一种混合的信用卡欺诈检测模型
2021-03-24毛铭泽
毛铭泽



摘要:信用卡欺诈检测是一个重要的问题,为了提升对于真实世界的信用卡欺诈数据的识别率,提出了一种混合的信用卡欺诈检测模型AWFD(Anomaly weight of credit card fraud detection),首先通过异常检测的方法将数据划分为可信和异常数据,然后利用半监督的方法训练一个集成模型,最终再利用异常检测进一步剔除检测结果中的异常结果。AWFD在保障对于可信数据的学习效果上,通过半监督集成学习的方法,利用异常数据进一步扩充集成模型的多样性,并将异常检测和集成模型融合。实验结果表明,比起一些传统的机器学习方法,AWFD可以提高整体的信用卡欺诈检测的识别率。
关键词:信用卡欺诈检测;异常检测;半监督;集成学习;多样性
中图分类号: TP311 文献标识码:A
文章编号:1009-3044(2021)02-0194-03
Abstract:Credit card fraud detection is a serious problem. In order to improve the recognition rate of real-world credit card fraud data, a hybrid credit card fraud detection model AWFD (Anomaly weight of credit card fraud detection) is proposed. Firstly, the data is divided into trusted and abnormal data by anomaly detection method, and then an ensemble model is trained by semi- supervised method. Finally, anomaly detection is used to further eliminate the abnormal results in the detection results. On the basis of guaranteeing the learning effect of trusted data, AWFD uses the abnormal data to further expand the diversity of the ensemble model by semi-supervised ensemble learning method, and integrates the anomaly detection and the ensemble model. The experimental results show that AWFD can improve the overall recognition rate of credit card fraud detection compared with some traditional machine learning methods.
Key words:credit card fraud detection; anomaly detection; semi-supervised; ensemble learning; diversity
由于社会的不断发展,人的消费需求也在不断上升,随之而来的支付手段的改变,导致信用卡欺诈问题日趋严重,每年会导致数十亿美元的损失[1],因此对于信用卡欺诈检测问题的研究也受到了广泛的关注[5-6],有许多学者从不同的角度研究这一问题,例如:有从数据不均衡角度研究采样问题的[2],有从模型融合角度研究的[1],也有从特征工程角度研究该问题的[3-10]。使用的算法也是多种多样,例如:随机森林Random Forest(RF)[4],Adaboost(ADB)[1]等。
除了以上这些常见问题之外,实际取得的标注训练数据中,可能会存在异常噪音数据,也就是指,在标注为欺诈样本的数据中,会存在一些实际不是欺诈的样本或是不处于数据空间分布密度……
