Profiling Analysis of Confiscated Methamphetamine using Non-Linear Classification Methods

Abstract

We have reported that classical methods such as the principal component analysis are insufficient for profiling illegally distributed drugs. Then we have used Livingstone-type hierarchical neural networks and the neural independent component analysis to classify confiscated methamphetamine. However, the two methods did not always show sufficient results for profiling such methamphetamine. In this study, other two methods, SOM and kernel principal component analysis, were applied in order to classify the confiscated methamphetamine using GC-MS data of impurities. To classify the data using SOM, not only winner neurons but also all the weights of connections were used like fingerprint matching. Thus, better classification results, which are consistent with the result from a synthetic investigation, were obtained. However, since SOM is a classification method, it is also desired to apply nonlinear component analyses. Kernel principal component analysis is one of such nonlinear analyses; it uses higher-dimensional feature space in order to find appropriate principal axes. This method gave the classification result that is easier to understand than SOM.