基于动态集成选择算法的信用卡审批异常检测  

An Anomaly Detection Model for Credit Card Approval Based on Dynamic Ensemble Selection

作　　者：程建华[1] 庞梦兰 CHENG Jianhua;PANG Menglan(School of Big Data and Statistics,Anhui University,Hefei 230601,China)

机构地区：[1]安徽大学大数据与统计学院,合肥230601

出　　处：《合肥学院学报（综合版）》2023年第5期86-94,共9页Journal of Hefei University：Comprehensive ED

基　　金：安徽省社会科学基金项目“大数据背景下数据驱动模式的经济运行预警体系研究”(AHSKF2019D019)。

摘　　要：针对数据集标签缺失且类别分布极不平衡的信用卡欺诈检测问题,提出一种基于动态集成选择算法的信用卡审批异常检测模型DES-HBOS(Dynamic Ensemble Selection based on Histogram-based Outlier Score)。首先,利用无监督异常检测算法构造训练集客户的伪标签;然后,确定待测客户能力区域,根据Pearson相关系数评估分类器性能;最后,选择一组较优的分类器对待测客户进行集成。在真实信用卡客户数据集上的实验表明,与其他6种经典异常检测模型相比,DES-HBOS的Recall更高,能将更多欺诈客户识别出来。在4个不平衡数据集上进行对比实验,实验结果表明与HBOS相比,DES-HBOS检测异常能力更强。Aiming at the problem of credit card fraud detection with missing labels and extremely unbal-anced distribution of categories in data sets,this paper proposes a credit card approval anomaly detec-tion model DES-HBOS(dynamic ensemble selection based on histogram based outlier score).Firstly,an unsupervised anomaly detection algorithm is used to construct customer pseudo labels in the train-ing set.Then,the competence regions of the tested customers are determined,and the performance of classifiers is evaluated based on the Pearson correlation coefficient.Finally,a better set of classifiers is selected for ensemble with the tested customers.Experiments on real credit card customer datasets have shown that compared to the other six classic anomaly detection models,DES-HBOS has the high-er recall,can identify more fraudulent customers.Comparative experiments were conducted on four im-balanced datasets,and the experimental results showed that compared with the HBOS,DES-HBOS has stronger ability to detect anomalies.

关 键 词：异常检测 动态集成选择 DES-HBOS 信用卡欺诈识别 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]