基于相似度量的核主成分序列极小化方法

A Sequence Minimization Method Based on Matrix Similarity Measure and KPCA

研究优化主成份序列分类精度,针对支持向量机在小样本情况下泛化能力差的问题,为提高训练的准确率,提出了一种基于相似度量的核主成分序列极小化方法,方法在进行核主成分分析时,使用混合核函数,权值和形式参数是通过遗传算法,以矩阵相似性度量作为适应度,优化求得的,得到最有利于分类的核主成分空间。使用序列极小化方法对主成分做进一步的选择,降低输入空间的维数,同时由于是线性的支持向量机,不会增加学习机的VC维,从而提高了小样本情况下分类的准确率。通过实验证明改进方法是有效的。

Aiming at the problem that SVM used in small size samples has poor generalization ability,a kernel principal component sequence minimization method based on matrices similarity measure was presented.This method used a hybird kernel in principal component analysis.The kernel＇s weight and paraments were determined by genetic algorithm,while matrices similarity measure served as the fitness.Thus the kernel principal component space is most conducive to classification.When implementing classification,sequence minimization method is used to choose principal component further.Because it is a linear support vector machine,the VC dimensions will not be increased,which ensures the accuracy of classification.The experiment proved this method is effective.