An Ultimately Simple Concept Drift Detector for Data Streams

This research presents a novel and simple approach to the concept drift detection problem, a fundamental issue in data stream mining because the generated models need to be updated when significant changes in the underlying data distribution occur. A number of drift detectors has been proposed but most of them have limitations such as limited effect on the accuracy results, high computational complexity, poor sensitivity to gradual change, or the opposite problem of high false positive rate. The proposed method, Ultimately Simple Drift Detector (USDD), has low false positive and negative rates as well as low computational complexity. Extensive experiments against seven drift detectors on a wide variety of datasets reveals that, using limited resources, USDD delivers high accuracy in classification and, at the same time, maintains a competitive true detection rate when compared to the other methods.