An ML engineer is developing a fraud detection model by using the Amazon SageMaker XGBoost algorithm. The model classifies transactions as either fraudulent or legitimate. During testing, the model ex

Question

Options

• AIncrease the learning rate.
• BRemove some irrelevant features from the training dataset.
• CIncrease the value of the max_depth hyperparameter.
• DDecrease the value of the max_depth hyperparameter.

• A
  9% (6)
• B
  3% (2)
• C
  4% (3)
• D
  84% (56)

Explanation

Decreasing max_depth reduces the complexity of each decision tree in the ensemble, preventing the model from memorizing the training data's noise and patterns that don't generalize - the classic fix for overfitting, which is exactly what the scenario describes (high train accuracy, low test accuracy). Option A (increasing learning rate) makes the model converge faster but amplifies overfitting rather than correcting it. Option B (removing irrelevant features) can help generalization in some cases, but the question points specifically to a depth/complexity problem, not a feature noise problem. Option C (increasing max_depth) is the opposite of what's needed - deeper trees capture more granular splits, making overfitting worse.

Memory tip: Think of max_depth as "how many questions the tree is allowed to ask." A model allowed to ask 20 questions can memorize your training set; one limited to 5 questions must learn the real patterns. When the model "knows the exam but fails the test," shorten the depth.

No community discussion yet for this question.