Online Active Learning Paired Ensemble for Concept Drift and Class Imbalance

Abstract

Practical applications often require learning algorithms capable of addressing data streams with concept drift and class imbalance. This paper proposes an online active learning paired ensemble for drifting streams with class imbalance. The paired ensemble consists of a long-term stable classifier and a dynamic classifier to address both sudden concept drift and gradual concept drift. To select the most representative instances for learning, a hybrid labeling strategy which includes an uncertainty strategy and an imbalance strategy is proposed. The uncertainty strategy applies a margin-based uncertainty criterion and a dynamic adjustment threshold. Based on the categorical distribution of the last data block, the imbalance strategy prefers to learn instances of the minority category. In addition, it also incorporates the advantages of the traditional random strategy and helps to capture the drifts away from the decision boundary. Experiments on real datasets and synthetic datasets utilize prequential AUC as an evaluation index, comparing the classification performance of our method with semi-supervised and supervised learning methods. The results show that the proposed methods can obtain higher AUC values at an even lower labeling cost. Moreover, it is noteworthy that the labeling cost can be dynamically allocated according to the concept drift and imbalance ratio.