EMZD: Equal Means Z-Test Concept Drift Detector
Danilo Rafael de Lima Cabral, Roberto Souto Maior de Barros · 2020
Algorithms for extracting information from data streams must learn online, as the data distribution may change with time, a phenomenon usually known as concept drift. Drift detectors are programs that approximate the positions of the drifts to replace the base learner with the aim of increasing accuracy. EDDM is a simple and traditional drift detector but its performance is often weak. This paper proposes EMZD, which is rooted in EDDM and uses the z-test for equal means in the detection of concept drift. Experiments using artificial dataset generators, with abrupt and gradual concept drift versions plus real-world datasets, run in MOA, show that EMZD improves the accuracy and the detections of EDDM and other detectors in many situations.