Research On Generative Adversarial Networks-based Approaches For Detecting Review Spammer Groups On E-commerce Websites

The reviewer spammer groups on e-commerce platforms can write a large number of fake reviews for the target products,which seriously affect the purchase decision behaviors of consumers.Relevant researches have shown that some merchants might hire a group of reviewers to write fake reviews,such spammer groups are more harmful to e-commerce websites than individual review spammers.To address this problem,many approaches rely on artificial feature engineering to design group spam indicators or features to capture the aggregated behavioral patterns exhibited by spammer groups,which have these drawbacks:costly,time-consuming,over-reliance on data sets.With this limitation in mind,this paper has some effective researches for detecting spammer groups on the e-commerce websites.Firstly,based on generative adversarial networks,one approach which with two stage to find spammer groups is proposed.The approach first perform multi-view sequence sampling on a given data set and use the Word2 Vec model to obtain the low-dimensional vector representations of users.This approach constructs a N-nearest neighbor user relationship graph according to the similarity of users in the embedding space.Then the approach employs the DBSCAN clustering algorithm to obtain the candidate groups in the closed neighborhood of each node in the graph.A generative adversarial network model uses the generative reconstruction loss and the discriminator loss to evaluate the spamicity of each candidate group.Secondly,this paper also proposes a spammer groups detection approach based on adversarial autoencoders.In the approach,a random walk method guided by meta-paths is used to generate network embeddings,and the approach construct a user relationship graph according the cosine similarity of users.Then the approach also employs the DBSCAN clustering algorithm to obtain the candidate groups.The approach uses the Doc2 Vec model to obtain the low-dimensional vector representations of the candidate groups.Since the AAEOC model is trained to fit the distribution of normal candidate groups,the spammer groups should have higher loss than normal candidate groups.Therefore,the proposed approach can use the loss the infer the spammer groups.Finally,the experimental results on two real-world review data sets show the proposed approaches outperform the baseline methods in detection performance.