基于一类支持向量机的财务数据异常模式识别  

Financial Data Abnormal Pattern Recognition Based on One-Class Support Vector Machines

作　　者：邓森林[1] 陈卫东[1] 

机构地区：[1]国防科学技术大学信息系统与管理学院,湖南长沙410073

出　　处：《信息工程大学学报》2015年第2期251-256,共6页Journal of Information Engineering University

摘　　要：财务监督在反腐倡廉建设中发挥着重要作用,但是财务数据具有非平稳、非线性、信噪比低等特点,且没有专门用于鉴别财务数据异常的训练集。将金融交易数据以周为时间跨度生成时间序列,选择交易总金额、交易离散系数、转账次数作为财务账户数据的特征,利用基于统计学习理论的一类支持向量机(one-class SVMs)实现金融交易数据的无监督分类,从而识别出可疑异常财务数据。采用径向基函数作为一类支持向量机的核函数,运用交叉验证法选择核参数γ和惩罚参数C。仿真数据异常检测实验证明了该方法的有效性,并检测出了实际财务账户数据中的可疑值。Financial supervision plays an important role in the construction of anti-corruption. But the financial data has the feature of being non-stationary, nonlinear, and low signal-to-noise ratio. And no specialized training set is used to identify abnormal financial data. This paper transforms the data of financial transactions to time series by a time span of a week. The total amount of transac- tion, the dispersion coefficient and the times of transaction are chosen as the features of the financial account data. In order to find abnormal values, the financial transaction data is classified with unsu- pervised method using one-class support vector machines （one-class SVMs） which is based on the statistic learning theory. We select radial basis function as the kernel function of one-class SVMs and the cross-validation method is adopted to choose the kernel parameters ~/ and C. Experiments on simulated data show promising results and find the abnormal values in real financial data.

关 键 词：财务数据 账户 异常检测 一类支持向量机 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]