Fast concept drift detection using unlabeled data

Streaming data mining is in use today in many industrial applications, but performance of the models is deteriorated by concept drift, especially when true labels are unavailable. This paper addresses the need of detecting concept drifts under unsupervised situation and proposes the Unsupervised Concept Drift Detection (UCDD) method. A cluster technique is first applied to determine artificial labels of the data set, then a fast drift detection algorithm is used to detect the boundary change between the labeled clusters. Through the empirical evaluation, the method demonstrates effectiveness on detecting various types of concept drifts.