A Novel Under-sampling Method with Gaussian Mixture and Jensen-Shannon Divergence
Xueling Pan, Guohe Li, Shunxin Liu, Qiuyue Yu, Ying Li · 2022
Due to bias towards major class samples, the traditional classifier cannot obtain excellent performance in imbalanced data. The under-sampling approach is one of the effective methods in imbalance classification. Nowadays, most of the under-sampling methods do not consider the global and local distribution of samples and the intra-class imbalance problem simultaneously. In addition, the sampling quantity cannot be adjusted adaptively in the practical application. To handle the above problems, in this paper, we propose a novel under-sampling approach with the Gaussian mixture model and Jensen-Shannon divergence, called GD-US. Firstly, it utilizes a Gaussian mixture model to cluster the majority class samples. And then, it calculates the number of sampling in each cluster according to its sparsity. In addition, GD-US adopts the global and local distribution of samples to decide the probability of deleting samples. Eventually, the Jensen-Shannon divergence is used to control the number of sampling. The experimental results show that the proposal can help solve the imbalance problem and enhance the classification accuracy.