What is cross-validation and which KFold variant should you use?

It scored 94% once. Your accuracy was one lucky split, and cross-validation is how you find out what the model actually does. One page. The print-ready A4 PDF is at the bottom.

Why

• One split: one lucky draw.
• k folds: k honest scores.
• Report mean ± std, not the best fold.

The honest five lines

from sklearn.pipeline import make_pipeline
pipe = make_pipeline(StandardScaler(),
                     LogisticRegression())
scores = cross_val_score(pipe, X, y,
           cv=StratifiedKFold(5, shuffle=True))
print(scores.mean(), scores.std())

Scaling lives inside the pipeline, so every fold learns preprocessing from its own train side only. Scale before the split instead, and the test rows have already whispered their mean to the model.

Variants

• StratifiedKFold: keeps the class ratio in every fold.
• GroupKFold: the same user never lands on both sides.
• TimeSeriesSplit: for anything dated.

Leakage

• Scale inside the folds, never before the split.
• Impute inside the folds too, same rule.
• Pipeline() does it right for free.

Numbers

• k = 5: the default.
• k = 10: small datasets.
• Leave-one-out: tiny data only.

The trap: the wrong splitter

The GroupKFold one is the interview favorite: leaking a user across the split inflates scores and nobody notices until production.