Google Professional Machine Learning Engineer Question 506
Select 3Google Cloud PlatformYou have deployed a machine learning model in production that uses a set of features for predictions. Over time, you notice a drop in the model's performance. You suspect feature attribution drift as the cause. Which steps should you take to monitor and address this issue effectively?
- A
Set up a pipeline to calculate feature attribution scores on a regular basis using SHAP or similar methods.
- B
Compare historical feature attribution distributions with the current distribution using statistical tests.
- C
Retrain the model immediately whenever feature attribution drift is detected, without further investigation.
- D
Implement alerts to notify stakeholders when feature attribution drift crosses a predefined threshold.
- E
Ignore feature attribution drift if the input feature values remain within their expected ranges.
Show answer and explanation
Correct answers: A, B, D
Explanation
Monitoring feature attribution drift is critical in maintaining the performance and reliability of a deployed machine learning model. Steps such as calculating feature attribution scores, comparing distributions, and setting up alerts allow you to identify and respond to drift effectively. Ignoring drift or retraining without proper analysis can lead to avoidable errors and inefficiencies.
- A. Correct.
Calculating feature attribution scores regularly enables you to detect changes in how features influence your model's predictions over time, which is critical for identifying drift.
- B. Correct.
Comparing historical and current feature attribution distributions allows you to quantify and verify the presence of drift using statistical methods.
- C. Incorrect.
Retraining the model immediately without investigating the root cause of drift is not a best practice, as it may lead to suboptimal solutions or mask deeper issues.
- D. Correct.
Setting up alerts ensures that stakeholders are notified when feature attribution drift exceeds acceptable levels, enabling timely action.
- E. Incorrect.
Ignoring feature attribution drift is not advisable, as it can indicate changes in the relationships between features and the target variable, even if feature values themselves remain within expected ranges.