Learning with Drift Detection based on k Time Sub-concept Windows

Concept drift detection has attracted much interest lately, due to the fact that a massive amount of streaming data are continuously being generated. Traditional concept drift detection methods, based on the monitoring performance of a base learner over instances from a data stream's whole time window, are not sensitive enough to sub-concept drifts and discover them late or not at all. This is because when aggregated together, the sub- concepts that form a concept are not carefully described. To solve this problem, we propose the kTSW based method that divides instances from a data stream's whole time window into \pmbk subconcept windows, and then builds a drift monitor for instances from each sub window. Once a sub window's instances have experienced a concept drift, we update the learned model. Two real data sets are used to verify the validity of our method in data stream classification. Experimental results show that our method obtained higher accuracy and recall than methods based on the whole time window.