Resampling algorithm based on sample similarity and variation coefficient

To solve the problem of imbalanced data classification, a resampling algorithm based on sample similarity and variation coefficient is proposed. For minority samples, new samples are synthesized with bias according to the similarity between samples. For majority samples, the boundary region is determined according to their variation coefficients, and the samples in the non-boundary region are deleted randomly. By the above two methods, original data set is reconstructed, and sample scales of two classes become more balanced. The identification ability of minority samples and overall classification accuracy can be improved effectively by scientifically guiding oversampling and undersampling. Finally, the imbalanced data sets were adopted for verification by using the kernel extreme learning machine (KELM) as classifier. Experimental results show that the overall performance of the proposed algorithm is superior to other comparison algorithms in dealing with imbalanced data classification.