摘要
该文对最佳鉴别特征的最佳维数问题进行了详细的讨论 .文章首先对最佳维数问题进行了界定 ,然后指出了两种最佳特征维数为c- 1维的情况即以某些基于矩的可分性判据 (准则函数 )为优化目标的最优特征和以某些特殊的分类器错误率为优化目标的最优特征 .最后该文运用方差分析法对最佳鉴别特征进行特征选择使之代入最小距离分类器后识别率最大 .
This paper makes a detail discussion on the optimal dimension problem of optimal discriminant feature set, gives the condition under which the dimension of optimal features sets is c -1. That is the optimal features set which optimizes the discriminant criterion with the form f(M 1, …,M c,S 1,…,S c ), where M i and S i are the first and second order moments, and the optimal features set which optimizes the error rate of some special classifiers, such as Bayesian linear classifier, minimum distance classifier. Both of the optimal features are the posterior probabilities or their function. At last, variance analysis is introduced against estimation error to select the optimal discriminant features that minimize the error rate of minimum distance classifier. This method can also be used in other feature selection problems.
出处
《计算机学报》
EI
CSCD
北大核心
2003年第7期825-830,共6页
Chinese Journal of Computers
关键词
模式类别
线性分类器
最佳鉴别特征
维数问题
特征选择
方差分析
optimal discriminant features
optimal features
optimal dimension problem
feature selection
variance analysis
