A dual evolutionary bagging for class imbalance learning

Bagging, as a commonly-used class imbalance learning method, combines resampling techniques with ensemble learning to provide a strong classifier with high generalization for a skewed dataset. However, integrating different numbers of base classifiers may obtain the same classification performance, called multi-modality. To seek the most compact ensemble structure with the highest accuracy, a dual evolutionary bagging framework composed of inner and outer ensemble models is proposed. In inner ensemble model, three sub-classifiers are built by SVM, MLP and DT, respectively, with the purpose of enhancing the diversity among them. For each sub-dataset, a classifier with the best performance is selected as a base classifier of outer ensemble model. Following that, all optimal combinations of base classifiers is found by a multi-modal genetic algorithm with a niche strategy in terms of their average G-mean. A combination that aggregates the smallest number of base classifiers by the weighted sum forms the final ensemble structure. Experimental results on 40 KEEL benchmark datasets and a practical one of coal burst show that dual ensemble framework proposed in the paper provides the simplest ensemble structure with the best classification accuracy for imbalance datasets and outperforms the state-of-the-art ensemble learning methods.