Increasing Efficiency of SVM by Adaptively Penalizing Outliers

In this paper, a novel training method is proposed to increase the classification efficiency of support vector machine (SVM). The efficiency of the SVM is determined by the number of support vectors, which is usually large for representing a highly convoluted separation hypersurface. We noted that the separation hypersurface is made unnecessarily over-convoluted around extreme outliers, which dominate the objective function of SVM. To suppress the domination from extreme outliers and thus relatively simplify the shape of separation hypersurface, we propose a method of adaptively penalizing the outliers in the objective function. Since our reformulated objective function has the similar format of the standard SVM, the idea of the existing SVM training algorithms is borrowed for training the proposed SVM. Our proposed method has been tested on the UCI machine learning repository, as well as a real clinical problem, i.e., tissue classification in prostate ultrasound images. Experimental results show that our method is able to dramatically increase the classification efficiency of the SVM, without losing its generalization ability.