Research On Review Spammer Detection Based On Review Graph

As an important part of opinion mining, the identification on view spammer detection has attracted lots of attention among domestic and overseas, and a lot of researches have achieved. However, the existing works ignored two features of spammer detection: the impact on reviews by the replies of credible respondents and the relations within the reviews data. These two features can have an impact on the recognition of the view spammer and ignoring them will reduce the identification accuracy. Therefore, we will focus on the identification of the product review spammer in this paper and the details of the research are proposed as follows:1) For the impact on reviews by the replies of credible respondents, we propose a new direction method as combing with characteristics of spammer. Firstly, we consider the relationship among reviewers, reviews, product, stores and respondents, and construct the review graph. Furthermore, we build up reviewer trustiness model, review honesty model and store reliability model. Finally, with the help of review trustiness by using the iterative algorithm, we test review spammer. As a huge advantage, this method divides the respondents into trusted and untrusted to avoid improve the influence degree of a review by spammer using the response information, and the response of untrusted respondents is regard as invalid. The experiment results show that the accuracy of this method has been improved by 4%, compared with the previous method without consideration of the respondents’trustiness2) Based on the internal relations of review data, we calculate store reliability with review honesty. The algorithm not only take the number of reviewer into account, but also combines with review honesty of each one, and thus avoid some bias of some reviewers with some stores. The experiment results show that the accuracy of this method has been improved by 3%, compared with the previous method without consideration of review honesty.3) For the inaccurate description of stores reliability, we optimize the model of store reliability by setting coefficient of proportional. For a store which has high proportion of trusted reviewers who write positive reviews, it’s reliability will be high. Conversely, the reliability of the store will be low if the store gets low proportion of trusted reviewers. The experiment results show that the accuracy of this method has been improved by 1%, compared with baseline method. Besides the accuracy of all features method has been improved by 8%.Combining the results of all, we design a prototype system based on review graph identify product review spammer. And we also brief each function module of the system. The system effectively used the relationship among reviewers, reviews, product, stores and respondents, and we can get even higher accuracy.