一种基于最小分类错误率的改进型LDA特征选择算法

LDA是目前常用的较好的特征选择方法。然而散布矩阵不同时,LDA分类效果往往不理想。本文提出一种基于分类错误率最小的改进型LDA特征选择算法,采用迭代计算使Bayes分类错误率上界最小,能取得比原LDA更好的分类效果。对高维数据提出基于PCA预处理的“快速改进型LDA特征选择”减少求解迭代计算时间。针对Marill.T.提供的典型数据和MINIST手写体数字库的实验证实以上论点是正确的。

An Improved LDA Feature Selection Algorithm Based on Minimizing Classification Error Probability

Common feature selection method as LDA is, it may not obtain ideal classification results when the scatter matricesare different. In this paper, an Improved LDA feature selection based on minimizing classification error probability is presentedand a recursive algorithm to obtain an accurate solution of minimum upper bound of classification error probability is adopted.For high dimension data, a fast algorithm that pre-processes all samples by PCA is introduced to reduce the heavy computationburden of the recursive algorithm. The experiments with standard data and MINIST database show that Improved LDA issuperior to LDA.