Using Machine Learning to Predict Wine Quality and Prices: A Demonstrative Case Using a Large Tabular Database

Abstract

The wine industry is currently experiencing a worldwide growing interest, playing a substantial role in the economy of numerous countries. In this paper, we address the prediction of both wine quality and price, which is a valuable tool for several wine stakeholders, including producers, sellers and consumers. In particular, we explore a large worldwide tabular database that was retrieved from the Vivino platform and that includes around 167 thousand price records and more than 4 million quality examples. Using these data and a 10-fold cross-validation evaluation scheme, a large set of ML computational experiments was executed, considering three Feature Selection (FS) scenarios, up to six ML algorithms and five modeling approaches: a single (global) model, two divide-and-conquer ML modeling strategies (country and clustering based) and two hyperparameter tuning (for the single model and clustering filtering approaches). Overall, high quality predictive results were achieved. In particular, the price prediction best results were obtained by the clustering tuned FS1 scenario, while the best quality predictions were achieved by the clustering filtering FS3 setup.