Improvement of Bagging performance for classification of imbalanced datasets using evolutionary multi-objective optimization

Today, classification of imbalanced datasets, in which the samples belonging to one class is more than the samples pertaining to other classes, has been paid much attention owing to its vast application in real-world problems. Bagging ensemble method, as one of the most favorite ensemble learning algorithms can provide better performance in solving imbalanced problems when is incorporated with undersampling methods. In Bagging method, diversity of classifiers, performance of classifiers, appropriate number of bags (classifiers) and balanced training sets to train the classifiers are important factors in successfulness of Bagging so as to deal with imbalanced problems. In this paper, through inspiring of evolutionary undersampling (the new undersampling method for seeking the subsets of majority class samples) and taking the mentioned factors into account, i.e., diversity, performance of classifiers, number of classifiers and balanced training set, a multi-objective optimization undersampling is proposed. The proposed method uses multi-objective evolutionary to produce set of diverse, well-performing and (near) balanced bags. Accordingly, the proposed method provides the possibility of generating diverse and well-performing classifiers and determining the number of classifiers in Bagging algorithm. Moreover, two different strategies are employed in the proposed method so as to improve the diversity. In order to confirm the proposed method's efficiency, its performance is measured over 33 imbalanced datasets using AUC and then compared with 6 well-known ensemble learning algorithms. Investigating the obtained results of such comparisons using non-parametric statistical analysis demonstrate the dominancy of the proposed method compared to other employed techniques, as well.