Research On Method Of Detecting Fake Reviews Via Dynamic Multimode Network

With the rapid development of Web2.0 technology and e-business,more and more users prefer to post reviews on products and services for sharing their opinions and experiences in the e-business web sites,such as Amazon and Tmall.These user-generated contents are important for both manufactures and consumers,because they contain rich user opinions involving the qualities of products and services.Driven by interests,some unscrupulous merchants hire spammer to post fake reviews to mislead the users by enhancing the ir reputation or diminishing the competitors' credit.Such behaviors not only mislead the user's shopping decisions but also seriously affect the healthy development of e-commerce.Therefore,it brings an urgent demand to detect fake reviews as early as possible for reducing their influence.In recent years,spam detection has become a hot research field.The common approaches to find these spam reviews are analyzing the text similarity or rating pattern.With these common approaches we can easily identify ordinary spammers,but we cannot find the unusual ones who manipulate their behavior to act just like genuine reviewers.The traditional algorithm based on single dimension cannot take into account the potential impact of multiple comment features,which leads to the lack of accuracy.In this paper,we propose a novel detecting method based on dynamic multimode and carried on the thorough research.The main work and innovations are as follows:(1)A novel spam detecting method based on dynamic multimode was proposed.Firstly,we construct a multimode network,which contains comment,reviewer,merchandise and store four kinds of nodes,and we use spectral clustering algorithm to explore the relationship between the four types of nodes.Secondly,we present four fundamental concepts,which are the quality of the merchandise,the honesty of the review,the trustworthiness of the reviewer and the reliability of the store,thus enabling us to identify the spam reviewers more efficiently.Thirdly,we design an iterative computation method to reveal the dynamic interaction between four dimensional networks.Eventually,we find that the multi-view spam detection based on the multimode network can detect more subtle false reviews according to our experiments.(2)An algorithm of detecting fake review based on emotion intensity was proposed.This method mainly uses the Natural Language Processing technology to analyze the sentiment polarity of the text.In this paper,our method includes the following innovations: Firstly,we utilize the domain dictionary to explore the category of the comment text and consider the sentiment polarity by connectives.Then,we simplify the collection and arrangement of the experimental data and puts forward 5 important detecting features according to analyze comment pattern.Finally,5 quantitative features are fused together by a logistic regression model and an effective classification model is established.This is an important step to compute the fidelity of comment in multimode network.(3)We improved a spam detecting method based on user's reputation.Firstly,the sparse user-project matrix is filled with the theory of sparse and low rank matrix completion,and then the model of reputation is constructed.Finally,we choose a more reasonable evaluation criteria and refine the reputation of users with the same group scale and different rating,and then use the Top-k algorithm to regard the lowest K users as spammer.This method is also crucial to compute the reputation of reviewer in multimode network.