An Imbalanced Data Classification Method Driven by Boundary Samples-Boundary-Boost

Abstract

Imbalanced data existed widely in various fields. The classification problem of imbalanced data is a hot issue in machine learning. Existing algorithms on imbalanced data classification processed data without putting enough attention on the positive boundary samples which were prone to misclassification, and the cost that the positive samples were misclassified is much larger than the cost that the negative samples were misclassified. Based on this background, this paper presented a new imbalanced data classification approach based on boundary samples named Boundary-Boost. At first, categorize positive sample as boundary sample and security sample. Then, synthesize the new positive samples according to the boundary sample to balance the imbalanced data. At the same time, classified all the samples by using Boosting. Next, deleted the synthesized samples which were misclassified to reduce the influence of synthesized error samples on the classification results. Experimental result shows that the approach of Boundary-Boost has advantages on imbalanced data classification problem.