基于条件生成对抗网络的不平衡学习研究

For the classification of imbalanced data, the imbalance ratio is not the only factor affecting the classification effect. The class overlapping, the separation of positive samples and the noise samples will all have impact on the classification effect. For the imbalanced data with class overlapping, a re-sampling method with the conditional generotive adversarial networks(CGAN) model(RECGAN) is proposed. It not only improves the recognition of positive samples in overlapping regions, but also overcomes the defect of previous sample synthesis based on the local neighborhood of samples. The experimental results show that the RECGAN method has obvious advantages in terms of the values of AUC and F₁ and the average ordering on the dataset.