An empirical study of learning from imbalanced data

No consistent conclusions have been drawn from existing studies regarding the effectiveness of different approaches to learning from imbalanced data. In this paper we apply bias-variance analysis to study the utility of different strategies for imbalanced learning. We conduct experiments on 15 real-world imbalanced datasets of applying various re-sampling and induction bias adjustment strategies to the standard decision tree, naive bayes and ?-nearest neighbour (?-NN) learning algorithms. Our main findings include: Imbalanced class distribution is primarily a high bias problem, which partly explains why it impedes the performance of many standard learning algorithms. Compared to the re-sampling strategies, adjusting induction bias can more significantly vary the bias and variance components of classification errors. Especially the inverse distance weighting strategy can significantly reduce the variance errors for ?-NN. Based on these findings we offer practical advice on applying the resampling and induction bias adjustment strategies to improve imbalanced learning.