Clustering of organoleptic quality evaluations of red and white wines by physicochemical parameters using statistica software

Abstract

Organoleptic evaluations in a ten-point scale of wine experts and experimental physicochemical parameters of red (1599 samples) and white (4898 samples) wines of Portuguese manufacturers were analyzed using STATIS-TICA sofware. Methods of agglomerative and iterative ( k -means algorithm) clustering revealed the grouping of similar wine samples into three, four and six clusters depending on the Euclidean distance of association. The quantitative filling of clusters with samples of bad wines (grades 3 and 4), normal wines (grades 5 and 6) and good wines (grades 7, 8, 9) was established. Neural network (MLP) and discriminant analyzes (DA) were performed; algorithms of classification trees (CT), support vector machines (SVM), naive Bayesian classification (NBC) and nearest neighbors (kNN) were involved. The best performance was demonstrated by neural network models. Multilayer perceptorons classifiers were trained: for red wines - MLP 11-7-3, MLP 11-13-4, MLP 11-14-6; for white wines - MLP 11-9-3, MLP 11-5-4, MLP 11-9-6. The properties of wines, whose contribution to the separating power of classifiers is decisive, are revealed. The ranges of changes in physicochemical parameters in three clusters of red and white wines for bad, normal and good wines were given.