Adaptive Resampling and Weighted Ensemble method for Dynamic Imbalance Data Stream Classification

Class imbalance inevitably occurs in dynamic data stream scenarios and can pose tremendous challenges for data stream mining. To address these challenges, an adaptive resampling and weighted ensemble method (ARWE) is proposed in this paper. First, the dynamic subdivision Poisson resampling (DSPR) module in ARWE is developed to address the class imbalance problem in thedata stream. DSPR combines local information from minority class samples with the imbalance rate to design a sample-weighting scheme that can enhance the visibility of minority class samples, particularly those at the boundaries. This approach ensures that the model pays more attention to minority class boundary samples. Second, to address drift, the detection mechanism is designed with a drift warning stage that extends the search for marginal samples when a drift warning occurs and recalculates the weights to cope with simultaneous changes in the data distribution and imbalance rate. Finally, the ensemble update and decision-making are achieved through a co-constructed weighting scheme that utilizes Hellinger distance and accuracy. To compare the proposed method with eight state-of-the-art methods, simulation experiments are conducted on 17 data streams with concept drift and class imbalance. The results demonstrate the superiority of the proposed model in handling dynamically unbalanced data streams.