基于共同交易行为的在线交易欺诈检测模型研究

发布时间：2018-07-13 17:21

【摘要】：随着在线交易的快速发展，在线交易欺诈已经越来越普遍，欺诈方式从传统的团伙欺诈发展成平台欺诈。在平台欺诈方式中，由于欺诈者广泛的分布在网络各地，他们之间并不会形成明显的欺诈团伙，所以目前流行的团伙欺诈检测模型并不能够很好的发现平台欺诈者。本文主要提取用户共同交易行为特征属性，，并结合社交网络分析和用户特征属性，提出了针对平台欺诈方式的检测模型。具体研究内容如下：首先，通过对平台欺诈中用户交易行为的分析与研究，提出了反转图和同盟对累积交易数，得到用户有关共同交易行为的特征属性。对上述特征属性进行分析，提出合理的属性值度量方法，并且设计并行算法用于计算特征属性值。然后，本文选取和设计了一些用户图级别的重要特征属性，反映用户在交易图中的诚信度和紧密度。通过对交易图进行社交网络分析来获得这些特征属性，考虑到交易图中海量的用户和交易，同样对特征属性值的计算设计了并行算法。文中还提出了一些用户级别的重要特征属性，并对这些特征属性进行了分析。受平台和数据集的限制，本文无法获取全部的用户级特征属性。最终的特征属性集包含了共同交易行为特征属性、图级别特征属性和用户级别特征属性。最后，本文设计了合理的欺诈检测模型，基于时间特性选取最优的数据集。针对类别不平衡分类问题和算法并行可行性问题，最终选择随机森林作为欺诈检测模型的分类算法。通过对比实验说明了选择基于时间特性的最优数据集和用户共同交易行为的特征属性能够提高检测性能，使用随机森林作为分类算法能够取得相对较优的性能。同时，通过实验将本文提出的检测模型与其它模型进行了对比，本文提出的模型能够用于对平台欺诈用户的检测，同时能够适用于真实交易平台中类别不平衡分类问题。本文最后对模型的缺点进行了说明，并提出了可行的解决方案。
[Abstract]:With the rapid development of online transactions, online transaction fraud has become more and more common, fraud methods have developed from traditional Gang fraud to platform fraud. In the way of fraud, the fraudsters are widely distributed across the network, and they do not form obvious fraud groups, so the popular fraud detection model is now popular. It is not good to find the platform frauds. This paper mainly extracts the characteristics of the user's common transaction behavior, and combines the social network analysis and the user characteristic attributes to propose a detection model for the platform fraud mode. The specific research content is as follows:
First, through the analysis and study of the user transaction behavior in the platform fraud, the reverse graph and the alliance against the cumulative transaction number are proposed, and the characteristic attributes of the user's common transaction behavior are obtained. The characteristics of the above attributes are analyzed, the reasonable attribute value measurement method is put forward, and the parallel algorithm is designed to calculate the characteristic attribute values.
Then, this paper selects and designs some important attribute attributes of the user diagram level, reflecting the integrity and tightness of the user in the transaction diagram. Through social network analysis of the transaction graph, these characteristics are obtained. Considering the mass users and transactions in the transaction diagram, the parallel algorithm is designed for the calculation of the characteristic attribute values. In this paper, some important attribute attributes of user level are also proposed, and the characteristics are analyzed. By the restriction of the platform and data sets, this paper can not obtain all the user level feature attributes. The final feature set contains the characteristic attributes of the common transaction behavior, the attribute attributes of the graph level and the attribute attribute of the user level.
Finally, this paper designs a reasonable fraud detection model and selects the optimal data set based on the time characteristics. Aiming at the classification problem and the parallel feasibility of the algorithm, the random forest is selected as the classification algorithm of the fraud detection model. The optimal data set and the user are selected by the comparison experiment. The characteristic properties of the common transaction behavior can improve the detection performance, and use the random forest as the classification algorithm to achieve relatively superior performance. At the same time, the test model proposed in this paper is compared with other models by experiments. The proposed model can be used for the detection of the flat table fraud users, and can be applied to the reality at the same time. Finally, the shortcomings of the model are explained and feasible solutions are proposed.
【学位授予单位】：重庆大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP393.08;D924.35

【参考文献】