Drift-detection Based Incremental Ensemble for Reacting to Different Kinds of Concept Drift

Abstract

Data stream mining has attracted attention in recent years due to its wide range of applications. Concept drift is a great challenge for learning data streams. The existing algorithms are generally designed for a particular type of concept drift. However, real-world data stream applications are always complex combinations of many types of concept drift. In this paper, We proposed a data-stream ensemble classifier for reacting to different types of concept drift, called Drift-detection based Incremental Ensemble (DIE). DIE combines the operators of concept-drift detection and component update mechanism to handle concept drift. In the chunk-based framework, a drift detector is used to monitor the dynamics of data distribution. When a concept drift is triggered, DIE uses the alternative tree of Hoeffding Adaptive Tree to replace the old one, rather than just updating the weights of ensemble members, which can enhance the ability of the model to deal with sudden drift. We also present a component update mechanism to adjust previous ensemble members using the latest examples. Thus, DIE is suitable for handling slow drift. Experimental studies demonstrate the effectiveness of DIE in dealing with different kinds of concept drift.