基于梯度惩罚生成对抗网络的过采样算法

Oversampling algorithm based on gradient penalty generative adversarial network

在不平衡数据分类问题中,为了更注重学习原始样本的概率密度分布,提出基于梯度惩罚生成对抗网络的过采样算法(OGPG)。该算法首先引入生成对抗网络(GAN),有效地学习原始数据的概率分布;其次,采用梯度惩罚对判别器输入项的梯度二范数进行约束,降低了GAN易出现的过拟合和梯度消失,合理地生成新样本。实验部分,在14个公开数据集上运用k近邻和决策树分类器对比其他过采样算法,在评价指标上均有显著提升,并利用Wilcoxon符号秩检验验证了该算法与对比算法在统计学上的差异。结果表明该算法具有良好的有效性和通用性。

In order to pay more attention to learning for probability density distribution of original samples in imbalanced data classification problem,an oversampling algorithm based on the gradient penalty generation adversarial network(OGPG)was proposed.Firstly,generation adversarial network(GAN)was adopted to effectively learn the probability density distribution of original data.Secondly,the gradient penalty was used to constrain the gradient two-norm of the input term of discriminator,which reduced the overfitting and gradient disappearance that appeared easily in GAN,so that the new samples were reasonably generated.In the experiment,the k-nearest neighbor and decision tree classifiers were adopted to compare the other oversampling algorithms,the evaluation indicators were significantly improved.The Wilcoxon signed-rank test was used to verify the statistical difference between this algorithm and the comparison algorithm.The results show that this algorithm has good effectiveness and generality.