Google Professional Machine Learning Engineer Question 517
Select 2Google Cloud PlatformYou have deployed a machine learning model using Google Cloud AI Platform Prediction. After deployment, you notice a significant drop in model performance as reported by your monitoring metrics. Upon investigation, you suspect that the input data distribution during serving has shifted compared to the training data. Which monitoring approaches should you use to identify and address this issue?
- A
Enable Vertex AI Model Monitoring to track feature attribution drift and label drift.
- B
Configure Vertex AI Model Monitoring to analyze feature value distribution drift between training and serving datasets.
- C
Use Explainable AI to generate feature importance during serving and compare it to the importance during training.
- D
Set up Cloud Monitoring alerts to notify you of prediction latencies exceeding expected thresholds.
- E
Log and compare serving data payloads with training data stored in BigQuery datasets.
Show answer and explanation
Correct answers: B, E
Explanation
To monitor and address serving input data distribution shifts, it is crucial to use tools that focus on comparing the distributions of training and serving data. Vertex AI Model Monitoring provides built-in capabilities to analyze feature value drift, while logging and comparing serving data with training data enables deeper investigation into discrepancies. These approaches are effective in diagnosing and addressing performance degradation caused by data drift.
- A. Incorrect.
Vertex AI Model Monitoring does not track feature attribution drift or label drift directly; it primarily focuses on monitoring data distribution drift, prediction drift, and feature skew.
- B. Correct.
This is a correct option. Vertex AI Model Monitoring allows you to configure alerts for feature value distribution drift, which helps identify changes in input data distribution during serving.
- C. Incorrect.
While Explainable AI can provide feature importance, it is not a direct tool for monitoring data distribution drift. It is more suited for interpretability rather than detecting serving issues related to data shifts.
- D. Incorrect.
Cloud Monitoring alerts for prediction latencies are useful for identifying serving infrastructure issues but are not directly related to data distribution drift.
- E. Correct.
This is a correct option. Logging and comparing serving data with training data can help you detect and analyze data distribution shifts, which is essential for resolving performance drops due to input data drift.