一种基于数据分布的SVM核选择方法

A SVM Kernel Selection Approach Based on the Characteristics of Data Distribution

针对目前支撑向量机(SVM)核函数的选择没有统一规则的现状,提出一种结合数据分布特征进行SVM核选择的方法.首先,采用多维尺度(MDS)分析方法对高维数据集合理降维,提出判断数据集是否呈圆球分布的算法;然后,在得到数据集分布特征的基础上进行SVM核选择,以达到结合数据分布特征合理选择SVM核函数的目的.实验结果表明:呈圆球分布的数据集采用球面坐标核进行分类,识别率达到100%,训练时间最短,优于采用高斯核SVM及多项式核SVM的分类效果.

The kernel selection has no unified rules for support vector machine （SVM）. Based on the characteristics of dataset distribution, a new way to select the kernel function was presented. First dimension reduction of the high dimen- sional dataset was processed with multidimensional scaling （MDS） method. Then an algorithm was put forward, it was judged whether dataset is sphericity distribution. On the basis of determining sphericity distribution, how to select the kernel function was discussed, to achieve the purpose of selecting SVM kernel function with data distribution characteris- tics. The experimental results illustrate that the classification recognition rate of sphericity datasets reaches 100% with sphere kernel and the training time is the shortest. The classification effect is better than that of using gaussian kernel SVM and polynomial kernel SVM.