Beyond Accuracy: A Leakage-Aware and Explainable Machine-Learning Framework for Binary Red-Wine Quality Prediction

Abstract

Predicting wine quality from physicochemical measurements is a useful benchmark for studying imbalanced classification, model comparison, and interpretable machine learning. This paper audits and extends the public Prediction-of-Wine-Quality project, which converts the ordinal quality score of red wine into a binary label: good quality (score ≥ 7) and not-good quality (score < 7). The supplied red-wine dataset contains 1,599 records and 11 physicochemical predictors. The positive class is uncommon (217 records; 13.57%), making accuracy alone an inadequate model-selection criterion. A source-code audit found that the original notebook uses a single non-stratified 70/30 split and reports a hard-coded summary table; therefore, the table is treated as a reported artifact rather than as independently verified output. Re-execution of the notebook’s model configurations produced a Random Forest and XGBoost accuracy of 89.38%, but their positive-class recall and balanced accuracy reveal limitations hidden by accuracy. Extended experiments used a stratified hold-out set, training-only randomized hyperparameter search with five-fold stratified cross-validation, imbalance-aware metrics, and model-agnostic and SHAP-based explanations. XGBoost achieved the highest cross-validation balanced accuracy (0.8008), while Extra Trees achieved the highest held-out balanced accuracy (0.8291). The selected XGBoost model attained 81.54% recall and 0.9197 ROC-AUC on the held-out test set, at the cost of 41.09% precision. Alcohol and sulphates were the most prominent global predictors in the explainability analyses. A duplicate-group sensitivity split reduced XGBoost balanced accuracy to 0.7518, underscoring the need to account for repeated feature vectors. These results constitute a transparent empirical reassessment, not evidence of causal effects or external clinical/industrial validity.