Google Professional Machine Learning Engineer Question 416
Single answerGoogle Cloud PlatformYou are managing a machine learning model deployed on Google Cloud AI Platform to predict customer churn. The model's performance is monitored through a dashboard that tracks key metrics such as accuracy, precision, recall, and data drift. Recently, you notice a slight decline in accuracy and a significant increase in data drift for the input features. What would be the most appropriate retraining policy in this scenario?
- A
Retrain the model immediately using the most recent data.
- B
Wait until the accuracy drops below a predefined threshold before retraining.
- C
Retrain the model on a fixed schedule, regardless of performance metrics.
- D
Retrain the model only after collecting additional labeled data to address the drift.
Show answer and explanation
Correct answer: A
Explanation
In this scenario, the model's performance metrics indicate both a decline in accuracy and significant data drift. Retraining the model immediately with the most recent data ensures that it adapts promptly to changes in the data distribution, maintaining its performance. Waiting too long or adhering to a fixed retraining schedule could lead to prolonged periods of suboptimal performance, while delaying retraining for additional labeled data may not be practical given the immediate need for adaptation.
- A. Correct.
Retraining the model immediately using the most recent data addresses both the decline in accuracy and the increase in data drift, ensuring the model adapts to the new data distribution.
- B. Incorrect.
Waiting until the accuracy drops below a predefined threshold might result in prolonged poor performance, especially since data drift indicates that the model's input distribution has changed.
- C. Incorrect.
Retraining on a fixed schedule ignores the immediate issue of data drift and performance decline, which may lead to suboptimal results if the model is not retrained in time.
- D. Incorrect.
While collecting additional labeled data to address the drift is useful, it may delay retraining unnecessarily and does not address the immediate need to adapt the model to changing data.