支持向量机的一个边界样本修剪方法

A method of pruning edge samples for support vector machines

支持向量机仅仅由支持向量所决定,而支持向量来自于边界的样本,如果样本集中存在较多的噪音或孤立点,特别是两类样本过分交叉,都会降低支持向量机的推广能力。为了改善支持向量机的推广性能,文章提出一个支持向量机的边界样本修剪方法:首先对边界样本进行抽取,然后用RemoveOnly算法对边界样本进行修剪,修剪后的边界样本就是最终的支持向量机训练样本。实验结果表明,修剪方法可以让支持向量机的推广能力有不同程度的提高。

As a support vector machine（SVM） is determined only by support vectors（SVs）, which are a part of edge samples, its generalization ability may be decreased if the noise is too much or outlier samples are too many, especially the samples from different classes are intermixed excessively. In order to improve generalization performance of the SV, M, a method of pruning edge samples is presented. Firstly, some edge samples near to the optimal hyperplane are extracted, including SVs likely. Secondly, these samples are pruned with the RemoveOnly algorithm. Lastly, these pruned edge samples are trained. Experiments show that pruning can partly improve generalization performance of the SVM.