Random forests: A machine learning methodology to highlight the volatile organic compounds involved in olfactory perception

Abstract

The purpose of this paper is to discuss the application of the Random Forest methodology to sensory analysis. A methodological point of view is mainly adopted to describe as simply as possible the construction of binary decision trees and, more precisely, Classification and Regression Trees (CART), as well as the generation of an ensemble of trees or, in other words, a Random Forest. The interest of the permutation accuracy criterion, as a measure of variable importance, is specifically emphasized as a way of identifying the most predictive variables and selecting a subset of these variables for parsimonious and efficient predictive models. A two-step procedure is proposed for choosing this subset of variables. The principle of the method is illustrated in a case study in which the aim was to better understand and predict the olfactory characteristics of red wines made of the Cabernet Franc grape variety, from their Volatile Organic Compound (VOC) content. For two main olfactory attributes, the bell pepper odor and the leather odor, it was possible to list the most important compounds and to highlight a very small number of compounds useful for estimating each of the olfactory attributes considered. For the latter, it was also observed that Random Forest models had a better predictive ability than Partial Least Squares (PLS) Regression models.