离群检测模型在反洗钱中的应用研究

发布时间：2018-08-24 20:59

【摘要】：金融交易网络是一个非常复杂的图网络,要从交易网络中海量的、复杂的、含有噪声的数据集中挖掘出不同于其它数据对象的小模式,也就是金融交易网络和反洗钱研究里的嫌疑账户的确定,这可以转化为数据挖掘中离群检测的问题。传统的基于聚类的离群检测方法有两个很大的不足之处,一方面,对于类内的数据对象的差异,没有很好的模型进行分析,另一方面,对于类之间的联系,没有合适的发现方法。针对上述出现的问题,可以将聚类、局部离群检测和链接发现三者结合起来,结合相关的理论,给出一种完整离群检测模型方案。对每个交易账户,根据交易次数和交易金额规整原始数据为适合离群挖掘的数据集,建立该数据集时,要注意一些无效数据的剔除,使规整后的数据集更加具有有效性和完整性。采用改进的两步聚类算法对规整后的数据集进行聚类,将具有相似交易特征的账户划分为一类,该聚类算法有很好的聚类效果,也可以发现一些异常数据对象,聚类之后的数据集使之后的模型发现分析更加具有针对性,然后对每个聚类分别建立基于聚类的异常检测模型:对类内的账户差异,采用基于马氏距离的局部离群检测算法计算离群指数并以升序的方式呈现出来,对类之间的交易联系,采取基于图熵的链接发现技术对聚类后的交易网络进行约减,给出交易网络中关键的交易节点。并且通过两组对比实验验证了改进的两步聚类算法的聚类精度以及基于马氏距离的密度离群检测算法的准确率。结合项目金融交易分析系统,依照之前的研究,实现了一个离群检测模型模块,针对真实案件的数据进行有效性验证。
[Abstract]:The financial transaction network is a very complex graph network, it is necessary to mine the small pattern which is different from other data objects from the massive, complex, noisy data set in the transaction network. That is, the identification of suspected accounts in financial transaction networks and anti-money laundering studies, which can be translated into outlier detection in data mining. The traditional outlier detection method based on clustering has two great shortcomings. On the one hand, there is no good model to analyze the difference of data objects in the class; on the other hand, the relationship between the classes is analyzed. There is no suitable method of discovery. In view of the above problems, we can combine clustering, local outlier detection and link discovery, and give a complete outlier detection model. For each transaction account, according to the number of transactions and the amount of transactions, the original data is suitable for outlier mining. When setting up the data set, we should pay attention to the elimination of some invalid data. The structured data set is more effective and complete. The improved two-step clustering algorithm is used to cluster the regular data sets, and the accounts with similar transaction characteristics are divided into a class. The clustering algorithm has a good clustering effect, and some abnormal data objects can also be found. The data set after clustering makes the analysis of model discovery more specific. Then the anomaly detection model based on clustering is established for each cluster. The local outlier detection algorithm based on Markov distance is used to calculate the outlier index and present it in ascending order. The link discovery technology based on graph entropy is used to reduce the transaction network after clustering. The key transaction nodes in the trading network are given. The clustering accuracy of the improved two-step clustering algorithm and the accuracy of the density outlier detection algorithm based on Markov distance are verified by two sets of comparative experiments. Combined with the project financial transaction analysis system, according to the previous research, a outlier detection model module is implemented to verify the validity of the real case data.
【学位授予单位】：华中科技大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP311.13;D924.3

【相似文献】