class modeling using multivariate curve resolution methods

Abstract: classification or supervised pattern recognition is a name given to a set of numerical techniques developed to solve the class membership problem. an empirical relationship or classification rule is developed from a set of samples for which the property of interest and the measurements are known. the classification rule is then used to predict this property in samples that are not part of the original training set. multivariate curve resolution (mcr) strategies are powerful tools allowing the description, species identification and system understanding, of totally or partly unknown chemical processes and reactions where species cannot be easily isolated and where unknown intermediate species may be present. these methods are also powerful techniques for quantification of complex mixtures. recently mcr-als was used enable discrimination between the constituents of two benchmark and two high-dimensional data sets and was compared with other commonly used techniques [1]. in this work, mcr methods, which already enjoy broad use in various fields, were applied for supervised learning. mcr is a bilinear decomposition approach that supports inclusion of additional system information within the form of numerical constraints. in order to develop an mcr algorithm for supervised learning, the target class of the samples use mcr subspace for two distance measures. they are the score distance (sd) and the orthogonal distance (od) [2]. mcr method was tested to enable classification between the constituents of data sets. several real experimental complex data have been used for evaluating the power of mcr methods for classification. for example, spectra of eighty corn samples (700 wavelengths) with reference value percent moisture, oil, protein, and starch were measured from 1100 to 2498 nm at 2-nm intervals on three nir instruments designated as m5, mp5, and mp6 [3]. the results were compared with the output of the application of different data classification methods like dd-simca and have reliable result. the proposed mcr approach, in comparison with other commonly used supervised techniques, has the advantages of improved accuracy from the inclusion of additional system information in the form of numerical constraints, mcr profiles and subspaces are excellent compressed information that represents adequately the initial information of the data set and, in difference with analogous representations coming from pca, the information enclosed in the mcr profiles is compound-specific and chemically meaningful, and the ability to resolve pure components signal weights.