Research Online Fraud Detection Model Based On Common Trading Behavior

With the rapid development of online transactions, online fraud has becomeincreasingly common. The form of fraud has changed from the traditional gangs fraudinto a kind of cheat based on network platform. In this kind of fraud, the deceivers whoare more widespread, and wouldn’t come into being a fraud gang obviously. Thus, thecurrent monitoring model for gangs can’t find them accurately. This paper has putforward a monitoring model against those deceivers by extract characteristics of usercommon transaction behavior, and combining the social network analysis and usercharacteristics. The details are as follows:Firstly, through the analysis and research of user activity in the platform fraudulent,proposed the Reverse Graph and Cumulative Trading of Alliance, and get thecharacteristics of user common transaction behavior. Then, put forward a reasonablemethod to measure the property value based on the analysis of the above characteristics,and designed parallel algorithms to calculate those values.Secondly, design and select some important attributes of users’ transaction diagram,which reflects the integrity and tightness of the users. These attributes are gainedthrough the social network analysis of the transaction diagram. Considering the massiveusers and transactions in the transaction diagram, the paper also adopts parallelalgorithm in the calculation of the attributes. Besides, has put forward the importantattributes at the user level and analyzed these attributes. Limited by the platform and thedatasets, unable to get all the user level attributes in this paper. Combined commontransaction behavior features with user and graph level features as the set of propertiesin this paper.Finally, designed a reasonable fraud detection model, and selecting the optimaldatasets based on time characteristics. Directing at the problem of unbalanced categoryclassification and the feasibility of parallel algorithm, the paper ultimately choosesrandom forest as the classification algorithm in fraud detection model. By comparingthe experiments, it shows that the selection of optimal datasets based on time and theattributes of user common transaction behavior can enhance the detection performanceof the whole system. For the selection of classification algorithm in the model, theexperiment indicates that the use of random forest can obtain relatively betterperformance. Meanwhile, comparing this model with other models by experiments, it can be used to detect the fraudsters in the network platform as well as those unbalancedcategory classification in real transactions. At the last, on the shortcoming of the modelare described and proposed possible solutions.