A sparse fused group lasso regression model for fourier-transform infrared spectroscopic data with application to purity prediction in olive oil blends

Abstract

The percentage of olive oil present in an oil blend is of interest in the quality control of oils sold to consumers. One way in which this can be measured is using infrared spectroscopy. The analysis of the resulting data is challenging due to the high-dimension of the data and multicollinearity caused by issues such as the similarities between the chemical constituents in vegetable oils. This paper develops a sparse fused group lasso model for simultaneous feature selection and model fitting on Fourier-transform infrared spectroscopic data, and applies it to the task of percentage purity prediction in oil blends. The arising optimization problem is solved via the alternating direction method of multipliers algorithm. The sparse fused group lasso method is seen to improve on the interpretability of the resultant models, while providing comparable predication performance. Most importantly, it provides a flexible model that can capture group structure and smoothness in the coefficient structure.