A credit card company needs to detect fraudulent transactions in real time. A labeled historical dataset (fraud vs. legitimate) is stored in Amazon S3. After removing missing values, a data scientist trains an XGBoost classifier in Amazon SageMaker. The model metrics are: TPR 0.700, FNR 0.300, TNR 0.977, FPR 0.023, and overall accuracy 0.949. Which action should the data scientist take to improve model performance given the apparent class imbalance?

The correct action is to apply SMOTE to the minority (fraud) class. The metrics show a high True Negative Rate (TNR) of 0.977 and a low False Positive Rate (FPR) of 0.023, indicating the model is good at identifying legitimate transactions. However, the True Positive Rate (TPR) of 0.700 and False Negative Rate (FNR) of 0.300 suggest it's missing a significant portion of fraudulent transactions. This pattern is typical of imbalanced datasets where the minority class (fraud) is underrepresented. SMOTE generates synthetic samples for the minority class, helping to balance the dataset and improve the model's ability to learn the characteristics of fraudulent transactions. Applying SMOTE to the majority class or oversampling the majority class would worsen the imbalance. Undersampling the minority class would further reduce the already limited data for fraud.