Answers are grounded in this paper and cite their sources.
Overall accuracy is the default headline metric for classifiers, yet on imbalanced data it can reward models that ignore the cases that matter. We study this problem on the red Vinho Verde wine dataset (n = 1,599), grouping expert grades into three classes: poor (3.9% of wines), average (82.5%) and premium (13.6%). Six classifier families were compared. Ranked by accuracy, a random forest (0.849) and a neural network (0.862) appear far ahead of logistic regression (0.612); yet the class-weighted random forest detected only 2 of 10 poor wines in the test set, whereas logistic regression detected 8 of 10. On this test set, a trivial classifier that always predicts “average” would score 0.822 accuracy and exceed the final model, while having a balanced accuracy of only 0.333 against the final model’s 0.710. Tuning logistic regression with GridSearchCV on balanced accuracy selected the default balanced class weights; the price of high minority recall was low precision (0.127 for the poor class) and misclassification of 42% of average wines. We argue that accuracy should never be the sole selection criterion on imbalanced data, and we document several evaluation weaknesses of our own pipeline as a transparent example. Code is available at github.com/YashDeepKunwar/Wine_prediction_Model.
No takes yet. Share an insight, caveat, or question.
Yash kunwar (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: