A Multiscale Concept Drift Detection Method for Learning from Data Streams

Concept drifts can occur in data streams, which disable any models built on static data distribution to fit dynamic or recurrent concepts. If there is a feature that can monitor the stableness of such distribution, then we have a proper reference to adapt the model. Based on this idea, we propose a novel approach named Multiscale Drift Detection Test (MDDT) that localizes abrupt drift points when detection feature values fluctuate. MDDT is based on a resampling scheme and a paired student t-test. It applies a detection procedure on a broad and a narrow scale. This multiscale structure not only reduces massive time of a constant checking process, but also filters noise in the detection features. Experiments are performed via synthetic and real-world datasets. The results indicate that the proposed method outperforms the state-of-art algorithms in terms of computation cost and average accuracy.