Understanding the Wine Judges and Evaluating the Consistency Through White-Box Classification Algorithms

Abstract

Wine is a broad field of study and is more and more popular today. However, limited amounts of data science and data mining research are applied on this topic to benefit wine producers, distributors, and consumers. According to the American Association of Wine Economics, “Who is a reliable wine judge?” and “Are wine judges consistent?” are typical questions that beg for formal statistical answers.

This paper proposes to use the white box classification algorithms to understand the wine judges and evaluate the consistency while they score a wine as 90+ or 90−. Three white box classification algorithms, Naïve Bayes, Decision Tree, and K-nearest neighbors are applied to wine sensory data derived from professional wine reviews. Each algorithm is able to tell how the judges make their decision. The extracted information is also useful to wine producers, distributors, and consumers. The data set includes 1000 wines with 500 scored as 90+ points (positive class) and 500 scored as 90− points (negative class). 5-fold cross validation is used to validate the performance of classification algorithms. The higher prediction accuracy indicates the higher consistency of the wine judge. The best white box classification algorithm prediction accuracy we produced is as high as 85.7 % from a modified version of Naïve Bayes algorithm.