Research On Review Spam Detection Based On Hierarchical Neural Network And Multivariate Features

In recent years,more and more consumers tend to share their consumption experience and refer to other reviews when making consumption decision on the Internet.However,some malicious merchants forge a large number of reviews in order to exaggerate their own brand or slander competitors.The review spam will not only influence the consumption decisions of users,but also lead to the adverse development of the business competitive environment.In order to effectively identify fake reviews,in this paper,rich linguistic and non-linguistic features are extracted from the review data,and the spam review recognition is carried out using MFNN(Neural Networks inverse Multivariate Features),an identification model based on hierarchical neural networks and multivariate features.The research contents of this paper are summarized as the following three points:(1)Construct multivariate features.In this paper,the potential clues of fake review detection are divided into linguistic feature information and non-linguistic feature information.And a set of multivariate feature set is extracted by using review text and other data information,which can be used to carry out fake review detection experiments based on different datasets.(2)Feature selection.The random forest and sequential backward selection strategy were used to trim the features according to their importance to the performance of the model,and the feature subset with the best recognition performance was selected to carry out experiments,so as to reduce noise interference and improve the simplicity of the model.(3)Review spam detection model is constructed based on neural network.This paper proposes a hierarchical neural network model that incorporates the attention mechanism to detect fake review by mining deep semantic information.At the same time,the convolution neural network is used to extract the local and global feature information through the multivariate discrete features,and finally the semantic feature vector is connected with the multivariate discrete feature vector to act on the spam review detection model.Experimental results show that the model proposed in this paper on multiple datasets is superior to the traditional discrete model and the existing neural network benchmark model,and has a good generalization ability.