93% accuracy is easy to brag about — class imbalance is harder

The notebook title says 93% accuracy. The interesting part is how we got there without lying to ourselves about the data. UCI red wine quality is tiny, skewed, and full of outliers — a perfect teaching set for messy classification.

Quality scores from 3 to 8 are not evenly filled. Most bottles sit in the middle. Collapsing into bad / average / good made the task clearer, but it also exposed the imbalance: after IQR cleanup I still had 826 average, 116 good, and only 32 bad.

SMOTE rebalanced to 826 per class before the train–test split. That helped models see the minority class, but it also means accuracy is measured on a synthetic-aware pipeline. I still report the notebook numbers honestly — Random Forest at 92.74% on 496 test samples — and I treat them as evidence of the workflow, not a claim about production wine scoring.

On this feature space, ensembles handled mixed physicochemical signals better than a margin-based SVM (64.52%). Decision Tree and KNN landed in between. The classification reports mattered more than the headline accuracy: recall on the rare “bad” band was where models showed their real character.

Always plot the label distribution before celebrating a metric. Clean outliers deliberately. And keep the notebook outputs in the repo — future-you will thank past-you for the receipts.