Using SageMaker's XGBoost, a fraud detection model fits the training data very well but fails to detect fraud on new, unseen transactions. What change should the ML engineer make to improve generalization to new transactions?

The model is overfitting, indicated by excellent training performance but poor generalization to new data. Decreasing the maxdepth hyperparameter in XGBoost reduces the complexity of individual trees, making the model less likely to memorize the training data and more likely to learn general patterns. This improves its ability to generalize to unseen transactions. Increasing the learning rate or maxdepth would further exacerbate overfitting. Removing irrelevant features is a good practice for model efficiency and interpretability, but directly addressing overfitting in a tree-based model often involves constraining tree growth.