基于距离核函数的除噪和减样方法  

A denosing and sample-reducing method based on kernel function with distance performance

作　　者：刘万里[1] 刘三阳[1] 薛贞霞[1] 

机构地区：[1]西安电子科技大学应用数学系

出　　处：《系统工程理论与实践》2008年第7期160-164,共5页Systems Engineering-Theory & Practice

基　　金：国家自然科学基金(60574075,60674108)

摘　　要：在使用支持向量机(SVM)分类时,存在以下两个问题:一是当存在噪点时,分类的精度低;二是对大规模样本集,训练时所需内存空间较大,运行时间较长.针对以上问题,给出一种基于具有距离性能的核函数的减样方法,称为删减法(DRM).该方法定位定量分析了噪点及多余样本点的一般比例.在应用时,分三步进行:首先根据小概率原理给出一小阈值删除噪点;然后给出一个较大阈值减去同类中心附近的大量多余的样本点;最后以另一个大的比例减去位于距异类中心较远的对分类不起作用的样本点,以便提取具有代表性的边界向量.试验结果检验了该方法的有效性,即,既减少了训练时间,又提高了分类精度.There exist two problems in using support vector machine （SVM） as follows： One is the lower classification accuracy when existing noises. The other is larger memory needed and longer time taken in training. For the above problems, a denosing and sample-reducting method, named deletion-reduction method （DRM） based on the kernel with distance performance, is proposed. The general proportion of the noises and excrescent sample points are analyzed in locality and quantity. The three steps are needed in application： Firstly, one small threshold is given to delete those noisy points lie on the adjacent boundary. Secondly, a large number of redundant sample points are reduced near the center of the same classes based on one large threshold. Finally, the other large proportion is decided to reduce those sample points lie on the further from the different class center so that the representative boundary vectors can be extracted. The results of experiments show the efficiency of the proposed method, i.e. it can reduce training time and can also improve classification accuracy.

关 键 词：距离核函数 噪点 减样 小概率原理 支持向量机 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]