A Study of Modeling Techniques for Prediction of Wine Quality

Abstract

The evaluation of wine quality is multifactorial and involves scoring characteristics that relate to a wine’s appearance, aroma and taste. These scores are, at least in part, subjective to individual expertise, experience and interpretation. Given this constraint, it is not surprising that the development of an objective, scientifically-based system for quantitatively scoring wine has remained elusive. In this work, multiple classification-based systems for predicting wine quality based on a set of measured physicochemical properties in the final product are evaluated. Several data pre-processing techniques are employed including, z-score normalization, outlier reduction, and principal component analysis (PCA). The resulting transformed data is then applied to various machine learning models (i.e., Multilayer Neural Networks (MNN), Random Forest (RF), basic Decision Tree (DT) and Support Vector Machines (SVM)). Standard metrics (i.e., confusion matrix, classification accuracies, precision, recall, F1-score), are compared for each model. A comparison of the results obtained in this study to contemporary works using similar metrics shows that two of the classifiers used in the current study (i.e., MNN with Adam optimizer and Random Forest) performed with higher classification accuracies and precision (70%/73 and 72%/73, respectively).