| Popis: |
Classifying samples into known categories is a common problem in analytical chemistry and other fields. For example, with spectroscopic data, samples are measured and the corresponding spectra are compared with existing spectral data sets of known classification (library sets) to determine the appropriate classification. Presented in this paper is a study of the simple and well known data analysis processes target factor analysis (TFA) and net analyte signal (NAS). Although TFA and NAS were originally derived for different purposes in analytical chemistry, they are based on the same calculation. The library set with the smallest TFA residual (smallest NAS) for a test sample spectrum can be used for classification purposes. Alternatively and equivalently, this paper uses the smallest angle (poorest selectivity in NAS terminology) between a new sample spectrum vector and the space spanned by each library loading vector basis set. The angle classification is compared with classifications by the Mahalanobis distance and k-nearest neighbors. The measures are evaluated with three spectroscopic data sets consisting of benchmark identification of plastic type (Raman) and gasoil plant source (ultraviolet) and a new extra virgin olive oil adulterant identification (fluorescence) data set. A fourth data set is the benchmark archeological data set. The Mahalanobis distance and k-nearest neighbors generally classify with 2%–40% and 0%–20% decreases in correct classifications, respectively, compared with TFA (NAS). Results from this study indicate that the simple TFA and NAS processes are useful underutilized classification and library searching tools. Copyright © 2012 John Wiley & Sons, Ltd. |
Plný text