Linear or non-linear multivariate calibration models? That is the question

Abstract

Concepts from data science, machine learning, deep learning and artificial neural networks are spreading in many disciplines. The general idea is to exploit the power of statistical tools to interpret complex and, in many cases, non-linear data. Specifically in analytical chemistry, many chemometrics tools are being developed. However, they tend to get more complex without necessarily improving the prediction ability, which conspires against parsimony. In this report, we show how non-linear analytical data sets can be solved with equal or better efficiency by easily interpretable modified linear models, based on the concept of local sample selection before model building. The latter activity is conducted by choosing a sub-set of samples located in the neighborhood of each unknown sample in the space spanned by the latent variables. Two experimental examples related to the use of near infrared spectroscopy for the analysis of target properties in food samples are examined. The comparison with seemingly more complex chemometric models reveals that local regression is able to achieve similar analytical performance, with considerably less computational burden.