Concept Drift and Model Decay in Machine Learning

Hello, fellow machine learning enthusiasts. Today, we are going to learn about Concept Drift and Model Decay in Machine Learning.

Well if you had been working in this field then you might be knowing concept drift leads to failure of even well-built machine learning model.

But, what is this Concept Drift?


Concept Drift:

Concept drift comes into the picture when we have a well-trained model that we have trained over data for a very long time.

But as the new data comes in overtime, our model begins to classify the previous data in a different way, which might lead to various flaws in the predictions and hence leads to a model decay.

This happens because over the course of time our thoughts about somethings change and hence we tend to label a set of data differently.

And hence over time, this change in ideology affects the previous dataset as well.

Consider the example of fraudulent transactions.

Over time the idea of what can be considered to be fraudulent transaction changes.

These change needs to be implemented by the model, but if a model fails to implement or cope up with the change, the model fails and such a problem is called concept drift.

But, how do we counter such an issue?


The Remedy:

Well, there are various ways that we can tackle this the issue on concept drift in machine learning:

  • Routinely Re-Fit the model:
  • -> Whenever new data is added we can update our model again to fit the most recent historical data.
  • Routinely update the model: 
  • -> We consider the existing model as the starting point and update it with the most recent data.
  • Weight data: 
  • ->We can configure models such that we assign more importance to the recent data than the older days.
  • Fit the Difference:
  • -> We can tackle this problem through the ensemble approach as well, in which what we do is we keep our model untouched.
  • -> And a new model can be trained to correct the prediction from the model we have based on the most recent data.
  • Data Formatting: 
  • -> Our data can be formatted in such a way that seasonal change in data could be eradicated.


So, there we have it “Concept Drift and Model Decay in Machine Learning”.

I hope, you enjoyed the read.

Thanks for reading.


Leave a Reply

Your email address will not be published. Required fields are marked *