KAPPA as Drift Detector in Data Stream Mining

Abstract

Abstract Concept Drift is considered a challenging problem that appears in data streaming. The classifier’s error rate and the ensemble are used in most of the previous works to manage classification accuracy as a criterion for judging whether concept drift is happening or not. KAPPA is an effective way to measure the level of agreement, and it may be suitable to detect concept drift in a reliable, fast, and computationally efficient way. In this paper, we propose a new concept drift detector, called KAPPA, which aims at reacting to detect concept drift in a reliable, fast, and computationally efficient way. Contrary the disagreement measure that we have already considered in our preliminary work (DMDDM), KAPPA would measure the level of agreement when different classifiers access data items is suitable to detect concept drifts. The performance of KAPPA has been experimentally compared with DMDDM on synthetic dataset streams, considering different performance measures, e.g., delay detection, true positives and the mean accuracy.