Abstract:

Based on the statistical learning theory (SLT), the margin scale reflects the generalization capability to a great extent. Inspired by oneclass support vector machine (SVM), doublemargin SVM is put forward to classify two classes by two margins separately. Instances can be classified correctly as well as margin maximization, and its superiority is theoretical proved by both generalization performance and imbalanced class distribution. Experiment on benchmark data sets shows that classification margin obtained by doublemargin SVM is larger than SVM, improving the generalization apparently, and analysis on imbalanced data shows that it has a higher recognition ratio. Finally real intrusion detection data shows that the detection precision is increased by 2% against boundary samples selection method.

Key words: classification margin, generalization capability, doublemargin support vector machine