Importance-SMOTE: a synthetic minority oversampling method for noisy imbalanced data

Synthetic minority oversampling methods have been proven to be an efficient solution for tackling imbalanced data classification issues. Different strategies have been proposed for generating synthetic minority samples. However, noisy samples which may cause the overlapping of minority and majority classes have not yet been properly treated for reducing their influence on the performance of a classification model. A new method, named Importance-SMOTE, is proposed in this paper. In this method, only borderline and edge samples in minority class are oversampled. The synthetic minority samples are generated proportionally to the importance of the minority samples which is calculated according to the composition and distribution of its nearest neighbors. The positions of the synthetic minority samples are determined by the relative importance of the paired neighbors. The proposed method is expected to obtain a more precise estimation of the true decision surface and reduce the influence of noisy samples. Various public imbalanced datasets and a real case study are considered in the experiments to prove the effectiveness of the proposed method.