Google Professional Machine Learning Engineer Question 502
Select 3Google Cloud PlatformYour team has deployed a machine learning model on Google Cloud that predicts customer churn. You notice that the model's performance has been degrading over the past few weeks. Upon investigation, you suspect feature attribution drift might be the cause. How can you monitor for feature attribution drift using Google Cloud tools?
- A
Use Vertex AI Model Monitoring to track changes in feature attribution distributions over time.
- B
Manually compute SHAP values for historical and recent predictions to identify changes in feature importance.
- C
Enable Vertex AI Explainable AI to automatically monitor feature attribution drift for deployed models.
- D
Use BigQuery to store feature attribution data and write custom SQL queries to compare historical and current explanations.
- E
Deploy a custom model in Vertex AI that specifically predicts feature attribution drift.
Show answer and explanation
Correct answers: A, B, D
Explanation
Monitoring feature attribution drift involves tracking changes in how features influence model predictions over time. Vertex AI Model Monitoring can help with input feature drift, but for attribution drift, you may need to use manual methods such as computing SHAP values or analyzing stored attribution data in BigQuery. Vertex AI Explainable AI provides explanations but does not directly track attribution drift.
- A. Correct.
Vertex AI Model Monitoring can track data drift and alert on changes in input features, but you must use custom setups or tools to monitor feature attribution drift specifically.
- B. Correct.
Manually computing SHAP values for historical and recent predictions is a valid approach to compare feature importance and identify drift in attribution.
- C. Incorrect.
Vertex AI Explainable AI provides explanations for model predictions, but it does not automatically monitor or alert you about feature attribution drift.
- D. Correct.
Using BigQuery to store feature attribution data and analyzing it with SQL queries is a feasible way to monitor feature attribution drift manually.
- E. Incorrect.
Deploying a custom model to predict feature attribution drift is unnecessary and not a standard practice for handling this issue.