GSNNB: A resampling technique of Gaussian Sampling within the Nearest Neighbor Boundary for class imbalance

Class imbalance is a common phenomenon in real world where the distribution of the training data across the classes is biased or skewed. Learning from class-imbalanced data has posed a great challenge in machine learning and thus several approaches were raised to deal with it such as resampling skills, feature selection and ensemble learning. In this study, a novel algorithm called Gaussian Sampling within Nearest Neighbor Boundary (GSNNB) is proposed which is a hybrid sampling technique with two stages. The first stage is to find the nearest neighbor boundary (NNB) between minority class and majority class. In the second stage, to balance the training data, minority samples are generated based on Gaussian sampling within the NNB, while majority samples are under sampled randomly to adaptive number after clustering. To demonstrate the performance of GSNNB, experiments on 40 two-class imbalanced datasets are given and the results show that GSNNB perform better than or comparable with some other existing methods concerning geometric mean (G-mean) and area under the receiver operating curve (ROC).