Research On Online Detection Method Of Reputation Fraud Campaign Based On Conditional Random Fields

Under the influence of economic environment,some businesses who are driven by interests make fake reviews on their products by hiring a large number of people to manipulate own user accounts,so as to deceive users.At present,the problem of fake reviews is still a challenge.Aiming at the problem,this paper proposes two methods,unsupervised learning and supervised learning,to further mine the fake reviews and related fraudulent activities in the review data.The paper proposes a framework that fraud reviews online detection based on markov random fields in the unsupervised learning-FraudMRF.This model detects fake reviews(reviewers)in online mode.To represent the relationship between reviewers and products,it defines a lightweight bipartite graph which can be real-time built.FraudMRF can catch fake reviewers in time by propagation of network of the temporary review and incorporation of real-time prior.Another method,this paper proposes an online mode framework that mine abnormal review sequence is based on dynamic programming and then label it using skip-chain CRFs in supervised learning-FraudGuard.The algorithm tracks the data stream of online review of each product in real-time.FraudGuard improves dynamic programing-the problem of maximum sum of subsequence,and find abnormal review subsequence recursively by online mode.Then,FraudGuard computes the temporary features of the abnormal review subsequence that reaches its deadline,and send it to CRFs.Finally,the framework labels the abnormal review subsequences and mines fraudulent activities.As the final experimental results shown,compared with the offline algorithm of the original model,FraudMRF improves the average precision of top 2000 reviews by 14-38%.Also,the F1 score of FraudGuard improves almost 10% than others classifier of without network structure.Last but not least,it is more specific than the previous similar framework on detection target.