Google Professional Machine Learning Engineer Question 521
Select 4Google Cloud PlatformYou are deploying a machine learning model on Google Cloud that predicts customer churn. After deployment, you observe that the model's predictions are significantly different from what was expected. Which of the following steps should you take to diagnose and monitor for common training and serving errors in this scenario?
- A
Compare feature distributions between training and serving datasets to detect data drift.
- B
Check for numerical precision mismatches between the training and serving environments.
- C
Re-train the model using the same serving data, even if the root cause is not identified.
- D
Set up Vertex AI Model Monitoring to track prediction quality and feature skew over time.
- E
Verify that the model's prediction schema matches the schema of the serving input data.
Show answer and explanation
Correct answers: A, B, D, E
Explanation
Common training and serving errors often arise from issues such as data drift, feature skew, numerical precision mismatches, or schema mismatches. Proper monitoring mechanisms, such as Vertex AI Model Monitoring, and proactive diagnostic measures, such as comparing feature distributions and verifying schemas, are essential to ensure consistent and reliable model performance in production. Re-training the model without investigating the root cause is not a best practice.
- A. Correct.
Comparing feature distributions can help detect data drift or feature skew, which are common causes of serving errors.
- B. Correct.
Numerical precision mismatches between training and serving environments can lead to discrepancies in model predictions.
- C. Incorrect.
Re-training the model without addressing the root cause is not a recommended approach as it may not solve the issue and wastes resources.
- D. Correct.
Vertex AI Model Monitoring enables tracking of feature and prediction drift, which is crucial for identifying and addressing serving issues over time.
- E. Correct.
Ensuring the model's prediction schema matches the serving input schema is essential to avoid errors caused by schema mismatches.