【作者】 刘宇；

【导师】 金升平；

【作者基本信息】 武汉理工大学 ， 统计学， 2018， 硕士

【摘要】 近年来随着金融业发展,利用商业银行业务进行非法交易的行为也越发猖獗,在大量的交易数据中识别异常交易账号已经成为商业银行维护正常金融秩序的重要工作之一。目前商业银行识别异常交易数据的方法是相关从业人员根据历史业务经验,人工制定识别准则并人工设定识别准则参数的方法。这种识别方法的准确性和有效性容易受到相关人员的业务素质以及往期数据中的小概率事件影响,所以本文采用统计学习的方法减少识别工作中的人为影响因素。决策树模型由于其可读性和分类速度快等优势而成为一种被广泛应用于识别工作中的统计学习方法。本文的主要工作是将决策树模型应用于识别商业银行异常交易账号的工作中,全文包括以下内容:首先,由商业银行交易数据构建交易网络,交易网络中的节点和边分别对应交易数据中的交易账号和交易账号间的交易行为。以交易网络中的节点为单位,在交易网络中提取节点的结构和性质,并对节点的结构和性质进行统计,将节点的结构和性质作为生成决策树模型的特征集,将节点结构和性质的统计数据作为生成决策树的数据集。将非现场监控模式中对交易记录的识别工作转变为对交易账号的识别工作,减少识别工作中的数据量,为生成决策树模型构建特征集和数据集;其次,构建适用于本文研究背景的噪声数据识别模型,以交易账号为单位,对于交易网络中的每一个交易账号,将其所有交易记录中的噪声数据进行有效的识别并清除,优化生成决策树模型的特征集和数据集;最后,将交易网络中节点的结构和性质作为特征集,并以节点结构和性质的统计数据作为数据集。对目前的C4.5算法进行改进,使连续型特征离散划分过程中对阈值的计算更加精确,在保留大部分数据信息量的前提下减少离散划分的区间数目。以改进的C4.5算法生成识别商业银行异常交易的决策树模型,并将该决策树模型应用于识别异常交易的工作中。更多还原

【Abstract】 In recent years,with the development of the financial industry,the illegal transactions in commercial banks’ business are also more and more serious.Identifying abnormal transaction accounts in a large amount of transaction data has become an important work of commercial banks in maintaining normal financial order.The method for commercial banks to identify abnormal transaction data is that relevant practitioners manually formulate the identification criteria and manually set the identification criteria parameters based on the historical business experience nowadays.The accuracy and effectiveness of such identification method is easily affected by the quality of the relevant practitioners and the small-probability events on the historic business transaction data,so this paper uses the statistical learning method to reduce the artificial factors in the identification work.Decision tree model has been widely used as a statistical learning method in recognition work because of its readability and fast classification speed.The main work of this paper is applying the decision tree model in identifying the abnormal transaction account of commercial banks.This paper includes the followings:Firstly,the transaction network that from the commercial bank transaction data is built.The nodes and edges in the transaction network correspond to the transaction accounts and the transaction activities between transaction accounts in the total transaction data.Taking every node in the transaction network as a unit,the structure and properties of the nodes could be extracted from the transaction network and count the structure and properties of the nodes.Then the structure and properties of the nodes are used as a feature set for generating the decision tree model,and the statistics of nodes’ structure and properties are used as the data set for generating decision trees.The identification of transaction records in off-site monitoring mode could be transformed into the identification of transaction accounts,then the amount of data in the job of identification could be reduced.These could build the feature set and the data set for generating the decision tree model.Then,a noise data identification model that could be suitable for the research background of this paper has been built.Taking every transaction account as a unit,for each transaction account in the transaction network,all the noise data in the transaction records can be effectively identified and cleared.These can optimize the feature set and the data set for generating the decision tree model.Finally,the structure and properties of nodes in the transaction network are used as the feature set,and the statistical data of nodes’ structure and properties are used as the data set.The improving in the current C4.5 algorithm can make the calculation of thresholds in discrete process of continuity characteristics more accurately and reduce the number of discrete divisions by preserving the most data.A decision tree model for identifying abnormal transactions in commercial banks can be generated by the improved C4.5 algorithm.The decision tree model can be applied in the work of identifying abnormal transactions.更多还原