SMOTE

SMOTE rebalances the training classes without duplicating rows, but the synthetic points are functions of existing positives and carry no new information. It shifts predicted probabilities towards a more balanced prior, can blur class boundaries, and leaks if applied before the train/validation split. Apply it inside training folds only, evaluate on real rows at the real prevalence, and correct probabilities before thresholding.