A Cluster-Aided Ensemble Classification Method for Data Stream with Mixed Concept Drift
P. R. Liu, Kui Ge, Yonghe Zhang, Bangxin Xu · 2024
Concept drift, i.e. the change of data distribution in data stream, is one of the main factors to reduce the accuracy of data stream classification. Among the methods to deal with concept drift, concept drift adaptation methods have been shown to be very competitive. However, in existing ones, the solution of solving mixed concept drift which means multiple concept drifts occur at the same time is seldom taken into account. To this end, in this paper, we suggest a multi-level cluster framework to reserve data and utilize the value of the historical data for respectively updating mixed drift, where two cluster-based updating strategies are proposed for updating the internal and external weights of instances and selecting the instances suitable for mixed concept drift. The two weights of instances are used to measure the contribution of instances to the quality and diversity of instance space. In multi-level cluster framework, the change of space caused by concept drift is decomposed into the change of cluster, and the range of cluster is used to simulate the range of drift. Accordingly, the cluster-aided ensemble classification is developed to adapt to data stream with mixed concept drift. Experimental results on synthetic and real-world datasets show that our algorithm outperforms existing concept drift adaptation methods.