Google Professional Machine Learning Engineer Question 503
Select 3Google Cloud PlatformYou are a Machine Learning Engineer responsible for maintaining a deployed ML model on Google Cloud. The model uses tabular data with multiple features, and you have noticed a degradation in the model's performance. You suspect that feature attribution drift might be the cause. Which actions should you take to monitor and address feature attribution drift effectively?
- A
Use Explainable AI to monitor feature attributions over time and identify changes in the importance of features.
- B
Implement Vertex AI Model Monitoring to track feature distribution statistics and set thresholds for drift detection.
- C
Retrain the model periodically on the same dataset used during initial training to ensure consistency.
- D
Compare feature attribution scores for new data against the baseline scores from the training dataset.
- E
Enable AutoML Tables' automatic retraining feature to address potential drift in feature attributions.
Show answer and explanation
Correct answers: A, B, D
Explanation
Monitoring feature attribution drift is crucial for maintaining the performance of a deployed model. By leveraging Explainable AI for feature importance analysis, using Vertex AI Model Monitoring for tracking drift, and comparing new feature attribution scores to baselines, you can effectively identify and address changes in how the model interprets features. Automatic retraining or retraining on the same dataset does not directly solve the problem of feature attribution drift, which is often caused by changes in the input data distribution.
- A. Correct.
Correct. Explainable AI tools in Google Cloud can help you inspect and monitor changes in feature importance over time, allowing you to detect attribution drift.
- B. Correct.
Correct. Vertex AI Model Monitoring provides robust tools for tracking feature distributions and detecting drift in real-time.
- C. Incorrect.
Incorrect. Retraining the model on the same dataset does not address feature attribution drift, as the drift likely arises from changes in the incoming data, not the stability of the original dataset.
- D. Correct.
Correct. Comparing new feature attribution scores with baseline scores can identify shifts in how the model interprets features, which is critical in detecting drift.
- E. Incorrect.
Incorrect. While AutoML Tables offers automatic retraining, it does not specifically monitor or handle feature attribution drift, which requires more targeted tools and analysis.