Automated, objective and optimized feature selection in chemometric modeling (cluster resolution)

Abstract

A novel metric, termed cluster resolution, which compares the separation of clusters of data points while simultaneously considering the shapes of the clusters and their relative orientations. This metric, in conjunction with an objective variable ranking metric, allows for the fully-automated determination of the optimal number of variables to be included in a chemometric model of a system. Cluster resolution is based upon considering the minimum distance between (or the extent of overlap of) confidence ellipses constructed around clusters of points representing different classes of objects. In one implementation, PCA was used to classify samples of gasoline based upon their octane rating. The entire GC-MS chromatogram of each sample comprising over 2×106 variables was considered. These variables were ranked by ANOVA and then a forward selection approach was used to find the optimal number of variables for inclusion. In this implementation, a model having 2761 variables was found to be optimal. This approach can be generally applied to feature selection for a variety of applications and represents a significant step towards the development of fully-automated, objective construction of chemometric models.

Dependencies are inferred from the claim text.