摘要： 针对目前金融支付行业交易中存在的欺诈风险复杂化和高频化的问题,提出一种基于密度分布演化的流式聚类算法(DDE-Stream)。利用CLIQUE算法对账户交易特征进行网格聚类,结合隐马尔可夫算法构建账户交易行为档案模型,根据该模型进行实时的欺诈侦测,并在模型自更新阶段,利用DDE-Stream算法对多维度交易特征进行实时聚类。实验结果表明,该算法能够实时侦测交易欺诈风险,且在验证集上获得的准确率相比传统随机森林分类算法超过50%。

关键词: 实时风控, 欺诈侦测, 行为档案, 流式聚类, 增量隐马尔可夫

Abstract: Aiming at the complexity and high frequency of fraud risk existing in the current financial payment industry transactions,a new method called the Density Distribution Evolution-based Stream(DDE-Stream) clustering algorithm is proposed.The CLIQUE algorithm is used to cluster the trading features of the account,and an account trading behavior archive model is constructed by combining with hidden Markov algorithm.According to the model,real-time fraud detection can be performed.In the stage of model self-renewal,DDE-Stream algorithm is used to perform effective real-time clustering of multi-dimensional trading features.Experimental results show that the algorithm can detect the risk of transaction fraud in real time,and the accuracy rate obtained on the verification set exceeds 50%,compared with the traditional random forest classification algorithm.

Key words: real-time wind control, fraud detection, behavior archives, stream clustering, incremental hidden Markov

中图分类号: 

• TP391