An Improved Algorithm of Unbalanced Data SVM

Abstract

Since SVM is unfair to the rare class for the classification of unbalanced data, a new balancing strategy based on common strategy of undersampling the training data is presented. Firstly, the fuzzy C-means clustering algorithm is used to cluster the unbalanced data sets, and choose the negative class samples whose memberships are greater than a certain threshold (supposing the number of positive class samples is less than that of negative class samples). The selected samples and the positive class of original samples are combined into a new training data set. After that, the new data set are used to train a support vector machine. At last, the simulations on unbalanced data show that the proposed algorithm can compensate the ill-effect of tendency when support vector machine are utilized to deal with the unbalanced data classification. Moreover, compared with the traditional support vector machine and some other improved algorithm, the proposed algorithm performs superior classification ability.