Feature Space Oversampling Technique for Imbalanced Classification

The classification problem with imbalanced data is very common in real world. With traditional classification methods, it is generally difficult to obtain satisfactory classification results. Oversampling provides a feasible solution to this kind of classification problems. Existing oversampling methods generally choose borderline minority samples to generate new samples. It would result in too many synthetic minority class samples are in the boundary region such that the original boundary between different classes is changed. To deal with this issue, a feature space oversampling technique (FSOTE) is presented in this study. By the FSOTE algorithm, the minority class clusters are indeed found from the feature space, and the synthetic minority class samples are uniformly filled in the interior of these clusters. Tested on some widely adopted imbalance data sets, it confirms that the classification accuracy is effectively improved by the proposed FSOTE than by some previous methods.