Research And Application Of Partial Least Square For Discrimination

The partial least square method is a widely used method for dimensionality reduction,classification and prediction of high-dimensional data,but in practice,this method usually produces linear combination of original variables that are difficult to explain.In this paper,we put discrimination into the framework of regression.In this paper,a sparse partial least square method with penalty term is proposed.Combined with singular value decomposition of matrix,an approximate method and algorithm of rank one matrix are proposed.The purpose of this method is to combine the variable selection of high-dimensional data and the linear combination of variables into one step at the same time.Find out the useful information contained in the original features as much as possible to reduce the dimension of the data.We simulated and generated high-dimensional binary classification data and multivariate classification data,and applied them to the data of esophageal squamous cell carcinoma that has been published,compared with the ordinary partial least square method.The results show that our method performs well in variable selection,and has better sensitivity and specificity.In high dimensional binary classification and multi category classification,it provides high prediction accuracy and stability,and has higher calculation efficiency.