Forecasting red wine quality: A comparative examination of machine learning approaches

Abstract

This research explores the forecast of red wine quality utilizing machine learning algorithms, with a particular emphasis on the impact of alcohol content, sulphates, total sulfur dioxide, and citric acid. The original dataset, comprised of Portuguese "Vinho Verde" red wine data from 2009, was bifurcated into binary classes to delineate low-quality (ratings 1-5) and high-quality (ratings 6-10) wines. A heatmap verified the potent correlation between the chosen variables and wine quality, paving the way for their inclusion in our analysis. Four machine learning techniques were employed: Logistic Regression, K-Nearest Neighbors (KNN), Decision Tree, and Naive Bayes. Each technique was trained and assessed through resulting metrics and graphical visualizations, with diverse proportions of data assigned for training and testing. Among these techniques, Logistic Regression achieved an accuracy score of 72.08%, while KNN slightly surpassed it with an accuracy rate of 74%. The Decision Tree technique rendered the peak accuracy of 74.7%, while Naive Bayes underperformed with a score of 60.2%. From a comparative viewpoint, the Decision Tree technique exhibited superior performance, positioning it as a viable instrument for future predictions of wine quality. The capacity to predict wine quality carries significant implications for wine production, marketing, customer satisfaction, and quality control. It enables the identification of factors contributing to high-quality wine, optimization of production processes, refinement of marketing strategies, enhancement of customer service, and potential early identification of substandard wines before reaching consumers, thereby safeguarding the brand reputation of wineries.