Machine Learning — Regression Modeling
This predictive analytics project uses physicochemical measurements to estimate red wine quality. Three regression approaches—Linear Regression, Support Vector Regressor, and Gradient Boosting Regressor—are compared, with hyperparameter tuning applied to the more advanced models.
The UCI Wine Quality red wine dataset contains 1,599 records, 11 physicochemical input features, and the quality score as the regression target. After removing 240 duplicate rows, 1,359 unique samples remain for modeling.
Data Understanding → Duplicate Removal → Feature/Target Split → 80/20 Train-Test Split → StandardScaler → Linear Regression, SVR & Gradient Boosting → GridSearchCV → MAE, RMSE & R² Evaluation
Wine quality scores range from 3 to 8, with most observations concentrated at quality 5 and 6.
Alcohol shows a positive relationship with quality, while volatile acidity shows a notable negative relationship.
After cleaning, 1,087 samples are used for training and 272 samples for testing.
Gradient Boosting Regressor provides the strongest final test performance among the evaluated models.
R² Score: 0.4560 — the model explains about 45.60% of the variance in red wine quality.
MAE: 0.4806 — predictions differ from the actual quality score by about 0.48 points on average.
RMSE: 0.6208 — this metric places greater weight on larger prediction errors.
Selected model: Gradient Boosting Regressor with learning_rate 0.05, max_depth 3, and n_estimators 100 after GridSearchCV.
