一种鲁棒的概率主成分分析方法

A Robust Probability Principle Component Analysis Method

针对传统主成分对实际样本的奇点不敏感的缺陷,提出了一种鲁棒概率主成分分析(RPP-CA)方法.首先引入连续的决策变量构造新能量函数,将事先给定的硬门限改为自适应确定的软门限,门限值由样本自动确定,再计算概率主成分进行特征提取.与线性主成分分析(LPCA)和概率主成分分析(PPCA)方法相比,RPPCA方法更为实用,有效地减小了奇点的影响,显示出比PPCA更强的稳健性,也扩大了实用范围.实验结果表明,RPPCA方法的分类准确率比LPCA方法平均提高了3.2%,比PPCA方法平均提高了0.7%.

In order to overcome the drawback that traditional principal component analysis fails to the outliers existing in the realistic data, a robust probability principal component analysis （RPPCA） method is proposed. A continuous decision variable is introduced into the energy function, and the preset hard threshold is replaced by a soft adaptive threshold which is automatically determined by the data. The algorithm is then embedded in the procedure of PPCA＇s principal component feature extraction. Compared with PCA and probability principal component analysis （PPCA）, the proposed RPPCA can resist outlier well, is more robust than PPCA, and enlarges the real application area. The simulation results show that the algorithm improves 3.2G classification accuracy to LPCA, and 0.7G to PPCA on average.