基于可信度的渐进直推式支持向量机算法  

Reliability-based Progressive Transductive Support Vector Machines

作　　者：薛贞霞[1,2] 刘三阳[1] 刘万里[1,3] 

机构地区：[1]西安电子科技大学应用数学系,陕西西安710071 [2]河南科技大学数学系,河南洛阳471003 [3]洛阳师范学院数学系,河南洛阳471022

出　　处：《厦门大学学报（自然科学版）》2008年第6期806-811,共6页Journal of Xiamen University：Natural Science

基　　金：国家自然科学基金(60674108,60705004)资助

摘　　要：针对渐进直推式支持向量机(Progressive transductive support vector machines,PTSVM)算法回溯式学习多,训练速度慢,学习性能不稳定的问题,提出一种基于可信度的渐进直推式支持向量机算法.该算法首先基于支持向量域描述(Support vector domain description,SVDD)对无标签样本点赋予一定的可信度,根据可信度选择新标注的无标签的样本点;其次利用支持向量预选取方法减少训练集的规模,对当前所有有标签的样本点用支持向量机(Support vector ma-chines,SVM)训练,最后重复上述过程从而求出最终的分类超平面.实验结果表明,与PTSVM相比,该算法不仅能较大幅度的提高算法的速度,更重要的是在一般情况下能提高算法的精度.Progressive transductive support vector machines （PTSVM） had some drawbacks such as more back learning steps, more slow training speed, and unstable learning performance. We proposed a reliability-based progressive transductive support vector machines learning algorithm. Firstly, based on support vector domain description （SVDD）, this algorithm applied reliability values to unlabeled samples. Next, new unlabeled samples was selected based on their reliability values. This approach also introduced pre-extracting support vector algorithm to reduce the calculation complexity,and then retrained the support vector machines（SVM） over current all labeled samples. Finally,we repeated the above process and obtained final classification hyperplane. Compared to PTSVM, the experiment results show that the method can improve greatly the computing speed. Moreover,it can enhance the classification accuracy in general.

关 键 词：半监督学习 支持向量机 直推式学习 支持向量域描述 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]