Cross-Validation Strategies: Which One to Use and When

Decision Tree

1. Time series data? → TimeSeriesSplit (never shuffle time!)
2. Groups that must stay together? → GroupKFold
3. Imbalanced classes? → StratifiedKFold
4. Default tabular → StratifiedKFold with 5 folds

TimeSeriesSplit Example

from sklearn.model_selection import TimeSeriesSplit
tscv = TimeSeriesSplit(n_splits=5)
for train_idx, val_idx in tscv.split(X):
    X_train, X_val = X.iloc[train_idx], X.iloc[val_idx]

GroupKFold: Preventing Leakage

If you have user IDs, patient IDs, or store IDs — always use GroupKFold so the same entity never appears in both train and validation.