A Synthetic Minority Oversampling Method Based on Local Densities in Low-Dimensional Space for Imbalanced Learning

Imbalanced class distribution is a challenging problem in many real-life classification problems. Existing synthetic oversampling do suffer from the curse of dimensionality because they rely heavily on Euclidean distance. This paper proposed a new method, called Minority Oversampling Technique based on Local Densities in Low-Dimensional Space (or MOT2LD in short). MOT2LD first maps each training sample into a low-dimensional space, and makes clustering of their low-dimensional representations. It then assigns weight to each minority sample as the product of two quantities: local minority density and local majority count, indicating its importance of sampling. The synthetic minority class samples are generated inside some minority cluster. MOT2LD has been evaluated on 15 real-world data sets. The experimental results have shown that our method outperforms some other existing methods including SMOTE, Borderline-SMOTE, ADASYN, and MWMOTE, in terms of G-mean and F-measure.