AD-DES: An adaptive dual dynamic ensemble selection for imbalanced data streams

Classification of imbalanced data streams in nonstationary environments poses a significant challenge in machine learning. Ensemble learning has demonstrated efficacy in managing imbalanced data streams and concept drift. However, most existing methods develop different strategies for each aspect, overlooking their relationships and interactions, which hinders the expected performance of these strategies. To address this issue, an adaptive dual dynamic ensemble selection (AD-DES) method was proposed for classifying imbalanced data streams with concept drift. First, an adaptive equalization resampling (AER) strategy was proposed to obtain balanced data chunks, which can help to reduce the risk of overfitting or insufficient sampling caused by excessive data imbalance. Following that, the data chunk obtained by the AER strategy was stored to balance the subsequent ones. Second, a dual dynamic ensemble selection (D-DES) strategy was introduced to perform two rounds of selection in the classifier pool to obtain the optimal ensemble model. Finally, an adaptive drift detector in AD-DES was integrated, which is beneficial for the model to adjust to newly emerging concepts in nonstationary environments. The research findings demonstrated that AD-DES outshines 9 comparison algorithms in terms of classification precision and robustness across 10 synthetic datasets and 5 real-world datasets featuring diverse forms of concept drift.