基于巴氏距离的特征选择

FEATURE SELECTION BASED ON BHATTACHARYYA DISTANCE

巴氏(Bhattacharyya)距离与分类错误概率上界有直接关系.本文提出面向不同分布的多类问题的基于巴氏距离的特征选择.在正态分布条件下,使错误概率上界最小.基于巴氏距离的特征选择首先把使总分类错误概率上界最小问题,转化为非线性矩阵方程求解问题.然后通过解矩阵方程的迭代算法和正交化处理,取得变换矩阵的最优解.通过分析和实例可见基于巴氏距离的特征选择是在一定条件下最好的特征选择.

A feature selection based on Bhattacharyya distance for minimizing the sum of upper bound of error probability of every two class pair in subspace is presented. The key of feature selection on Bhattacharyya distance is to change the problem of minimizing the criterion to a problem of solving nonlinear matrix equation with a recursive algorithm. The theoretical analysis and experimental results show that the performance of proposed feature selection is conditionally superior to the performance of any previous one.