Using contrasts as data pretreatment method in pattern recognition of multivariate data

A contrast method originally proposed by Spiegelman [C.H. Spiegelman, Calibration: a look at the mix of theory, methods and experimental data, presented at Compana ʹ95, Wuerzburg, Germany.] is modified to pretreat multivariate data for classification. Three NIR data sets and one pollution data set are used as examples. Our results show that the contrast method greatly improves the ratios of between- to within-class variance. It is more powerful than offset correction, SNV, first- and second-derivative methods in the cases studied. This conclusion does not depend on the type of classifier used. Regularised discriminant analysis (RDA) and partial least squares (PLS2) with univariate feature selection based on Fisherʹs ratio were applied here. There is a risk that chance correlations occur after the contrast pretreatment. The chance correlation decreases after first eliminating un-informative variables using the modified Uninformative Variable Elimination (UVE)-PLS method.