Handling imbalanced data

Imbalance isn't a modeling problem until metrics say it is: pick a metric that reflects business cost - PR-AUC, F-beta with beta tied to FN:FP cost ratio, or top-k recall. Training-side options: class weights or focal loss; SMOTE/ADASYN on features only inside CV folds to prevent leakage; balanced bagging. Inference-side: threshold calibration on a held-out set with the true operating point. Cautions: SMOTE on time-ordered data leaks future into past; up-weighting rare classes without re-tuning learning rate can hurt; for ranking problems the right fix is often a calibration layer, not retraining. Pick the cheapest fix tied to the decision boundary.