ARSE: Adaptive Regression via Subspace Elimination, a novel algorithm for eliminating the contribution of uncalibrated interferents

Abstract

Chemometrics represents an extremely effective data analysis tool. Through its judicious application, information that would otherwise be obscured within a data set can be discovered. This dissertation investigates the development of novel chemometric algorithms for accommodating the presence of uncalibrated spectral components. It will also discuss building models for both the classification of edible oils as well as predicting the peroxide value of those edible oils. ☐ Adaptive Regression via Subspace Elimination, ARSE, is demonstrated to be able to effectively handle the presences of uncalibrated chemical components within the prediction set. This is demonstrated first with a model system consisting of Gaussian “model spectra”. It is then expand to two different artificial data sets based upon actual pure component spectra. Across all the data sets a maximum of 4.2x improvement in prediction compared to just PLS is observed. ☐ Also shown within this dissertation is the ability to build models to accurately predict the type of edible oil as well as the peroxide value. The classification models demonstrate an overall 92.75% accuracy. The error for predicting peroxide value depends on both the type of spectroscopy used as well as the composition of the data set. The prediction error varies from 3.60 to 8.72 on the same data set for Near IR and Raman spectroscopy, respectively.