| 姓名 |
靖兆丰 |
| 论文答辩日期 |
2007.06.05 |
| 论文提交日期 |
2008.06.16 |
| 论文级别 |
硕士 |
| 中文题名 |
动态金融网络中异常交易预测方法研究 |
| 英文题名 |
Research on Prediction Method in Exceptional Exchange of the Dynamic Financial network |
| 导师1 |
李玉华 |
| 导师2 |
|
| 中文关键词 |
Dynamic Financial network;Markov chain;transition probability matrix;link-prediction;predictor |
| 英文关键词 |
动态金融网络;马尔可夫链;转移概率矩阵;链接预测;预测子 |
| 中文文摘 |
金融犯罪具有隐蔽化、智能化、专业化等特点，特别是在洗钱活动中表现得尤为突出。目前大多数的研究方法，采用的是传统知识发现方法，首先对数据集进行调查，形成感兴趣的假设，然后再设计分析方法来解决这些问题。这样的方法能够发现可疑交易的节点或路径，但是仍然有一个很大的不利之处即我们使用的是静态的方法，在动态变化的金融网络中，不能确定未来的交易情况，以便提前预防这些非法交易，针对这一问题，可以采用马尔可夫链理论和链接预测技术设计一套预测方法来解决。根据反洗钱领域的实际情况和许多专家对现阶段洗钱案例的总结，着重关注了以下五类节点对象类型：账号、交易金额、账户所属组织、交易时间、交易地点，通过这五类节点和它们彼此之间的联系设计一种预测方法，整个预测流程分为两个阶段：交易节点状态预测阶段和交易路径链接预测阶段。交易节点状态预测阶段，它基于的思想是马尔可夫链理论。马尔可夫链是一个有着广泛应用的随机过程模型，它对一个系统由一种状态转移到另一种状态的现状提出了定量分析。首先从马尔可夫链的基本理论入手，接着讨论了马尔可夫链转移概率的计算方法，最后运用基于绝对分布的马尔可夫链预测方法设计预测算法，并通过实验表明了此预测方法的可行性和实用性。交易路径链接预测阶段，它基于的思想是链接预测理论。在总结了众多链接预测所使用的基于图的拓扑结构的属性的基础上，使用了一种适用于金融网络的基于共同邻居交易次数的新属性，然后通过构建链接预测子设计预测算法，并通过与传统的基于节点邻居的属性作对比实验，说明了此预测方法的可行性和实用性。 |
| 英文文摘 |
Financial crimes always have the attributes of covertness， intelligence and professionalism， which are more serious in money-laundering. At present， most systems take use of traditional knowledge discovery. Firstly， we propose interested suppose by investigating data set; and then， analysis approach will be designed to solve these problems. The approach can always find the abnormal exchange paths and the abnormal exchange nodes， but there is still a big disadvantage that we always make use of the static method and in the dynamic financial networks we can’t predict the situation of the exchange in the future， so it is hard to prevent the illegal trade. For this problem， we make use of Markov chain theory and Link-prediction technology to design a prediction method. According to anti-money-laundering knowledge and experts' summarization， the question for discussion focuses on five node object type as follows： accounts， money， organization， trade time， trade address. We make use of these five node type and relationship to design a prediction method. The flow includes two phases as follow： phase of predicting abnormal exchange node， phase of predicting abnormal exchange path. The phase of predicting abnormal exchange node bases on Markov chain theory. Markov chain is an extensively applied stochastic process model which is to quantitatively analysis a system transferring from one state to another， this paper first introduces the basic theory， after that it talks about the transition probability， finally， it uses the Markov chain prediction method based on absolute distribution to design a prediction method and makes experiments to prove the method is practical and can be applied in this field. The phase of predicting abnormal exchange path bases on Link-prediction theory. This paper summarizes so many attributes based on the topology of the graph which can be used in the link-prediction problem， we use a new attribute based on the exchange times of the common neighbors which is suitable for the exchange networks and build the predictor from the new attribute. Finally we make experiments to prove the new attribute is better than the traditional common neighbors attribute and the method is practical and can be applied in this field. |
