Abstract

This paper proposes an approach for feature selection aimed at classifying wines samples according to place of origin. The method relies on Kruskal-Wallis non-parametric test to remove non significant features, and Linear Discriminant Analysis to derive a feature importance index. The ranked features according that index are iteratively added and classification performance is assessed after each insertion. The number of selected features is chosen according the maximum accuracy in a repeated 10-fold cross-validation. Aiming at improving cate-gorization accuracy, different classification techniques are tested. When applied to a wine dataset comprised of 53 samples from four South America countries (Argentina, Brazil, Chile, and Uruguay) and 45 chemical elements concentrations determined by ICP-OES and ICP-MS, the proposed framework yielded average 99.9% accurate classifications in the testing set, and retained average 6.73 of the 45 original elements. Retained chemical ele-ments were then qualitatively assessed.