Your team trained and tested a DNN regression model with good results.
Six months after deployment, the model is performing poorly due to a change in the distribution of the input data.
How should you address the input differences in production?
Click on the arrows to vote for the correct answer
A. B. C. D.C.
The correct answer to this question is A. Create alerts to monitor for skew, and retrain the model.
When a model is deployed into production, it can encounter new data with different statistical properties from the data it was trained on, which is referred to as a change in the data distribution. As a result, the model may perform poorly, and the model may require adaptation to the new data distribution.
Option A, Create alerts to monitor for skew, and retrain the model, is the most appropriate solution. This solution involves creating a monitoring system that detects the distributional changes in the input data and notifies the team responsible for retraining the model. The team can then use the new data to retrain the model and improve its performance. The use of alerts can help to ensure that the model's performance is constantly monitored, and changes in the data distribution are quickly identified, and the model is updated as necessary.
Option B, Perform feature selection on the model, and retrain the model with fewer features, is not the best solution to address input differences in production. Feature selection can help to reduce the complexity of the model and improve its generalization ability. However, it is not a solution to address changes in the data distribution.
Option C, Retrain the model, and select an L2 regularization parameter with a hyperparameter tuning service, is not a direct solution to address changes in the data distribution. L2 regularization is a technique used to prevent overfitting in the model. While hyperparameter tuning can help to optimize the model's parameters, it may not necessarily address changes in the data distribution.
Option D, Perform feature selection on the model, and retrain the model on a monthly basis with fewer features, is not an appropriate solution to address changes in the data distribution. Retraining the model on a monthly basis can help to improve its performance. However, feature selection alone may not be sufficient to address changes in the data distribution. Additionally, retraining the model every month can be costly and time-consuming.
In conclusion, the best solution to address changes in the data distribution in production is to create alerts to monitor for skew, and retrain the model. This approach ensures that the model is updated when the distribution of the input data changes, thereby maintaining its accuracy and usefulness.