A novel synthetic minority oversampling technique based on relative and absolute densities for imbalanced classification

Learning a classifier from class-imbalance data is an important challenge. Among the existing solutions, SMOTE has received great praise and features an extensive range of practical applications. However, SMOTE and its extensions usually degrade due to noise generation and within-class imbalances. Although multiple variations of SMOTE are developed, few of them can solve the above problems at the same time. Besides, many improvements of SMOTE are based on advanced models with introducing external parameters. To solve imbalances between and within classes while overcoming noise generation, a novel synthetic minority oversampling technique based on relative and absolute densities is proposed. First, a novel noise filter based on relative density is proposed to remove noise and smooth class boundary. Second, sparsity and boundary weights are proposed and calculated by relative and absolute densities, respectively. Third, normalized weights based on absolute and sparse weights are proposed to generate more synthetic minority class samples in the class boundary and sparse regions. The main advantages of the proposed algorithm are that: (a) It can effectively avoid noise generation while removing noise and smoothing class the boundary in original data. (b) It generates more synthetic samples in class boundaries and sparse regions; (c) No additional parameters are introduced. Intensive experiments prove that SMOTE-RD outperforms 7 popular oversampling methods in average AUC, average F-measure and average G-mean on real data sets with the acceptable time cost.