model learning in the presence of noise

what is the effect of class label noise, in training data, on classification model learning??..

I have a training dataset and added class label noise in it. Now I have passed that noisy training data to the classifier ensemble. But the problem is, when the model learning starts, it doesn't take the same time to learn the model as it took without noise. It takes a very small time to learn the model in the presence of noise. I am unable to find the reasons why is this happening?? I have checked out all the things in the code but couldn't figure out the problem. I think model learning should take the same time or more in the presence of noise. Can u plz tell what should I do??..

Answer 1

[Ilya]

When you say "classifier ensemble", I assume you mean "decision tree ensemble". The time it takes to grow a tree depends on the data, both predictors and labels. If the classes are barely separable after you add noise, the tree may bail out early because it fails to find a good split. Individual trees are saved in the Trained property of the ensemble object. You can check how deep they are with and without label noise.

If the classification accuracy of the ensemble with added noise is at the level you would expect, I don't see why you need to worry about training being too fast.