Location bagging-based undersampling for imbalanced classification problems
Tongwen Rong, Xing Tian, Wing W. Y. Ng · 2016
Random-based UnderSampling (RUS) methods for imbalanced pattern classification problems suffer from high variance problems. Therefore, the Inverse RUS (IRUS) is proposed to relieve this problem using an ensemble of classifier with bagging undersampling on majority samples to create an inverse imbalanced dataset. However, both the IRUS and the RUS do not consider the distribution of dataset. So, in this work, we propose the Location Bagging-based UnderSampling (LBUS) to divide the input space using the ITQ hashing method and undersampling according to location of hash buckets. In this way, the LBUS enjoys benefits of fast random undersampling and distribution information preservation. Experimental results show that the LBUS outperforms the IRUS.