Google Professional Machine Learning Engineer Question 415
Single answerGoogle Cloud PlatformYou are managing a machine learning model deployed on Google Cloud that predicts inventory demand for a retail company. The model's performance is monitored using a validation dataset, and you observe a gradual decline in its accuracy over time. Additionally, the company introduces new product lines every quarter, which significantly changes the data distribution. What would be the most appropriate retraining policy for this model?
- A
Retrain the model at fixed intervals, such as every month, regardless of performance metrics.
- B
Retrain the model when there is a statistically significant drop in performance metrics on the validation dataset.
- C
Retrain the model whenever new data is added to the dataset, regardless of its impact on the data distribution.
- D
Retrain the model quarterly to align with the introduction of new product lines and monitor performance in between.
Show answer and explanation
Correct answer: D
Explanation
The correct retraining policy should balance computational efficiency with ensuring the model remains effective. In this scenario, retraining quarterly aligns with the known data distribution changes caused by the introduction of new product lines, while monitoring performance in between ensures any unexpected performance issues are promptly addressed. This approach strikes an effective balance between proactive and reactive retraining strategies.
- A. Incorrect.
Retraining at fixed intervals, such as every month, could be inefficient and unnecessary if the model's performance remains stable. This approach does not account for changes in the data distribution or performance metrics.
- B. Incorrect.
Retraining based on a statistically significant drop in performance is a good practice, but it does not address the known regular changes in data distribution due to the introduction of new product lines every quarter.
- C. Incorrect.
Retraining whenever new data is added is not practical and can lead to excessive computational costs, especially if the new data does not significantly impact the model's performance or the data distribution.
- D. Correct.
Retraining quarterly aligns with the known changes in data distribution caused by the introduction of new product lines. Monitoring performance in between allows you to address unexpected issues with the model’s accuracy.