Model Drift

A Simple Explanation - By Varsha Saini

The machine learning model which is predicting amazing results today may not perform well after some time span. This degradation of model performance over a period of time is called Model Drift or Model Decay.

Model drift is an important concern to the business as their decisions rely so much on machine learning models today. If the prediction quality degrades, it can have a huge impact on business.

What causes the model to drift?

  • The relationship between independent variables and dependent variables may change over time.
  • Changes in the business environment, customer habits, economic pressure, or a natural disaster such as Covid-19.

How to prevent model drift

  • Regularly monitor the performance of the model.
  • Checking the quality of data.
  • Retraining the model with updated data at regular intervals.
  • Online learning, means making a machine learning model learn in real-time.

Types of Model Drift

There are two types of Model Drift.

1. Concept Drift

Concept drift happens when the properties of the target variable change i.e relationship between input variables and the target variable changes.

An example is a change in the template of spam email. The email which wasn’t considered spam before may be considered spam today.

2. Data Drift

Data Drift happens when the properties of independent variables change.

For example changes in the data due to seasonality, changes in consumer preferences, the addition of new products, etc.