基于图的链接发现在反洗钱中的应用研究-智能与分布计算实验室

基于图的链接发现在反洗钱中的应用研究

姓名	罗汉果
论文答辩日期	2007.01.26
论文提交日期	2007.02.01
论文级别	硕士
中文题名	基于图的链接发现在反洗钱中的应用研究
英文题名	An Application Research on Graph Based Link Discovery in Anti-Money Laundering
导师1	李玉华
导师2
中文关键词	链接发现;频繁子图;正向集;负向集;洗钱链路
英文关键词	Link discovery;Frequent subgraph;Positive set;Negative set;Money-laundering flow
中文文摘	金融犯罪一般隐藏于大量的正常账户交易当中，而目前洗钱犯罪趋于集团化规模化，犯罪分子不再通过少量的账户来进行洗钱交易，而是通过一些账户群体来掩盖其犯罪模式，现有的一些金融监管技术和手段对许多集团性的异常资金流动方式还无法进行有效的监管。针对这一问题，采用基于图的链接发现技术，利用频繁子图对资金交易图进行压缩，可以有效地侦测群体的异常资金交易结构。根据社会网络分析理论，异常行为总是隐藏在大量的正常行为模式中。对应于金融监管领域，就是剔除交易中的正常结构，从而凸现异常结构，即是一个从资金交易图裁剪频繁子图的过程。因此频繁子图发现算法，是基于图的链接发现技术的基础。在Apriori思想的基础上，研究了一种频繁子图的发现算法，该算法从单个节点出发，通过迭代来生成一个k阶频繁子图候选集。在频繁子图候选集的基础上，研究了两种异常交易结构的发现方法：无监督模式方法和监督模式方法。无监督模式方法在对源数据进行预处理后，构建资金交易图，通过采取一种度量原则对频繁子图候选集进行度量，得到一个最佳候选子图，利用该最佳候选子图对资金交易图进行压缩，即将资金交易图中该子图所对应的实例压缩为一个节点，生成一个新的交易图。迭代此过程直到最后的异常交易结构，即洗钱链路的形成。通过正向集和负向集对领域知识进行描述，监督模式方法则引入了金融监管的领域知识对其进行指导。监督模式也是通过多次迭代压缩逐步形成洗钱链路，相对于无监督模式，监督模式对频繁子图的度量基于该子图对正向集和负向集的压缩程度，两个集合的压缩也参与整个迭代过程。根据上述方法，设计了一个原型系统以说明方法的有效性，并对结果可视化技术进行了研究。
英文文摘	Financial crimes always hide among large number of exchanges between normal accounts. Nowadays money laundering crimes trend to be collectivized, criminals not only use several accounts to transfer black money, but also use account groups to conceal their criminal pattern. The existing technologies and approaches in financial supervising can‘t reveal these exceptional fund flow efficiently. Aiming at this problem, graph based link discovery technology, which compresses the fund exchange graph by frequent subgraphs, can detect the criminal pattern of groups with illegal fund exchanges. According to social network analysis theory, a large number of abnormal behaviors are always hidden in the normal mode of behavior. Corresponding to the field of financial supervising is to remove the normal structure of the transaction, which highlights unusual structure. It’s a proceeding of cutting frequent subgraphs from the fund exchange graph. So the approach of graph based link analysis is based on frequent subgraph searching algorithm. In the basis of Apriori idea, this paper presents a frequent subgraph finding algorithm. The algorithm begins from a single vertex, generate a set of candidates in k ranks through iteration. When candidates of frequent subgraphs generated, there are two kinds of illicit transactions discovering approaches: unsupervised mode and method of supervised mode. After source data preprocessing and financial transaction graph building, unsupervised method adopt a measure to calculate the value of frequent subgraphs in candidates. The best candidate is the one which compress the graph to the least size. The method of compressing is replacing the instances of the best subgraph in the graph by a vertex. Finally the unusual structure of fund transactions, called money-laundering flow, formation after repeating this proceeding. Supervised method let the domain knowledge in the field of financial supervising, which is described as positive set and negative set, induct the compress proceeding. Supervised method also formations the money-laundering flow by iteration. Compared with unsupervised mode, supervised method measures the candidates of frequent subgraphs by their compression level to positive set and negative set. And the two set also take part in the iteration of compressing. According to these approaches, a application system prototype is designed to prove the validity of the approaches，and result visualization is also studied in the system.