An Efficient Active Training Algorithm for SVM-Based Binary Classifiers

Pattern classification has been a key task in many scientific and engineering applications. In this paper, the problem of support vector machine (SVM) based binary classification is considered. Though the SVM is an effective supervised learning algorithm, its high computational complexity often limits its real-time implementation. The support vector machines require enormous amount of training time for problems of large magnitude. In this paper, a new efficient active training approach is proposed, which makes use of the fact that the SVM decision boundary is completely characterized by a small subset of the training data. The proposed method is evaluated both theoretically and experimentally using the benchmark KDD 1999 intrusion detection datasets. The proposed method is also compared with some of the existing training methods. The experimental results show that the proposed method can successfully reduce the training time without significantly degrading the classification performance.