Work / School / 2026
Wine Origin Classification
University of Florida · with Deep Patel
Identified which of three cultivars a wine came from using its chemistry. Cut 13 chemical predictors down to 5 and still hit 95.3% cross-validated accuracy.
The question
Can a handful of lab measurements reliably tell you where a wine came from, and which measurements actually matter?
Approach
Multinomial logistic regression across three classes. Used stepwise AIC and LASSO to drop redundant predictors, then validated with 10-fold cross-validation instead of a single lucky split.
What we found
Five of the thirteen predictors carried the signal. The reduced model scored 95.3% CV accuracy with a log-loss of 0.119, so it was both right and confident.
Why it matters
Fewer inputs means a cheaper, simpler test. Spending effort only on the variables that matter is the same instinct that cuts costs in a real business process.