Feature selection versus feature compression in the building of calibration models from FTIR-spectrophotometry datasets

Abstract

Undoubtedly, FTIR-spectrophotometry has become a standard in chemical industry for monitoring, on-the-fly, the different concentrations of reagents and by-products. However, representing chemical samples by FTIR spectra, which spectra are characterized by hundreds if not thousands of variables, conveys their own set of particular challenges because they necessitate to be analyzed in a high-dimensional feature space, where many of these features are likely to be highly correlated and many others surely affected by noise. Therefore, identifying a subset of features that preserves the classifier/regressor performance seems imperative prior any attempt to build an appropriate pattern recognition method. In this context, we investigate the benefit of utilizing two different dimensionality reduction methods, namely the minimum Redundancy-Maximum Relevance (mRMR) feature selection scheme and a new self-organized map (SOM) based feature compression, coupled to regression methods to quantitatively analyze two-component liquid samples utilizing FTIR spectrophotometry. Since these methods give us the possibility of selecting a small subset of relevant features from FTIR spectra preserving the statistical characteristics of the target variable being analyzed, we claim that expressing the FTIR spectra by these dimensionality-reduced set of features may be beneficial. We demonstrate the utility of these novel feature selection schemes in quantifying the distinct analytes within their binary mixtures utilizing a FTIR-spectrophotometer.