Abstract:

Outlier detection is an important problem, and it is necessary to solve both problems of  imbalanced dataset as well as cost sensitive instantly. First a new re-sampling algorithm based on a modified oneclass support vector machine (SVM) is presented, and then a two-stage outlier detection approach combining with cost sensitive SVM is designed. In the first stage, various low weights are set for outliers, and some common points are removed proportionally by the hyper-plane in feature space, and could  overcome the effect of  overlapping data points. In the second stage, a receiver operating characteristic (ROC) analysis is applied to  select  the optimum parameters of cost sensitive SVM in limited grid scope, and finally the detection decision function is obtained after adjusting the threshold. Experiment results show that the proposed method can improve  the classification accuracy and decrease the misclassification cost effectively.