Learning with Concept Drift Detection based on Sub-concepts from k Time Sub Windows

Learning with Concept Drift Detection based on Sub-concepts from k Time Sub Windows

Concept drift detection has attracted much interest recently, due to its pervasive nature in the massive amount of streaming data available for analysis. Traditional concept drift detection methods, based on the monitoring performance of a base learner on a whole time window of data stream, are not sensitive enough to sub-concept drifts and discover them late or not at all. This is because, when aggregated together, the subconcepts that form a concept are not precisely described. To solve this problem, we propose the kTSW (k Time Sub-concepts Window) based framework that divides instances from a whole time window into k sub-concept windows, and then builds a drift monitor for each subconcept window. Once a sub-concept window’s instances have experienced a concept drift, we update the learned model. We propose three schemes with different base learner numbers for our framework. Each of the schemes takes advantage of a different degree of sub-concept knowledge. Two real data sets are used to verify the validity of our method in data stream classification. Experimental results show that our method is able to obtain higher accuracy and recall than methods based on a whole time window.