How to handle class imbalance in production ML models - cost-sensitive learning vs resampling?