Ensemble based on Constraint Projection and Under-Sampling for Imbalanced Learning

Ensemble method is widely used in imbalanced learning, and it is well known that diverse and accurate base classifiers are the keys to the success of an ensemble. This paper proposes a novel ensemble method based on constraint projection and under-sampling (ECPUS) for imbalanced learning. Unlike conventional ensemble methods, ECPUS learns an ensemble through the following two steps: (1) constructing pairwise constraints by drawing pairs of examples from the minority and/or majority class set, and learning a projection matrix from the constructed pairwise constraints, (2) projecting original training set into a new data space which is defined by the learned matrix, and under-sampling the projected training set to obtain a balanced distribution for learning a base classifier. For the former step, constraint projection introduces diversity to base classifiers, and also increases the separability between examples involved in different classes. For the latter step, under-sampling improves the performance of base classifiers on minority class examples. Experimental results show that, compared with some other state-of-the-art imbalanced learning methods, ECPUS shows better performance on evaluation measures of recall, g-mean, f-measure and AUC.