From the correlation matrix, we identified strong correlations between fixed acidity, citric acid, and pH. There is also a moderate correlation between alcohol and density.

Since fixed acidity and pH are fairly linearly dependent, we risk the presence of multicollinearity, which would hamper our ability to interpret the impacts of specific variables on the response in an ordinary least square regression setting. Therefore, by performing shrinkage methods, we may be able to improve overall test MSE.

We built a tree using recursive binary splitting to predict alcohol content. The tree indicates that the most important predictor in predicting alcohol content is density, followed by citric acid, fixed acidity, and residual sugar.

We performed random forests on our regression objective, and found the results to be consistent with our trees. Density, residual sugar, and fixed acidity contribute the most information to the MSE, making them the strongest predictors.

Prior to our classification analysis on the quality of wine, we performed EDA to identify trends in the distribution of variables by quality of wine class. Interestingly, quality of wine rated “good” (1) tends to have a greater distribution of alcohol content.

We created a pruned tree with the quality of wine outcome. The tree considers only two variables: alcohol and volatile acidity. If the alcoholic content is at or above 10.25% then the wine is going to be predicted as above average quality, whereas if the alcoholic content is below this level, we need to look at volatile acidity.