基于密度的空间数据聚类的正常用户筛选方法  

Filtration of normal users based on density-based spatial clustering of applications with noise

作　　者：徐小凤[1] 刘家芬[1] 郑宇卫 

机构地区：[1]西南财经大学经济信息工程学院,成都611130 [2]同济大学电子与信息工程学院,上海201804

出　　处：《计算机应用》2015年第A01期43-46,共4页journal of Computer Applications

基　　金：国家自然科学基金重大研究计划项目(91218301);国家自然科学基金青年项目(60903201);中央高校基本科研业务专项(JBK140129)

摘　　要：第三方支付需要通过有效的欺诈识别方法去进行风险控制,但通过复杂的案件识别模型对每次交易都执行案件判别会降低正常用户的体验。因此第三方支付希望对于正常用户,能够不通过复杂的案件识别系统而通过一个简单的模型系统直接放行以减少对正常用户的打扰。在样本极不均衡的情况下,针对第三方支付的正常用户识别问题,提出了一种基于DBSCAN算法的过滤方法。该方法首先利用信息值(IV)筛选特征,利用信息增益率对特征进行加权,再利用DBSCAN算法来识别案件的分布特征并排除异常案件,计算所有样本与案件聚类质心的距离来筛选出正常用户。实验表明,在保证漏过案件不超过总案件数5%且筛选出的样本中案件占比不大于0.03%的指标前提下,能直接筛选出比指标下限30%更多的正常用户,可达到42.518%,即接近42%的用户可以不用进行案件识别而直接继续其下一步操作,有效提升了总体交易效率。Third-party payment needs fraud detection method to make risk under control. As transaction fraud is small probability event, detecting each transaction thoroughly will bother many normal users with latency. If we can filter those normal users, third party payment can just approve their transactions to give them better experience. To solve this problem, a normal users filtering method based on Density-Based Spatial Clustering of Applications with Noise （ DBSCAN ） . Firstly information value was used to select features and weight was assigned by information gain rati0． DBSCAN was used to exclude some abnormal cases and study the distribution pattern of fraud transactions. Then normal and suspicious users were distinguished by the Euclidean distance to the centroid. The experiment shows that the filtering method let 42. 518%transactions go with leakage rate under 5% and abnormal cases density no more than 0. 03%, which improves the overall efficiency of transactions greatly.

关 键 词：第三方支付 风险控制 基于密度的空间数据聚类 特征加权 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]