The Reseach Of Online Fraudulent Transactions Recognition Based On Deep Learning

With the rapid development of China's e-commerce,more and more merchants begin to sell products in the electronic commerce mode,especially the online retail business,the competition is becoming increasingly fierce.In order to increase the sales volume,merchants promote goods through illicit competition such as fraudulent transaction to attract the attention of consumers.Fraudulent transaction is not only a kind of fraudulent propaganda,but also on the one hand harms consumers' right to know and right of fair trade what has an extremely negative influence on consumers'online shopping experiences,on the other hand it has become a key factor that hinder the healthy development of China'e-commerce seriously.The main research works of this paper are expounded in the following parts:(1)There are two participants involved in fraudulent transaction,which are merchants in charge of releasing fraudulent transaction tasks and fraudulent consumers that excute fraudulent transactions.Merchants acquire fraudulent sales volume,credit scores,reviews and so on through fraudulent transactions,meanwile,fraudulent consumers are hired by the merchants and take their commissions through fraudulent purchasing.Because of the products in fraudulent transaction,merchants and fraudulent customers are connected with each other,consequently,the research is divided into two parts which are the fraudulent transaction recognition based on commodities and the fraudulent transaction recognition based on consumers,meanwhile we use deep belief network to recognize fraudulent transactions.(2)Fraudulent transaction recognition based on commodities aims to recognize the commodities online that the sales volume and the number of reviews are increased by fraudulent transaction.During the process of describing the features of commodities,in addition to considering the merchant attributes,the commodity attributes and the review attributes,commodity sales index is used to analyze the sales record,which is used to describe the changes of the sales volume during a period of time.The model chosen to work with in this thesis is a deep belief network which is coupled with a multi-layered perceptron(MLP).The MLP module and the DBN module construct the fraudulent transaction recognition model,where the features of input dataset are learned by the DBN module that the higher level abstract features can be got,and the MLP module perform the classification task in the recognition model,so as to realize the recognition of fraudulent transaction based on commodities.(3)Fraudulent transaction recognition based on consumers aims to distinguish the fraudulent consumers that post fraudulent comments from the normal consumers.In the process of analyzing the features of consumers,we mainly consider the consumer attributes,the commodity attributes,the review attributes and the merchant attributes.When learning features of consumers by deep belief network,considering the problem that falling into the local minimum point in the learning process of DBNs may result in the absence of the superiority in the aspect of feature learning,alternating supervised learning and unsupervised feedback fine-turning process in DBNs to solve the problem that the performance of deep network is not good enough because the derivative of the weight becomes small in the back-propagation process.Meanwhile,fuzzy sets are used to describe the grade of membership that whether users are fraudulent users or not and will be introduced into.DBNs to improve the accuracy of classification.(4)Finally,the thesis tests and verifies the performance according to the data of commodities and consumers collected from taobao.com,and compares our methods presented in this thesis with both traditional deep belief network and common shallow machine learning(Beysian Network,Random Forset and Suport Vector Machine),in order to verify the performance of the recognition.