CAJ | 학술논문

目前,在基于SVM的分解多分类方法中,各个子分类器都采用相同的核参数进行整体优化.如果采用不同的核函数参数分别优化各个子分类器,相当于在不同的特征空间进行分类,此时,能否直接采用求输出最大值(MaxWin)等判别策略,需要研究各分类器的输出是否可比.利用相对间隔对此问题进行深入的研究,说明在基于SVM的"一对多"等多分类方法中,采用不同的核参数分别优化各个子分类器,其决策函数的输出结果仍是可比的,且具有更好的泛化能力.
목전,재기우SVM적분해다분류방법중,각개자분류기도채용상동적핵삼수진행정체우화.여과채용불동적핵함수삼수분별우화각개자분류기,상당우재불동적특정공간진행분류,차시,능부직접채용구수출최대치(MaxWin)등판별책략,수요연구각분류기적수출시부가비.이용상대간격대차문제진행심입적연구,설명재기우SVM적"일대다"등다분류방법중,채용불동적핵삼수분별우화각개자분류기,기결책함수적수출결과잉시가비적,차구유경호적범화능력.
There usually using the same kernel parameter in multi-classification methods based on SVM such as one versus one(1-v-1)and one versus rest(1-v-r).Optimizing each sub-classifier under different kernel parameter is equal to classifying in different feature space.Then the comparability of each sub-classifier′s output should be analyzed before using the MaxWin strategy directly.In this paper,a relative margin is introduced to analyse the outputs of sub-classifiers.Results show that using different kernel parameters to optimizing each sub-classifier respectively,the outputs of decision function are still comparablk,and furthermore,it performs the better generalization ability.