범주 불균형 문제 해결을 위한 새로운 형태의 SVM 기반 AdaBoost

Abstract

We proposed the AdaBoost-type boosting algorithm applied to SVM by categorizing instances based on the margin of SVM and assigning the cost differently to the related instances by considering the categories divided, classes, and the classification result of a previous SVM. We defined borderline and noisy instances based on the margin of SVM and set the additional cost to borderline instances and positive noise. The cost for a borderline instance was inversely proportional to the portion of its own class in the overlapped region. The cost for a positive noise was proportional to the portion of positive noise among the minority class. The cost was used additionally to learning a SVM, apart from instances weight in an AdaBoost algorithm. We compared the proposed method with various 10 methods for 10 class-imbalanced datasets. The proposed method showed high values of F-measure and AUC for most data sets. In other words, the proposed method improved the classification performance more than the trade-off between true positive and false negative. Therefore we concluded that the proposed method relaxed the imbalance ratio between classes as well as three degradation factors, the overlap, small disjunct and data shift problems significantly.