Google Professional Machine Learning Engineer Question 519
Select 3Google Cloud PlatformYou are responsible for maintaining a machine learning model deployed on Google Cloud AI Platform. Users report that the model's predictions have become unreliable in production. Upon investigation, you identify a significant drift in the input feature distribution compared to the training data. Which actions should you take to address this issue?
- A
Retrain the model using updated data that reflects the current input distribution.
- B
Implement data monitoring to detect feature distribution changes in real time.
- C
Enable autoscaling for the serving infrastructure to handle increased traffic.
- D
Analyze the serving logs to identify patterns and potential root causes of the drift.
- E
Revert the model to a previous version that performed well in production.
Show answer and explanation
Correct answers: A, B, D
Explanation
Input feature distribution drift occurs when the data the model encounters in production deviates from the data it was trained on. To address this, you should retrain the model with updated data, monitor feature distributions in real time to detect future drift, and analyze serving logs to understand the root cause. These steps ensure the model remains reliable and performance issues are resolved effectively.
- A. Correct.
Retraining the model with updated data that reflects the current input distribution can help improve its performance and adapt it to the changed environment.
- B. Correct.
Implementing data monitoring helps detect feature distribution changes in real time, allowing you to address drift issues proactively.
- C. Incorrect.
Autoscaling is related to handling traffic load and does not address input feature distribution drift.
- D. Correct.
Analyzing serving logs can provide valuable insights into patterns or sources of the drift, helping in diagnosing the problem.
- E. Incorrect.
Reverting to a previous model version may temporarily solve the issue but does not address the root cause, which is the input feature distribution drift.