Unsupervised Online Concept Drift Detection Based on Divergence and EWMA.

Concept drift problem is a common challenge for data stream mining, while the underlying distribution of incoming data unpredictably changes over time. The classifier model in data stream mining must be self-adjustable to the concept drift, otherwise it will get terrible classification results. To detect concept drift timely and accurately, this paper proposes an unsupervised online Concept Drift Detection algorithm based on Jensen-Shannon Divergence and EWMA(CDDDE), which detects concept drift through measuring the difference of data distribution within sliding windows and calculating the drift threshold dynamically by Exponentially Weighted Moving Average (EWMA), during the detection without the use of labels. Once concept drift is detected, a new classifier would be trained using the current and subsequent data. Experiments on artificial and real-world datasets show that CDDDE algorithm can efficiently detect the concept drift, and the retrained classifier effectively improves the classification accuracy for the subsequent data. Compared with some supervised algorithms, the detection accuracy and classification accuracy are higher for most datasets.