核极化优化多参数高斯核的特征选择算法

Feature Selection Algorithm of Gaussian Kernel with Multiple Parameters Optimized by Kernel Polarization

为了解决支持向量机中多特征样本的特征选择问题,鉴于多参数高斯核中的多参数的不同取值可以区分和体现样本中各个特征的重要性差异,在深入分析核极化的几何意义和多参数高斯核特点的基础上,提出了基于核极化梯度优化多参数高斯核的特征选择算法。首先,利用核极化的梯度迭代算法来寻求多参数高斯核的最优多参数值,然后,以优化的多参数大小为基准,进行样本特征的重要性程度强弱标定,进而,采用特征重要性指标来执行SVM的特征选择。最后,将选择出的样本特征子集应用于SVM分类器中。UCI数据的实验结果表明,相较于PCA-SVM、KPCA-SVM和经典SVM方法,所提出算法的分类正确率更高,验证了核极化与多参数模型特征选择算法的有效性。

For the issue of feature selection for support vector machine（SVM）,considering that multiple parameters Gaussian kernel can differentiable and emphasize the different importance of each feature,the paper proposed a feature selection algorithm of Gaussian kernel with multiple parameters based on the optimization of the gradient of kernel polarization,at the base of the geometrical meaning of kernel polarization. First,search the optimal parameter values of Gaussian kernel with multiple parameters using the gradient of kernel polarization,which is a kernel evaluation criterion. Then,demarcate the important degree of the sample features as the benchmark of the optimized multiple parameters values,and then,perform the feature selection of SVM by using the importance index. Last,apply the selected feature subset to SVM classifier. The proposed algorithm was tested on several different datasets of UCI. From the experimental results,we can know that the proposed algorithm outperforms PCA-SVM,KPCA-SVM and SVM. In other words,more accurate outcomes of identifying and classification can be obtained by the proposed algorithm,demonstrating the effectiveness of kernel polarization and the feature selection algorithm with multiple parameters.