基于VAE-GAN算法的信用卡欺诈检测模型

Credit Card Fraud Detection Model Based on VAE-GAN Algorithm

信用卡欺诈检测数据集是典型的离群点分布极度不平衡的高维数据集,信用卡交易中被盗刷的交易占比非常小,但每一笔被盗刷的交易都影响重大。针对传统离群点检测算法难以学习到极度不平衡的高维数据集中离群点的分布模式,导致检测率低的问题,本文应用一种基于变分自编码器(Variational Auto-Encoder,VAE)和生成对抗网络(Generative Adversarial Network,GAN)相结合的VAE-GAN算法进行无监督学习,算法首先将数据集输入VAE型生成器中进行训练,生成大量潜在的离群点,然后令判别器学习正常点与离群点的分类边界,最后将测试数据输入训练后的模型中,将离群值高的测试数据判定为离群点。在信用卡欺诈检测数据集上与现有的无监督学习所得结果相比,VAE-GAN在尽可能更多地检测出离群值的同时,尽量减少误判,AUC达到0.9581,Recall达到0.9118,ACC为0.9468,优于目前的最优模型,证明VAE-GAN算法在信用卡欺诈检测中的优越性。

Dataset of credit card fraud detection is a typical high-dimensional dataset with excessively unbalanced outlier distribution,i.e.,percentage of fraud in total credit card transaction is low,but each fraud causes enormous implications.Traditional outlier detection algorithms have difficulty in comprehending the outlier distribution in extremely unbalanced high-dimensional dataset,resulting in low detection rate.To address the issue,a VAE-GAN algorithm for unsupervised learning based on combining the Variational Auto-Encoder(VAE)and the Generative Adversarial Network(GAN)is proposed in this paper.In the VAE-GAN,dataset is first inputted into a VAE-type generator for training to generate plenty of potential outliers.Then,discriminator is trained to learn the classification boundary between the inlier and the outlier.Finally,test data are inputted into the trained model to determine those with high outlier values as the outliers.Compared with existing unsupervised learning on credit card fraud detection dataset,the VAE-GAN detects as many outliers as possible while minimizing the false positives,with an AUC of 0.9581,Recall of 0.9118,and ACC of 0.9468,outperforming the current optimal model,which indicates that the VAE-GAN algorithm is superior in credit card fraud detection.