Research On Fraud Detection Based On Heterogeneous Graph Neural Network

With the advancement of information communication technology and digital technology,especially the development of 4G and the emergence of 5G,a highly interactive world has been created,thus establishing a huge digital society,which makes human communication methods simpler,and also greatly improves efficiency.However,driven by malicious competition or huge commercial interests,a large number of fraud behaviors have also appeared on the Internet,such as review fraud,financial fraud and Internet advertising fraud.Because the fraud behavior itself has good camouflage and is not easy to identify,coupled with the increasing amount of information on the network,it is difficult to detect fraud manually,the cost is high,and the accuracy is not ideal.Therefore,the Internet field urgently needs some scientific and efficient detection methods to detect these fraud.Recently,Graph Neural Networks(GNNs)have been widely used in fraud detection tasks.GNNs first generate node embedding by aggregating neighboring information under different relations,and then use the final node embedding to detect the node’s suspiciousness.However,traditional GNNs employing only a single type of neighborhood aggregator fail to capture neighbor information from multiple perspectives and treating different relations equally inevitably weakens the semantic information of heterogeneous graphs.Meanwhile,expressive ability of GNNs is limited by using conventional concatenating or averaging operations to update the center node.To handle these problems,This paper proposes a heterogeneous graph fraud detection model based on multiple neighborhood aggregators to conduct fraud detection tasks.In addition,fraudsters often connect with benign entities through camouflage,which will seriously damage the performance of the model.We further design a fraud camouflage behavior detection model based on trainable neighbor sampler to perform fraud detection tasks.The main research content and innovation work of this paper are as follows:(1)Heterogeneous graph fraud detection model based on multiple neighborhood aggregators is proposed.(MAFD)Most of the existing fraud detection methods based on graph neural network only use a single type of aggregator to aggregate the neighborhood information,which leads to the problem that the model can only capture one aspect of the neighbor information.In addition,the expression ability of graph neural network is limited by using traditional averaging or concatenating methods in the updating phase.Therefore,this paper proposes a heterogeneous graph fraud detection model based on multiple neighborhood aggregators.Concretely,using multiple types of aggregators to aggregate neighbor information and using aggregator-level attention to learn the importance of different aggregators.Also,relation-level attention is leveraged to learn the importance of each relation.Besides,conventional update operations are replaced with vector-wise implicit and explicit feature interactions.Experimental results on AMAZON and YELP datasets show that the model can effectively improve the accuracy of fraud detection.(2)Fraud camouflage behavior detection model based on trainable neighbor sampler is proposed.(TNS)Fraudsters often connect with benign entities through camouflage,which will seriously affect the performance of fraud detection model.Although most of the existing fraud detection methods based on graph neural network can achieve good effect,the vast majority of the work ignore the fraudster’s camouflage behavior,Therefore,this paper proposes a fraud camouflage behavior detection model based on trainable neighbor sampler.The model designs a trainable neighbor sampler to identify and filter camouflaged fraudsters.Experimental results on AMAZON and YELP datasets show that this method can effectively identify and filter the camouflaged fraudsters and improves the accuracy of fraud detection.