MLA-C01 Question 363
Select 3A company has deployed a machine learning model on Amazon SageMaker for real-time inference. The model's predictions are becoming less accurate over time, and the data distribution in the input features appears to have changed. What actions should a Machine Learning Engineer take to address this issue?
- A
Enable Amazon SageMaker Model Monitor to monitor data drift and set up alerts.
- B
Retrain the model with the current data and redeploy it to SageMaker.
- C
Use Amazon CloudWatch to directly log the model's predictions and monitor data drift.
- D
Implement input data validation to detect anomalies and filter invalid data before inference.
- E
Replace the current model with a pre-trained model from AWS Marketplace without retraining.
Show answer and explanation
Correct answers: A, B, D
Explanation
The observed drop in accuracy is likely due to data drift, which occurs when the data distribution changes over time. Addressing this issue involves monitoring the data for drift, retraining the model with updated data to reflect the new distribution, and implementing input data validation to maintain data quality. These steps ensure that the model remains accurate and reliable. Amazon CloudWatch and replacing the model without retraining are not effective solutions in this scenario.
- A. Correct.
Correct: Enabling Amazon SageMaker Model Monitor allows you to track data drift and receive alerts when significant changes to the data distribution are detected.
- B. Correct.
Correct: Retraining the model with current data ensures that the model adapts to the new data distribution, improving accuracy over time.
- C. Incorrect.
Incorrect: Amazon CloudWatch is primarily used for logging and monitoring system performance, but it does not natively support data drift detection for machine learning models.
- D. Correct.
Correct: Input data validation helps ensure the quality of the data being fed into the model, reducing the likelihood of inaccuracies caused by invalid or anomalous data.
- E. Incorrect.
Incorrect: Replacing the model with a pre-trained model without retraining may not address the specific data drift issue, as the pre-trained model may not be tailored to the company’s data.