基于超球结构的渐进直推式支持向量机

Progressive transductive Support Vector Machines based on hypersphere structure

针对渐进直推式支持向量机(Progressive Transductive Support Vector Machines,PTSVM)算法训练速度慢,学习性能不稳定的问题,提出一种基于超球结构的渐进直推式支持向量机算法。该算法首先利用支持向量域描述(Support Vector Domain Descrip-tion,SVDD)得到包含每个类别的有标签样本点的最小包球,选择这个包球边界附近的无标签样本点进行标注,然后对目前所有有标签的样本点进行SVM训练。实验结果表明该算法不仅能保持甚至提高算法的精度,更重要的是能大大提高训练速度。

Progressive Transductive Support Vector Machines（PTSVM） has some drawbacks such as slower training speed,and unstable leaming performance,a Progressive Transductive Support Vector Machines learning algorithm based on hypersphere structure is proposed.Labeled data of each class are described by using Support Vector Domain Description（SVDD） and the corresponding smallest enclosing hypersphere is obtained.Then,the method selects new unlabeled samples located near the boundary of the hypersphere.Finally,labeled data available is used to train standard SVM.Experiment results show the method can improve greatly the computing speed.Moreover,it can keep,in fact sometimes improve,the classification accuracy in general.