A data stream classification methods adaptive to concept drift
GE Mao-song · Caai Transactions on Intelligent Systems · 2007
At present, most classification methods for data streams are developed with the assumption of steady data distribution. However, the data collected from the real world will change over a period of time in the underlying concepts (known as concept drifting). This lowers the predictive precision of a classification model. This paper proposes a classification algorithm that can identify and adapt to occurrences of concept drifting according to the characteristics of the data stream. Experiments show that the proposed algorithm dynamically adjusts the size of the training window and the number of new examples during model reconstruction according to the current rate of concept drifting.