1
$\begingroup$

I'm currently struggling with concept drift problem in on-line learing. I read some papers "Ikonomovska Gama - Regression Trees from Data Streams with Drift Detection " and check their implementation of AM Rules Regresor in MOA framework and their concept drift detector based on knowing real output of each sample.

In real-word cases I don't know real output (or know after some time). So is there some statistical test that tells if there is a change in data distribution based only on sample features?

$\endgroup$
1
$\begingroup$

Most existing approaches for handling concept drift make the bit unrealistic (for most real world application imo) assumption that the labeled data will be available at no labeling cost shortly after classification. However there are some proposed alternatives, I've come across this one :

  • Lindstrom, Patrick, Brian Mac Namee, and Sarah Jane Delany. "Drift detection using uncertainty distribution divergence." Evolving Systems 4.1 (2013): 13-25

So maybe it's a good starting point. hth.

$\endgroup$
  • $\begingroup$ Welcome to our site! Could you complete the citation by adding year and publisher/journal information, as it would appear in a bibliography? $\endgroup$ – Silverfish Mar 30 '16 at 10:26

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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