Finding the most important sensory descriptors to differentiate some Vitis vinifera L. South American wines using support vector machines

• Published in: European food research & technology Vol. 245; no. 6; pp. 1207 - 1228
• Main Authors: Costa, Nattane Luíza, Llobodanin, Laura Andrea García, Castro, Inar Alves, Barbosa, Rommel
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01-06-2019; Springer Nature B.V
• Subjects: Agriculture; Analytical Chemistry; Biotechnology; Chemistry; Chemistry and Materials Science; Feature selection; Food Science; Forestry; Learning algorithms; Machine learning; Organic chemistry; OriginalPaper; Sensory properties; Support vector machines; Vitamin E; Wine; Wines; Support vector machines; South American wines; Feature selection; Sensory descriptors; Machine learning; Classification
• ISSN: 1438-2377; 1438-2385
• DOI: 10.1007/s00217-019-03245-9

The geographical recognition of wines has been extensively attempted based on chemical parameters. However, few studies have used wine sensory properties to characterize wines according to their geographical origin. This paper presents a machine learning study to classify and to find the most important sensory descriptors of Cabernet Sauvignon, Syrah, Tannat, and Merlot wines from Argentina, Brazil, Chile and Uruguay. Four feature selection methods (
F
score, relief,
χ
2
, and random forest importance) were used to generate the order of importance of the sensory descriptors. The feature subsets were generated based on the feature selection ranking order to use as input features for the support vector machines classifier. Very good results with 85–100% accuracy were achieved, and the results showed that few sensory descriptors discriminate the origin of wines better than when using all the descriptors and that there is a specific subset of most important features to each wine variety. As far as we know, this is the first study to analyze South American wines based solely on sensory descriptors and support vector machines along with feature selection methods.