Reactive Concept Drift Detection Using Coresets Over Sliding Windows

The change of underlying data is one of the biggest challenges in non-stationary environments. While several algorithms have been proposed to detect these changes, substantial problems remain in the case of higher dimensional data. Thus, we propose a novel Concept Drift detector based on Minimum Enclosing Balls, with the capability to quickly process higher dimensional data. Additionally a kernelized version of this detector is derived, to process non-linear streaming data. We also propose a method to measure the performance of drift detectors with a binary classification evaluation technique, the confusion matrix, which enables calculating statistics like the F1 score. Our experiments show, that this novel technique is superior to existing state-of-the-art Concept Drift detectors regarding its true positive detection rate.