Gaussian prior based adaptive synthetic sampling with non-linear sample space for imbalanced learning

In the presence of skewed category distribution, most learning algorithms fail to provide favorable performance on the representation about data characteristics. Thus learning from imbalanced data is a crucial challenge in the field of data engineering and knowledge discovery. In this work, we proposed an imbalanced learning method to generate minority samples for the compensation of class distribution skews. Different from existing synthetic over-sampling techniques, the data generation is conducted within the hyperplane rather than on the hyperline, thus the proposed method breaks down the ties imposed by the linear interpolation. In addition, this proposed method minimizes the sampling uncertain and risk by integrating a prior knowledge about the minority class instances. Moreover, a multi-objective optimization combined with error bound model develops this proposed method into an adaptive imbalanced learning. Extensive experiments have been performed on imbalanced issues, and the experimental results demonstrate that this method can improve the performance of different classification algorithms.