PLS and dimension reduction for classification (Q964596)

Different techniques of dimension reduction in classification problems are compared. The main attention is paid to the partial least squares (PLS) and its modification, oriented PLS (OPLS). These techniques are compared to the principal components analysis (PCA) used as pre-processing before linear discriminant analysis (LDA). A ridge-like technique is proposed for OPLS dimension reduction to handle multicollinear data. It is shown that the OPLS-LDA technique leads to classifiers with lower misclassification rates then PLS-LDA and PCA-LDA ones. A real data example is presented.