Wine Quality Grading from Physicochemical Features: Comparing SVM and Random Forest with PCA/t-SNE Visualization

Abstract

This study investigates multi-class prediction of wine quality scores (3–8) using physicochemical measurements from a Kaggle dataset. After removing the identifier field and standardizing features, we examine the data structure with correlation analysis and low-dimensional embeddings. PCA and t-SNE visualizations reveal substantial overlap among adjacent quality levels, implying gradual transitions rather than sharply separable classes. We then benchmark an RBF-kernel Support Vector Machine (SVM) and a Random Forest (RF) using an 80/20 stratified train/test split and 5-fold stratified cross-validation, tuning hyperparameters using Macro-F1 to account for class imbalance. Cross-validation slightly favors RF (Macro-F1 0.3271) over SVM (0.3078). On the held-out test set, RF achieves higher overall performance (Accuracy 0.7118, Macro-F1 0.3442, Weighted-F1 0.6897) than SVM (Accuracy 0.6332, Macro-F1 0.3072, Weighted-F1 0.6157). However, the relatively low Macro-F1 indicates uneven per-class performance, with limited effectiveness at rare quality levels, as reflected in confusion patterns that concentrate between neighboring grades (e.g., 5↔6 and 6↔7). RF feature importance highlights alcohol, sulphates, and volatile acidity as key predictors, offering interpretable evidence for the dominant decision cues.