核分段逆回归集成线性判别分析用于质谱数据分类

针对高维小样本质谱数据在构造模型时易产生的过拟合现象、变量间的严重共线性、及结构与性质间的非线性关系,采用了核分段逆回归(KSIR)特征提取集成线性判别分析(LDA)新技术。首先以KSIR算法完成质谱数据的非线性特征提取,然后在由新特征矢量张成的低维空间构造样本类别的线性判别函数,负责各样本个体类别的判定。将KSIR-LDA方法应用于软饮料的质谱数据分类,结果表明:该方法不仅适应质谱数据与性质间的非线性关系,而且可以更少、解释能力更强的特征变量取得更高的分类精度,并能实现在低维特征空间对数据的解释及可视化。

Classification of Mass Spectrometric Data Based on Kernel Sliced Inverse Regression and Linear Discriminant Analysis

Extracting the most discriminatory features was important in mass spectrometry recognition tasks.In the case of a small number of mass spectrometric samples,the existed methods for extracting discriminatory features encountered various problems,such as the over-fitting,the multicollinearity within numerous predictor variables,nonlinear quantitative relationship between the structure and properties and etc.A novel classification method was constructed by combination of the kernel sliced inverse regression(KSIR) with linear discriminant analysis(LDA).The resulting discriminate model based on this proposed approach(KSIR-LDA) was divided into two steps: the first step was to perform the SIR algorithm in the reproducing kernel Hilbert space(RKHS) induced by applying the kernel trick to the original high dimensional mass spectrometric data for nonlinear dimension reduction and feature extraction,and the second step was that to build the LDA discriminate model in making use of the extracted feature variates.Finally,application to the soft drinks four-group classification problem was presented with a comparison to some other methods.The results show that it is an effective classification method in the structural risk minimization,non-linear characteristics,avoiding the over-fitting and strong generalization ability.At the same time,the KSIR-LDA discriminate model is more concise and can be used to observe the structure of the set of samples.