Handling severe class imbalance in production

Layered approach: (1) data-level: resampling (oversample minority via SMOTE/ADASYN, undersample majority), class-balanced mini-batches; (2) loss-level: class-weighted loss, focal loss for hard examples, label smoothing; (3) metric-level: favor PR-AUC, F-beta (F2 if recall matters), not accuracy; (4) model-level: calibration (Platt, isotonic), probability thresholds tuned on validation; (5) deployment-level: two-stage cascade (cheap filter -> expensive model). Senior nuance: oversampling can leak duplicates into validation; use stratified splits. SMOTE assumes feature-space density assumptions that don't always hold. Class weighting is usually the cleanest lever. Threshold-tuning at deployment is often more impactful than any training change.