When to Retrain?

A model's performance can degrade over time as the data it's seeing in production drifts away from the data it was trained on. To maintain a high-quality model, you need to have a strategy for retraining it on a regular basis.

---

1. Retraining Triggers

There are three main types of triggers for retraining your model:

• Scheduled Triggers: Retrain your model on a regular schedule (e.g., every day, every week). This is the simplest approach and is a good choice if your data distribution changes at a predictable rate.
• Performance-based Triggers: Retrain your model when its performance drops below a certain threshold. This is a more sophisticated approach that requires you to monitor your model's performance in production.
• Data-based Triggers: Retrain your model when new data becomes available. This is a good choice if your data is constantly changing and you want to keep your model as up-to-date as possible.

---

2. Defining a Retraining Policy

When defining a retraining policy, you should consider the following factors:

• The rate of data drift: How quickly does your data distribution change? If your data drifts quickly, you will need to retrain your model more frequently.
• The cost of retraining: Retraining a model can be expensive, so you need to balance the cost of retraining with the benefit of having a more accurate model.
• The importance of model freshness: How important is it for your model to be up-to-date? For some applications (e.g., fraud detection), it's crucial to have a model that is trained on the latest data.

---

Knowledge Check