Near-perfect evaluation scores can hide overfitting, especially on a small dataset. Cross-validation checks whether performance holds across different partitions rather than relying on one evaluation. section 4.6.5
In five-fold cross-validation, split the data into five folds. Train on four and test on the remaining fold, repeating until each fold has served as the test set. Compare the fold scores as well as their average; substantial variation can signal instability, though it does not by itself prove overfitting. section 4.6.5
Fold 1 Accuracy: 0.9499 Fold 2 Accuracy: 0.9680 Fold 3 Accuracy: 0.9566 Fold 4 Accuracy: 0.9339 Fold 5 Accuracy: 0.9633 Average CV Accuracy: 0.9543 (+/- 0.0119)
These five accuracy scores show why both the average and variation across folds matter. section 4.6.5
Cross-validation complements threshold selection: an F1 score can help choose a balance between precision and recall, but a strong threshold score alone does not establish that the model generalizes. section 4.6.5














