Abstract: In order to study each feature difference of sample in the Least Squares Support Vector Machine （LSSVM）, Gaussian kernel with multiple parameters is introduced. Through analyzing the geometric significance of kernel polarization, a feature selection algorithm based on kernel polarization gradient iteration optimization of multiple parameters Gaussian kernel is proposed. Firstly, by using the gradient iteration of kernel polarization algorithm, the important degree of each feature in the sample is measured. Then, the feature selection of LSSVM is selected by the importance of feature characteristics. Finally, the selected features in the LSSVM is trained and tested. The name of the method is called KP_LSSVM. The experimental results on the UCI dataset show that the proposed method can obtain more accurate classification results than the PCA_LSSVM, KPCA_LSSVM and LSSVM methods, and verify the effectiveness of the proposed method.

Key words: feature selection, Least Squares Support Vector Machine（LSSVM）, kernel polarization

摘要： 为了对最小二乘支持向量机中样本的各个特征的差异性进行研究，引入了多参数高斯核，在分析核极化几何意义的基础上，提出了基于核极化梯度迭代优化多参数高斯核的特征选择算法。利用核极化梯度迭代算法对样本中每个特征的重要性程度进行测定；按特征的重要性大小进行LSSVM样本的特征选择；运用LSSVM对选出的特征子集进行训练和测试，称该方法为KP_LSSVM。UCI数据集上的实验结果表明，相较于PCA_LSSVM、KPCA_LSSVM和LSSVM方法，提出的方法可以取得更为准确的分类结果，验证了该方法的有效性。

关键词: 特征选择, 最小二乘支持向量机, 核极化