Improved automatic filtering algorithm for imbalanced classification based on SVM-RFE

Xiaoqiang Li, Qing Shao, Jingjing Wang

Abstract: Almost all unbalanced classification algorithms focus on how to maximize the balance degree of the data set, which means to remove those negative samples that are useless for classifier training while keeping the positive samples and useful samples as many as possible. However, we find that the best balance degree is not necessary with the highest classification accuracy. In this paper, we propose a new method for imbalanced classification combined the SVM-REF (Support Vector Machine Recursive Feature Elimination) with automatic filtering algorithm. First, the SVM-RFE is applied to select the most discrimination features. Second, the combination of these features are used in the automatic filtering algorithm to extract the filtering rules, which will remove the samples that have no or negative effects on classifier training and testing on imbalanced data sets. Experimental results demonstrated that the proposed method can get higher classification accuracy. In addition, our approach can significantly shorten the training time.