Variable versus Interval selection in Partial Least Squares

Among multivariate techniques, partial least squares (PLS) is a wide implemented technique. Genetic algorithm (GA) and synergy interval-PLS (Si-PLS) were compared as two methods for improvement of PLS modeling. Using the visualization ability of self-organizing map-partial least squares (SOMPLS), we introduced a reference for spectrum regions of high importance. Corn and Metabolite data were used. On the whole, SiPLS resulted in better errors of cross validation and prediction but model complexity in GA-PLS was simpler. In the case of Metabolite data, containing components of relatively low overlapping spectra, the prediction of some vectors of Y matrix was only slightly or never better than PLS without preprocessing