Combined Support Vector Machines For Classification

摘要

Support vector machines are an important method of knowledge discovery. Standard support vector machine have two drawbacks. One is that standard support vector machines are very sensitive to outliers or noises because of overfitting problem. The other one is that when training sets with uneven class sizes are used, the classification result based on standard support vector machines is undesirably biased towards the class with more samples in the training sets. In this paper, a new classification algorithm, called combined support vector machines which consists of a dual-weight support vector machine and a detection algorithm, is proposed to deal with these problems. The dual-weight support vector machine treats training data points and classes with different weights. In addition, the outliers and support vectors are detected by the proposed detection algorithm that is a hybrid method based on a proximal support vector machine cascaded with an algorithm of pruning redundant examples, called golden section rule. Experimental results indicate that the proposed combined support vector machines actually reduce the effect of outliers and uneven class sizes and yield higher classification rate than standard support vector machines do.