Bagging and boosting machine learning algorithms for modelling sensory perception from simple chemical variables: Wine mouthfeel as a case study

Aiming to predict sensory properties from chemical data, the application of bagging and boosting machine learning (ML) algorithms was comprehensively investigated and applied to modelling of red wine mouthfeel from simple chemical measurements. A panel of 15 Australian winemakers described the mouthfeel properties of a total of 30 commercial red wines from Australia and Spain using rate-all-that-apply sensory methodology. In parallel, linear sweep voltammetry signals and excitation-emission matrix (EEM) and absorbance data were acquired for the wines. Data were analysed following unsupervised statistical strategies including principal component analysis (PCA with varimax rotation) to simplify the interpretation of sensory variables, along with supervised regression models based on ML, namely random forest (RF) and extreme gradient boosting (XGBoost). PCA results showed that four independent and uncorrelated sensory dimensions mainly related to perceptions of ‘drying’, ‘full body’, ‘velvety’, and ‘gummy’ differentiated among the wines. The RF and XGBoost algorithms yielded superior validated regression models compared to classical PLS modelling. The ML algorithms exhibited strong predictive performance on test data, with an average value exceeding 80 % accuracy for any of the three sets of chemical variables employed. Although XGBoost provided slightly better models, the low computational effort required by RF is advantageous. Important variables included in the models are discussed along with the importance of controlling overfitting. Overall, absorbance, voltammetric or EEM signals coupled with RF or XGBoost algorithms are presented as cheap, easy-to-use, and rapid approaches to predicting sensory properties from chemical signals in complex matrices such as wine. • Modelling of wine mouthfeel from simple chemical measurements was achieved. • Bagging and boosting machine learning algorithms were comprehensively investigated. • Random forest and XGBoost algorithms yielded validated models with >80 % accuracy. • XGBoost provided slightly better prediction models than random forest. • The lower computational effort required by random forest is advantageous.