LDAMSS: Fast and efficient undersampling method for imbalanced learning

In this article, a novel undersampling method based on linear discriminant analysis (LDA) and Markov selective sampling (MSS) is proposed. This method contains two stages. The first stage is to adjust the position of classification boundary according to the G-mean of LDA classifier for many times. The second stage is to extract the “important” training samples from the current majority class by MSS. We apply the proposed undersampling method to Xgboost and study its learning performance. The experimental results of binary class datasets show that compared to other methods, Xgboost based on LDAMSS (X-LDAMSS) not only has better performance in three metrics (F-measure, G-mean, and AUC), but also has less total time. We also apply X-LDAMSS to multi-classification problem and present some useful discussions.