Google Professional Machine Learning Engineer Question 414
Single answerGoogle Cloud PlatformYou are leading a team responsible for deploying a machine learning model that predicts product demand. The model was trained on historical data and has been in production for three months. Recently, you noticed a decline in prediction accuracy, and an analysis shows that the data distributions in the new data differ significantly from the training data. What would be the most appropriate retraining policy in this scenario?
- A
Retrain the model on a fixed schedule, such as every six months, regardless of data changes.
- B
Retrain the model only when there is a significant performance degradation or data drift detected.
- C
Retrain the model daily to ensure it always accounts for the latest data.
- D
Do not retrain the model until at least one year has passed since the initial deployment.
Show answer and explanation
Correct answer: B
Explanation
In this scenario, the decline in prediction accuracy and observed data drift indicate that the model is no longer well-aligned with the current data distribution. The most appropriate retraining policy is to retrain the model when significant performance degradation or data drift is detected, as this ensures the model adapts to changes in the data while avoiding unnecessary retraining.
- A. Incorrect.
Retraining on a fixed schedule might not address the current issue because the model needs immediate attention due to significant data drift. Fixed schedules are not ideal for dynamic environments.
- B. Correct.
Retraining when significant performance degradation or data drift is detected is the most appropriate approach in this case, as the issue is directly tied to changes in the data distribution.
- C. Incorrect.
Retraining daily is unnecessary and inefficient unless the use case requires extremely frequent updates, which is not mentioned in this scenario. It could also lead to overfitting or resource wastage.
- D. Incorrect.
Waiting for a year before retraining is not advisable, as the model's performance has already degraded significantly, and the data drift needs to be addressed promptly.