Noise-Robust Gaussian Distribution Based Imbalanced Oversampling

Imbalanced data classification has become one of the hot topics in the field of data mining and machine learning. Oversampling is one of the mainstream methods to solve the imbalance problem by synthesizing new samples to balance the data distribution. However, due to the limited sample local information, the data synthetic process is risky in deteriorating the class overlap phenomenon, showing a vulnerable robustness with respect to data noise. In this paper, we propose a noise robust gaussian distribution based imbalanced oversampling (NGOS). NGOS first determines the neighborhood radius based on the global information, and then assigns sampling weights to minority class samples based on the density and the distance information within each of the neighborhoods. Finally, NGOS generates new samples with a Gaussian distribution model. We validate the effectiveness of our proposed method on the 38 KEEL datasets, DT classifier and eleven comparison methods. Experimental results show that our method outperforms the other compared methods in terms of Fmeasure, AUC, Gmean. The codes of NGOS are released in https://github.com/ytyancp/NGOS.