Theoretical Window Size for Classification in the Presence of Sudden Concept Dr ift
Ludmila Ilieva Kuncheva · 2010
In classifying sequential data, a new classifier is needed af ter a sudden concept change. However, the old classifier may b e better than a new classifier trained on a small window of new data. We d erive a general formula for the size of this window, with a closed-form expression for two equiprobable Gaussian classes. Numerical experiments demonstrate that swapping the classifiers after the window has been acquired is better than using the new classifier right after the change or not modifying the class ifier at all.