Wine Quality Prediction with Ensemble Trees: A Unified, Leak-Free Comparative Study

Accurate and repeatable wine quality assessment remains indispensable, but to a large extent, it relies on subjective tasting teams. This study compares five advanced ensemble models—Random Forest, Gradient Boosting, XGBoost, LightGBM, and CatBoost—using the standard Vinho Verde red wine and white wine datasets. Adopt leak-free pipes stratified 80:20 training-test segmentation, five stratified GroupKFolds of the training data, each StandardScaler scaling, SMOTE-Tomek resampling, inverse-frequency cost weighting, Optuna-driven hyperparameter search (120–200 trials per model), and two-stage feature selection modification. Performance is reported on untouched test set, with weighted F1 as the primary metric. Gradient Boosting provides the strongest test scores weighted F1 = 0.693 ± 0.028 (red) and 0.664 ± 0.016 (white)—followed closely by Random Forest and XGBoost within 3% points. The input space was compressed to the five most important variables of each model, reducing the dimension by 55%. At the same time, the weight F1 of red wine was reduced by 2.6% points, and the wright F1 of white wine was reduced by 3.0% points. This confirmed that alcohol, volatile acidity, sulphates, free SO₂, and chlorides occupied the majority of the predictive signal. Runtime analysis on the AMD EPYC 9K84/NVIDIA H20 node revealed a distinct efficiency gradient: Gradient Boosting averages 12 hours per five studies, XGBoost and LightGBM complete within 2-3 hours, CatBoost within approximately 1hour, and Random Forest within 50 minutes. These findings indicate that Random Forest is the most cost-effective production model, XGBoost and LightGBM as GPU-efficient alternatives, and Gradient Boosting as the upper limit of accuracy for offline benchmark tests. A fully documented pipeline and comprehensive metric set establish a repeatable baseline for future work on imbalanced, and multi-grade wine quality prediction.