MLA-C01 Question 376
Select 3You are an AWS Certified Machine Learning Engineer responsible for maintaining an e-commerce recommendation system deployed on Amazon SageMaker. Over time, you notice that the click-through rate (CTR) of the recommendations has significantly decreased. You suspect that the model’s performance is affected by data drift. Which of the following actions should you take to address potential data drift?
- A
Analyze recent input data and compare its distribution with the training dataset.
- B
Use Amazon SageMaker Model Monitor to detect changes in input data features.
- C
Increase the training dataset size by collecting more historical data.
- D
Retrain the model with updated input data that reflects the current distributions.
- E
Deploy a new Amazon SageMaker endpoint immediately to reset the model.
Show answer and explanation
Correct answers: A, B, D
Explanation
Data drift occurs when the distribution of input data changes over time, leading to degraded model performance. To address this, you must first identify whether drift has occurred by analyzing the input data and leveraging tools like Amazon SageMaker Model Monitor. If drift is confirmed, retraining the model with updated data that reflects the current distribution is crucial to restore performance. Increasing the training dataset indiscriminately or deploying a new endpoint without addressing the underlying issue will not resolve the problem.
- A. Correct.
Analyzing recent input data and comparing it with the training dataset is essential to identify whether data drift has occurred, as drift indicates a change in the data distribution over time.
- B. Correct.
Amazon SageMaker Model Monitor is specifically designed to track and detect changes in input data and can be used to diagnose data drift in real-time.
- C. Incorrect.
While increasing the training dataset size might be helpful in some cases, using outdated or irrelevant historical data may not address the issue of data drift caused by recent changes in input data distributions.
- D. Correct.
Retraining the model with updated input data that reflects current distributions is a key step to mitigate the effects of data drift and improve model performance.
- E. Incorrect.
Deploying a new endpoint will not address the root cause of data drift, as the issue lies in the model’s training data and its ability to generalize to the current input data.