Monitoring Online Reviews For Group Spamming

Around the world,fake reviews occur frequently on online consumption platforms,which seriously disrupt users' shopping judgment.So fake reviews need to be detected in a timely manner.Since most of the previous studies focused on offline detection of false reviews,this paper proposes a novel online reputation fraud detection model,Spam Guard,which mainly detects emerging group cheating activities in the comment data stream.Specifically,the Spam Guard model is built in three steps.First of all,this paper USES Markov random field model to build and maintain comment similarity graph incresively by establishing the similarity relationship between comments.The innovation of this paper is to create multilateral undirected weighted graph by dividing it into three types,that is to say,comment similarity graph is composed of a collection of three types of sub-graphs.Secondly,the cheating degree of each comment is calculated and updated by LBP iteration of the comment similarity graph and the message passing between nodes.Finally,an unsupervised likelihood estimation method,also known as CUSUM method,is used to detect online the sharp rise and subsequent significant fall of cheating degree in the review time series in a manner of slight error accumulation.A sequence in the middle of these two points is recorded as a group cheating activity.This study is different from the previous offline detection in the detection of false comments.In the online environment,it is unrealistic to extract reliable false comment indicators from a small number of recent comments in operation.Therefore,no prior knowledge of personal comments is required in this study.In addition,as a kind of incremental updating algorithm,through the three different types of similarity relation is built figure,with LBP during iteration,and all of the nodes are involved in LBP static graph in different ways,but only to the new comments and connect with new comments form relationship of degree of node involved in cheating update,this property makes Spam Guard a fast online algorithm.Double verification was conducted in this experiment through Yelp Zip and Amazon data sets respectively.In terms of the processing of Yelp experimental data set,DBSCAN parameters were used to divide the data set into three groups with different degrees of cheating,the optimistic group,and the recall rate reached 88.63%.On the manual analysis of Amazon data set,three classic cheating cases were found,which verified the effectiveness of the experiment.