Concept Drift Data Stream Classification Algorithm Based on McDiarmid Bound
LI Guanghui LIANG Bin · DOAJ (DOAJ: Directory of Open Access Journals) · 2021
Concept drift in data streams can cause significant performance degradation of existing classification models. Most current data stream algorithms for concept drift only aim at a certain type of concept drift (such as abrupt, gradual, or recurring drift), which is difficult to adapt to different scenarios. Therefore, this paper proposes a new data stream algorithm suitable for different types of concept drift. The proposed algorithm saves the latest classification results through a two-layer window, assigns weights to it based on the membership function and calculates the weighted error rate. Then the McDiarmid bound is used to analyze the difference [δ] between the error rates of current window and the past window, and the concept drift is detected according to the significance of[δ]. After detecting drift, the semi-parametric log-likelihood algorithm is used to check whether the current new concept is a recurrence of the past concept, and then whether to reuse the old classifier is decided. Experimental results show that, the proposed algorithm outperforms the similar existing algorithms in terms of average detecting delay, false positive rate, classification accuracy and running time.