A synthesized sampling approach for improving the prediction of imbalanced classification

Abstract

Imbalanced dataset is an important factor influencing the effect of learning algorithms. Its influence on the classification learner is even more universal. To deal with imbalanced classification problem, sampling strategy is always an efficient method, however some other aspects of this strategy need to be solved. What distribution should be regulated among the classes and within the class? Which sampling strategy, over-sampling or under-sampling, is more acceptable in specific issues? What metric should be used to measure the classification results? In this paper we propose a general rule to select sampling strategy and design a novel metric V-measure, putting more attention to the minority. As for the distribution between the classes our choice of them is based on the standard whether the selected distribution will lead to significant improvement of the evaluation criteria.