Combined Synthetic Minority Oversampling Technique and Deep Neural Network for Red Wine Quality Prediction

Abstract

Red wine is an alcoholic drink made from the fermentation of grapes. With the continuous increase in the market of red wine, quality assessment of red wine is vital to meet the required quality. Prediction of red wine quality holds significant reasons such as consumer satisfaction, building a strong reputation for wine producers, identifying high-quality wine batches, and determining problems during wine-making process. Formulating predictive models for wine quality classification are already explored in past researches but, improvements of techniques and performance for these models are still in front of wine production research. This paper discusses the utilization of Deep Neural Network (DNN) algorithm combined with Synthetic Minority Oversampling Technique (SMOTE) in predicting red wine quality into ‘low’, ‘moderate’ and ‘high’ quality. The red wine dataset is obtained from UCI machine learning repository. The dataset records physiochemical parameters of red wines and the corresponding quality level. Results have shown that the formulated predictive model via DNN integrated with SMOTE for predicting wine quality yielded a considerably very high performance with an accuracy = 97.81 %, kappa coefficient = 0.967, and f - score = 0.976. Future research direction may include (1) feature importance analysis of wines' physicochemical parameters and their interactions; (2) sensitivity analysis of input parameters (physiochemical properties) with respect to the output categories (wine quality); and (3) exploration of other machine learning algorithms and other techniques to improve prediction performance.