I am implementing a clustering-based algorithm for non-stationary data stream. Most concept drift techniques are based on change in classification output (or on its accuracy). Is their a way for detecting concept drift using only the distance between centroid of clustering that evolved over time?

Do you think it is possible to to monitor the change in the values of consecutive centroids of some cluster, and deduce the if the distance between them exceeds some preset threshold, there is some concept change?



1 Answer 1


Yes, several papers have been published on clustering data with drifts. They are easy to find.

I just don't think there is any real data to test or compare these algorithms, so it's not clear if they solve any real problem.

Where does your "drift" come from? Do you have data with labeled drift, or do you just plan to simulate this with synthetic data, too?

  • $\begingroup$ Thanks for your answer. I used both synthetic datasets (like SEA or hyperplane) along with real ones (like "Airlines"). I understand that the drift is related to the distribution of the class label, right? $\endgroup$
    – Michael
    Oct 16, 2017 at 16:24
  • 3
    $\begingroup$ In real clustering, you don't have a class label. $\endgroup$ Oct 16, 2017 at 18:37

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.