基于支持向量域的分离超平面

Separation hyperplane based on support vector domain description

为了提高支持向量机(support vector machines,SVM)和支持向量域分类器(support vector domainclassifier,SVDC)的精度,减少SVM的训练时间,建立一种分离超平面。该算法首先通过确定参数以减少每类的野点。然后分别对每类样本应用support vector domain description(SVDD)算法分别进行描述以求取两个超球的球心和边界向量;根据到这两个超球心的最大距离和为准则来确定出分类超平面的法向量。最后在两球相邻边界中间点建立一个分离超平面。该方法是从整体上考虑分类信息,是尝试SVDD和SVM的结合。实验结果表明,提出的算法与SVDC相比,精度有了显著提高;与SVM相比,不仅精度有所提高,而且训练速度随着样本容量的增大也有很大提高。

To improve the precision of both support vector machines （SVM） and support vector domain classifiers （SVDC） and to reduce training time that SVM taken in training, a separation hyperplane is constructed. Firstly, by determing the parameters to reduce outliers. Then, the description of the training samples from each class respectively is presented to obtain two hypersphere centers and boundary vectors by using support vector domain description （SVDD）. The normal vector of separation hyperplane is determined based on the rule of the maximal sum of distance to the two classes sphere centers. Finally, a separation hyperplane is constructed at the middle point between these two hypersphere boundaries. The method that considers wholly the classification information from the two classes of samples is an attempt to incorporating SVDD and SVM. The experiments on several real data show the proposed algorithm has much higher accuracy than support vector domain classifier （SVDC）. Compared with support vector machines （SVM）, the proposed algorithm has higher accuracy and can improve the training speed of greatly as increase of the size of sample.