Question:

The question involves configuring an XGBoost model deployed on Google Cloud's Vertex AI for online prediction, focusing on explanation methods and model monitoring for minimal latency and drift detection.

Options:

• A: Uses sampled Shapley with a path count of 5, deploys to Vertex AI Endpoints, and creates a Model Monitoring job using prediction drift.
• B: Uses Integrated Gradients with a path count of 5, deploys to Vertex AI Endpoints, and creates a Model Monitoring job using prediction drift.
• C: Uses sampled Shapley with a path count of 50, deploys to Vertex AI Endpoints, and creates a Model Monitoring job using training-serving skew.
• D: Uses Integrated Gradients with a path count of 50, deploys to Vertex AI Endpoints, and creates a Model Monitoring job using training-serving skew.

Answer:

The suggested answer is A. This option balances speed and accuracy by using sampled Shapley with a lower path count (5) for faster online explanations and monitors for prediction drift, which is more suitable for online predictions. Higher path counts in sampled Shapley (like 50 in options C and D) improve accuracy at the cost of speed. Monitoring for training-serving skew is unnecessary for online prediction scenarios.