Research On Network Spammers Detection Based On Graph Neural Networks And User Rating Behavior Characteristics

In the era of e-commerce,with the development of the network and logistics,more and more users tend to choose online shopping.The true quality of goods often depends on the objective and true ratings of users.However,many businesses,for their own benefit,often maliciously control the ratings of target goods by employing a large number of spammers.These behaviors not only disturb the judgment and purchasing power of users,but also seriously affect the credibility and market order of online rating system.Therefore,how to design an efficient and robust fraud detection algorithm has become very important.To solve this problem,some scholars have built online rating networks based on user relationships with products and proposed online reputation ranking methods.This kind of method allocates the user’s initial reputation through the user’s rating of the product,and identifies users with low reputation value as malicious users.Although such methods can find a large number of malicious users in a short time,when the number of spammers exceed a certain threshold,the stability of the algorithm will decline.In addition,with the rise of graph neural networks in the field of fraud detection,researchers tend to improve the recognition of suspicious user nodes by changing the number of layers and internal aggregation mechanisms of graph neural networks,ignoring the impact of noise nodes on the generalization ability of fraud detection models.In the field of fraud detection,how to solve these problems has become an important research direction.Therefore,this thesis first proposes a reputation ranking method based on user rating patterns and rating deviation(RPRD).This method allocates initial reputation through the user’s criterion function,and identifies malicious users through user rating behavior features.In order to verify the performance of the RPRD method,this thesis uses three real datasets and three reputation ranking methods to test and compare.The experimental results show that this method not only effectively resists the attack and identification of spammers,but also has high accuracy and robustness with the number of spammers increases.It also can be applied to large sparse two-part rating networks in a short time.In addition,On the basis of this method,this thesis proposes a general spammer indicator of rating systems uncovering rating preferences and bias(GNR)to supplement the existing reputation ranking methods.we carry out research from the basic assumption that normal users and malicious users can be fundamentally divided into two groups by a general indicator.Such an indicator reflects the ladder rating behaviors from normal users and non-ladder rating behaviors from spammers.In this thesis,the GNR indicator is applied to three reputation ranking methods.The experimental results show that this indicator can significantly improve the accuracy and robustness of existing fraud detection methods.Finally,to enhance the efficiency of fraud detection models in identifying suspicious user nodes,this thesis also introduces a method of using adversarial data augmentation technology(FLAG)method to improve the generalization ability of fraud detection models,This method generates gradient based adversarial disturbances by inputting node features,and then adds adversarial samples to the original dataset,enabling the model to achieve higher accuracy and classification performance during the training process.The experimental results show that the model achieved excellent performance on the dataset after applying the FLAG method,and the accuracy and generalization ability of the model were significantly improved.