核函数选择方法研究

核函数的选择对支持向量机的分类结果有着重要的影响,为了提高核函数选择的客观性,提出了一种以错分实例到支持向量所在界面的距离来表示错分程度,并基于此进行秩和检验的核函数选择方法.通过与K-折交叉验证、配对t测试等参数检验的统计方法进行对比分析,对9种常用核函数的分类能力在15个数据集进行了定量研究.与参数检验方法不同,秩和检验并未假定数据的分布情况(很多情况下数据并不满足假定的分布),而且数据实验证明,秩和检验不但能够对核函数的分类能力进行客观评估,而且在某些数据集上还能产生更好的核函数选择效果.

Research on Selection Method of Kernel Function

The selection of kernel functions has an important influence on the classification results of support vector machines. This paper proposed a kernel functions selection method based on rank sum test in order to enhance the selection objectivity, where the error degree adopted in the rank sum test was represented by the distance between the error instance and the interface of support vectors. By comparing with other statistical methods, such as K-folding cross validation and paired t test, the classification abilities of nine common kernel functions were quantitatively studied based on 15 datasets. Different from parameter test methods, the rank sum test does not assume the data distribution(in some cases data cannot satisfy the assumed distribution), the experimental data proves that the rank sum test not only can objectively evaluate the classification abilities of kernel functions, but also can produce better selection results on some data sets.