification algorithms assume that class distribution is balanced and the effect is not ideal when handling imbalanced datasets. In this paper, the support vector machine is used as basic classifier and a virtual sample generation method based on support vector is proposed to solve the problem of imbalanced classification and to improve the recognition rate of the minority class according to the characteristic that support vector machine is a classifier that relies heavily on support vectors. Firstly, support vector machine is used to learn training set to obtain support vectors of the minority class. Then, a certain number of virtual samples are generated around the support vector of the minority samples through the smoothness hypothesis to balance the data set. The generated samples can conform to the statistical characteristics of the original training data, which proves the rationality of the generated virtual samples. Finally, the new dataset is learned by support vector machine. Experimental results show that the method is effective in both artificial datasets and UCI standard datasets.