Adaptive Regression via Subspace Elimination

Abstract

The two primary goals in the creation of any multivariate calibration model are effectiveness and longevity. A model must predict accurately and be able to do so over an extended period of time. The primary reason models fail when applied to future samples is the presence of uncalibrated interferents. Uncalibrated interferents represent any change in the future samples not described by the original calibration set. Most often uncalibrated interferents result from the addition of chemical constituents, changes to analytical instrumentation, and changes to environmental conditions. Presented within this chapter is a novel algorithm, Adaptive Regression via Subspace Elimination, for handling uncalibrated chemical interferents via an adaptive variable selection approach. Results are presented for synthetic Near Infrared (NIR) and Infrared (IR) data.