A multi-level weighted concept drift detection method

The concept drift detection method is an online learner. Its main task to determine the position of drifts in data stream, so as reset classifier after detecting improve learning performance, which very important practical applications such user interest prediction or financial transaction fraud detection. Aiming at inability existing methods balance delay, false positives, negatives, and space–time efficiency, a new level transition threshold parameter proposed, multi-level weighted mechanism including "Stable Level-Warning Level-Drift Level" innovatively introduced instances window are levels, double sliding also applied. Based on this, (MWDDM) proposed. In particular, two variants MWDDM_H MWDDM_M proposed based Hoeffding inequality Mcdiarmid inequality, respectively. Experiments artificial datasets show that can detect abrupt gradual faster than other comparison algorithms while maintaining low positive ratio negative ratio. real-world MWDDM has highest classification accuracy most cases good space time efficiency.