Fraud Clique Detection Model Based On Cellular Automata

There existsome dishonest users in C2C platform. They increase their credit scorerapidly by generatinga large number of unreal transactions, and thenunder the cover ofhigh credit rating, theycommit fraudulenttransactions with innocent buyers, whichmakeC2C e-commerce suffering a severe credit crisis. Brushing credit score and thencommitting fraud transactionsare the conventional process for fraud cliques, which iscomposed of fraudulent sellers and accomplice buyers.It is a difficult job to identifysuch cliques as members of fraud clique cover each other.There is great significance toreevaluate the credit state of users and pick out the suspect cliques for maintaining a fairtrading environment,relieving the credit crisis and reducing troubles and losses forconsumers.Current researches on online fraudulent detectionare mainly focused on the auction,stock and futures markets, and few fraud detection researches are about e-market withconfirmed price. After carefully analyzingthe literatureon fraud detection, we findoutsome problems in current employed methods.In this paper, we try to explore a new wayto detect fraud cliques by combining characteristic attributes associated with newattributes mining from trading networks.Cellular automata (CA) are capable of expressingthe complexity of a systembymicroscopicindividualself-organization behaviors.Irregularcellular automata (ICA),which areageneralization of the standardCA,cansimulatecomplextransactionnetwork.Learningautomata (LA) canautomaticallyadjust itsstatebased ontheenvironment’sfeedbacks.Theyarecombined to formulateaclassification modelwithexcellent adjustment capacity toclassify users by credit state, which is calledFD_ICLA.Machine learningalgorithm is employed to discoverlocalrulesbased on users’elementaryattributesand transactingstatistical properties. Reinforcement signal isgenerated by local ruleswith its neighborhood and the action selected by embedded LAas input.Based on the reinforcement signal,the embedded LAin the specified CAadjuststhestatethrough thelearning algorithm.The proposed FD_ICLAmodelis a typical“bottom-up” simulation method, which determinesmacro-state of a system byexecuting microscopicreference repeatedly.It is a time-consuming job to handle millions of nodes with single FD_ICLAprocess.We propose the parallel FD_ICLA model to enhance the scalability of the original model based on K-partition of graph algorithm, which is capable of dispersingthe computation pressure to multiple machines, considering the local dependencyofCA.Also we achieve a visual prototype system of fraud cliques detectionbased on Gephiwhich is an open source project, to make the entire detecting process visible.Finally, we conduct groups of comparison experiments ona real transaction data setwhich is collected from Kongfz trading platform to test the detecting effect and timeperformance of the proposed model. The experimental results show that FD_ICLAmodel achieveshigher precision rate in detecting accomplices andact more effectively inmining fraud cliques within complex transaction networks. Also the proposing parallelmodel can perfect the original single-process DF_ICLA model by solving the weaknessof time-consuming.