Research On Spam Review Detection Method Based On Heterogeneous Graph Neural Networks

With the rapid development of the Internet and mobile Internet,e-commerce is developing rapidly in China,and product reviews are becoming more and more critical in e-commerce.Consumers tend to record and share their shopping experiences by posting reviews after shopping online,and potential consumers also use these reviews as a reference for their purchases.In e-commerce shopping,many consumers believe that their purchase decisions are influenced by product reviews,making it profitable to falsify product reviews.It is a significant research issue to detect spam reviews accurately and effectively,improve the credibility of product reviews,and clean up the e-commerce environment.Spam review detection is a subclass of text classification that aims to distinguish genuine reviews from spam reviews.Previous studies have focused on review text analysis,abnormal behavior detection,and intrinsic relationship identification.However,these spam review detection methods suffer from several shortcomings,mainly including(1)the detection techniques are inadequate,and existing models have difficulty modeling the semantics of long text reviews.Also,they do not exploit the semantic similarity of reviews for the same item and fail to detect spam reviews using the similarity of the user’s writing style.(2)Many reviews are carefully camouflaged,and existing work is inadequate in detecting fraudulent camouflage behavior.This paper constructs a heterogeneous spam review graph to address the above problems.In addition,it proposes two spam review detection methods based on heterogeneous graph neural networks.The main work of this study is as follows:(1)This work extracts the statistical features of users,reviews,and items from the experimental dataset using a feature engineering approach to address the shortcomings of insufficient detection techniques.It constructs a heterogeneous spam review graph using the behavior of users reviewing items,enhancing the hidden connections of users,reviews,and items through graph structure and enriching graph properties with statistical features and syntactic tree nodes.This study constructs a syntactic encoder and a semantic coder to enhance the capability of representing long reviews using the text’s syntactic information,semantic information,and statistical features.For the stability of the same user’s writing style and the semantic similarity of the same item reviews,this study proposes a sampling method based on user and item contexts on the heterogeneous graph.The neural network is trained by sampling sequence,and spam comments are identified by their context information.The experimental results demonstrate that the model based on text representation and heterogeneous spam review graphs can effectively detect spam reviews in the contextual environment.(2)There is a wide variety of fraudulent camouflage phenomena in e-commerce.This study proposes a heterogeneous subgraph neural network Spam-MHSNN,based on spam review graphs,to address the deficiency of existing work in detecting camouflage,which is used to detect spam reviews.The model constructs different metapath-based subgraphs for different camouflages,ignores the hard-to-identify nodes in the different subgraphs,and uses the rest of the information to detect spam reviews.In particular,this study constructs a User-Item subgraph for review camouflage,a Review subgraph for behavior camouflage,and a User-Review-Item subgraph for relation-based cocamouflage.This study proposes a node representation method for the User-Item subgraph that is based on metapaths.For the Review subgraph,this study suggests a subgraph representation method.And for the User-Review-Item subgraph,this study indicates an inter-metapath and intra-metapath node aggregation method.Finally,this study proposes a subgraph aggregation method to detect different camouflages and identify spam reviews simultaneously and automatically.Experiments demonstrate that heterogeneous subgraph neural networks can effectively detect different camouflage methods and improve spam review detection on different datasets.This paper investigates the problem of e-commerce spam review detection.This work has theoretical significance and practical value for reducing spam reviews,improving consumers’ shopping experience,increasing the income of honest merchants,and raising the reputation and competitiveness of the e-commerce platform.