교차 검증 제대로 하기

A single train-test split can flatter or punish your model by luck. Cross-validation averages many splits for a score you can trust.

K-fold cross-validation slices your data into k equal parts, then trains on k-1 and tests on the one held out, rotating each time.

Across the k rounds, every example is used for testing exactly once. You get k scores instead of one fragile number.

Report the mean of the k scores as your estimate, and the standard deviation to show how stable that performance really is.

scikit-learn handles the whole loop with one helper that returns a score per fold.

For classification use stratified folds so each fold keeps the same class ratio, which matters most on imbalanced data.

Leakage means test information sneaks into training, inflating your score. Fit scalers and vectorizers inside each fold, never before.

Wrap preprocessing and the model in a Pipeline so every fold refits the transforms only on its own training portion. Clean and safe.

Five or ten folds are common. More folds give a steadier estimate but cost more compute, since the model retrains for every fold.

When picking hyperparameters, search them inside cross-validation. GridSearchCV ties tuning and validation together so results stay honest.

Even with cross-validation, keep one untouched test set for a final check. It confirms the estimate on data the search never saw.

Why does cross-validation beat a single split?

Cross-validation rotates folds for a reliable score. Use stratified folds, guard against leakage with pipelines, and keep a final test set. ✅