Detecting E-commerce Review Spammers Based On The Analysis Of Behaviors

With the development of web2.0 technology,those single type portals will be gradually replaced by the open and interactive network.More and more people begin to communicate in the network world,also people can tell each other their idea and viewpoint in different network platform.There is no doubt that all these human activities in internet have become an integrant part of our lives.For instance,in e-commence websites such as Taobao and Dangdang,people can give a mark and write their feelings after purchasing a product.Also,users can raise a question or show their support to each other.In the present e-commence system,products with more positive comments will be showed in the front pages.Meanwhile,negative reviews can not be deleted by the sellers and it will affect consumers' shopping plan.Consumers usually give up their shopping plan after reading many negative reviews about their target product,also consumers prefer to buy those products with positive comments.As a result,online product reviews can significantly affect quantity of sale and more and more online sellers began to hire or act as review spammers to write untruthful product reviews.There are two types of spam behavior,the first one is writing favorable reviews to their own products and the second one is writing negative comments to their competitors.Both two cases have bad effect on consumers' judgement and people can not get satisfied shopping experience.In the complicated network shopping environment,finding a scientific and effective method to detect spammers or spam reviews is necessary and many experts all over the world have tried to solve the problem recently years.This thesis put forward a supervised learning method to detect review spammers based on user's behavior.More precisely,there are three following points:(1)For the phenomenon that large amount of spurious reviews existing in the online e-commerce websites,we analyze review spammer's cheating motivation and summarize their cheating actions.Then we use multiple-unit indexes to detect spammers.(2)For the supervised learning method,we regard reviewer's behavior as evidence and build a D-S evidence theory model.Meanwhile,we use SVM model to detect spammers in single behavior pattern and get mass function value by inputting the unthresholded output of an SVM into the sigmoid function.(3)We crawl some real consumers information in the Amazon China as our data set,also we hire three evaluators to label spammers for our data set.Then we calculate the accuracy rate,recall rate and other index by means of our D-S evidence theory model,also we compare our method with three single user behavior models.