select TWO: DetectEye operates a fraud model with 0.3% positive class and plans to train an XGBoost model on SageMaker. They want to address severe class imbalance while preserving as much signal as possible and avoiding label noise amplification. Which TWO interventions are most appropriate to try first in this production context?

SMOTE (Synthetic Minority Over-sampling Technique) is appropriate because it generates synthetic samples for the minority class, directly addressing the imbalance by increasing its representation without simply duplicating existing data, thus preserving signal. Validating synthetic sample quality is crucial to avoid amplifying label noise. Setting scaleposweight in XGBoost is also highly effective. This parameter adjusts the weight of positive samples during training, making the model pay more attention to the rare class without altering the dataset distribution itself. This prevents discarding valuable majority class information while still mitigating the imbalance. Randomly undersampling to a 1:1 ratio would discard too much majority class data (99%), potentially losing critical information. Using Athena for SMOTE-like generation at query time is not a standard or efficient practice for synthetic data generation. Disabling early stopping and increasing the learning rate on imbalanced data would likely lead to overfitting or unstable training, not better minority class learning.