random subspace ensemble-multivariate curve resolution-alternating least squares (rse-mcr-als) as a new technique for first-order calibration

Abstract:one of the most important aspects of any analytical method is quantitation especially using first-order multivariate calibration techniques which allows the simultaneous determination of multiple components in the mixed sample. multivariate curve resolution-alternating least squares with correlation constraint (mcr-als-cc) has played an important role in this area, provided both the qualitative and quantitative determination and prediction of components in complex mixtures [1-3]. due to this importance, in this study a novel strategy called random subspace ensemble multivariate curve resolution-alternating least square (rse-mcr-als) was proposed in order to improve the proficiency of the conventional mcr-als with correlation constraint for calibration purposes. ensemble regression with purpose of combining several models to make a prediction has been the object of many researches in recent years, considerably improved both the robustness and prediction accuracy of the models [4]. in this regard, three different near infrared (nir) data sets (including meat (fat), milk (water adulteration), and gasoline (octane numbers)) were chosen to check the feasibility of the proposed method. two parameters/factors namely the number of base learners, and the number of subspaces, which are the two vital parameters of rse-mcr-als were optimized using central composite design (ccd) and simplex optimization algorithm. the weighted average of root mean square error of prediction (rmsep) for mcr-als-cc was considered as response of the experimental design. due to the results obtained by analysis of variance (anova) of three nir data sets, the number of subspaces was not a significant parameter, but the number of base learners was significant. finally, by considering the optimum number of base learners and subspaces (contains the segments of features in order of their arrival) and applying them on mcr-als-cc model and weighted averaging over the rmsep values of base learners, considerably better prediction results based on rmsep were obtained (0.07, 0.20, and 0.02 for prediction of fat, water adulterant, and octane number respectively) in comparison to conventional mcr-als-cc model (with rmsep values of 7.74, 9.14, and 9.14 respectively). moreover, by considering the same procedure on plsr as the most used first order multivariate calibration technique in analytical chemistry, better prediction results were obtained for rse-plsr (0.30, 0.56, and 0.05 respectively). however, these results were still poor in comparison to rse-mcr-als.