A Parallel Oversampling Algorithm Based on NRSBoundary-SMOTE

Abstract

Classification with imbalanced class distribution data set has always been regarded as a difficult problem in knowledge discovery, and re-sampling is an effective way to deal with imbalanced data set. SMOTE is a widely used over-sampling algorithm, however it does not consider the distribution of the original data while generates new synthetic samples. NRSBoundary-SMOTE is based on Neighborhood RoughSet Theory. It only chooses the minority class samples, which belong to the boundary region, to generate synthetic samples, and it can improve the accuracy of the minority class effectively. But it needs to compute the distance of any two samples when parting the data set. That process will take a long time on large data sets. So we propose a parallel over-sampling method based on NRSBoundary-SMOTE, Parallel-NRSBoundary-SMOTE. And we apply our method by mapreduce programming paradigm. The experimental results that running on hadoop clusters shows that our method can maintain the accuracy of minority class and improve the efficiency on large data sets.