Dynamic weighted majority based on over-sampling for imbalanced data streams
Hongle Du, Palaoag Thelma · 2021
Imbalanced data stream with concept drift mining have gained the significant popular among researchers in recent years. The complexity of data stream mining algorithm with concept drift and imbalance will be greatly increased. In this paper, dynamic weighted majority based on over-sampling algorithm for imbalanced data stream with concept drift is proposed. Combining the ensemble learning and resampling method, the proposed algorithm is designed based on chunk-based ensemble and under-sampling methods. The previous minority samples are selectively stored to amply the minority samples in current chunk. The method of update the stored minority samples is designed to ensure the diversity among the base-classifiers and the informativity of the selected samples. In addition, in order to improve the performance of ensemble classifier, dynamic weight is calculated for every base-classifier according to a certain evaluation indicator on every data chunk. The efficiency and classification performance of the proposed algorithm have been confirmed on the basis of extensive experimental research carried out on the all kind of the data streams.