Study Of Reviewer Spammer Group Detection Based On Graph Clustering

With the development of Internet technology,the Internet has been integrated into every aspect of our lives.Comment system as a new product in this background,which is widely used in all major social network sites,e-commerce platform as well as Forum,pastes and other Internet applications.Especially in e-commerce,because of the virtuality of the network,other users comments on items in the system are customers important references when they purchasing goods.A lot of bad businessmen employing spammer groups exaggerated comments on their goods,or employ spammer groups comment derogatory comments on their rivals for personal reasons.Fake comment problems appear,greatly advancing the research in this area.Many domestic and foreign researchers solving this problem from a different angle,using different methods to study this problem.In this paper,we present a method detecting review spammer groups based on graph clustering.This paper argues that only from the individual commentators to find spammer groups will gets a low accuracy and efficiency.As long as spammer in their review process mimicked normal comments in people's behavior or comment as groups,such spammer are not easy to find.Proposed algorithm is good at finding hidden deep spammers and cooperation spammer groups.This paper presents a similarity formula based on both properties and structures to calculate the similarity between customers.We believe that true spammer groups should be a set within a small group of closely connection with each other,therefore,in the phase of finding the spammer groups we use biconnected to constraint the tightness of the groups.In subsequent phase,we use graph partitioning.For the problems using graph division in graph clustering,we solve them combined with our scenarios.In the final experimental phase,we makes a number of characteristics for artificial evaluation,and use these characteristics to check the results of our algorithm.With the same Data Set,we select 500 groups from the results of our algorithm and 500 groups from the compared one for comparing to improve the accuracy of our algorithm is high.