Google Professional Machine Learning Engineer Question 418
Single answerGoogle Cloud PlatformYou are managing a machine learning model deployed on Google Cloud AI Platform that predicts customer churn for a subscription-based service. The model’s performance has been stable for the past six months, but recently, you noticed a significant drop in its accuracy. After investigating, you find that customer behavior has changed due to a new marketing campaign and the introduction of new subscription tiers. How should you define a retraining policy to address this issue while ensuring the model remains accurate over time?
- A
Schedule retraining of the model at fixed intervals, such as every 6 months, regardless of performance.
- B
Implement a monitoring system to track the model's performance and trigger retraining when key metrics like accuracy or precision fall below a defined threshold.
- C
Retrain the model only when new features are added to the dataset.
- D
Set up a daily retraining schedule to ensure the model is always using the latest data.
Show answer and explanation
Correct answer: B
Explanation
The correct approach is to implement a monitoring system that tracks the model's performance and triggers retraining when key metrics fall below a threshold. This ensures the model adapts to data distribution changes, such as those caused by the new marketing campaign and subscription tiers, while avoiding unnecessary retraining and resource expenditure.
- A. Incorrect.
Scheduling retraining at fixed intervals may not address the issue of performance deterioration caused by changing data patterns. Retraining should be driven by performance metrics and data changes.
- B. Correct.
Tracking key metrics and triggering retraining when performance degrades ensures that the model adapts to changes in data patterns, such as those caused by new marketing campaigns or subscription tiers. This approach is dynamic and efficient.
- C. Incorrect.
Retraining only when new features are added focuses on model complexity but does not address performance drops caused by shifts in data distribution or customer behavior.
- D. Incorrect.
Daily retraining is unnecessary and computationally expensive unless the data changes at very high frequency, which is not the case here. This approach can lead to wasted resources without significant performance gains.