MLS-C01 Question 374
Select 3You are a Machine Learning Engineer managing a fraud detection model hosted in Amazon SageMaker. The model's performance has started to degrade over time due to data drift in the incoming transactional data. You want to set up an automated pipeline to retrain the model when performance metrics fall below a certain threshold. Which components and services should you use to build this solution?
- A
Amazon SageMaker Model Monitor to detect data drift and trigger retraining
- B
Amazon SageMaker Pipelines to orchestrate the retraining workflow
- C
AWS Lambda to manually monitor metrics and invoke the retraining pipeline
- D
Amazon CloudWatch to monitor model performance and trigger alerts
- E
Amazon SageMaker Ground Truth to label new incoming data for retraining
Show answer and explanation
Correct answers: A, B, D
Explanation
To automate retraining pipelines effectively, you need to monitor the model's performance using tools like SageMaker Model Monitor and CloudWatch. Amazon SageMaker Pipelines is then used to orchestrate the retraining workflow when thresholds are breached, ensuring an end-to-end automated solution. AWS Lambda and Ground Truth are not essential components for this specific use case.
- A. Correct.
Amazon SageMaker Model Monitor is designed to detect data drift or anomalies in input data or model predictions. It can trigger notifications or pipeline actions, making it a critical component for retraining.
- B. Correct.
Amazon SageMaker Pipelines is a fully managed service that helps automate and orchestrate ML workflows, including retraining pipelines.
- C. Incorrect.
AWS Lambda is not typically required for this scenario since SageMaker Pipelines and CloudWatch can handle monitoring and triggering actions natively.
- D. Correct.
Amazon CloudWatch can monitor metrics such as model performance and invoke alarms or events when thresholds are breached, which can trigger the retraining pipeline.
- E. Incorrect.
Amazon SageMaker Ground Truth is used for data labeling, but labeling may not be necessary in this scenario if existing labels are available or unsupervised methods are used.