E-commerce Transaction Fraud Detection Based On Heterogeneous Graph Neural Networks

With the rapid development of mobile payment,the phenomenon of fraud transactions in e-commerce platform is becoming more and more serious.Fraudulent trading detection has been an important part of risk control system of e-commerce platform.However,while the platform is cracking down on these fraudulent transactions,the criminals’ fraudulent methods and fraudulent abilities are constantly improving,leading to a continuously high fraud risk and expanding economic losses.Therefore,how to accurately identify fraudulent transactions has become an important research topic.Starting from the subjects involved in transactions,this paper constructs a heterogeneous graph of nodes including transaction,address,IP and mobile phone number,aiming to solve the detection problem of fraudulent transactions on e-commerce platforms.This paper summarizes the following four major analyses as the fundamental basis for carrying out the work:(1)the behavior of malicious users is limited by the cost of resources,and most malicious users will only submit orders frequently on a few resources;(2)the browsing behaviour of users has good differentiation ability in determining whether users are malicious;(3)malicious users tend to generate more clicks and will be active at some unusual times;(4)some immediate characteristics of transactions can visualise the degree of anomalies in transactions.The first of these analyses inspires us to construct an information network using a transaction-resource heterogeneous graph,and analyses two and three and four portray a transaction from three aspects,respectively.To exploit user browsing behavior data,we treat clickstream data as a sequence of behaviors,learn the sequence embedding using Transformer model,and combine it with user and transaction features to form a hybrid feature of transaction representation.In terms of model,this paper further improves on the HAN model(Heterogeneous Graph Attention Network)and proposes the HA AN model(Heterogeneous Feature Attention and Graph Attention Network),which is based on heterogeneous graph neural network to extract transaction information,aiming to solve the problem of e-commerce transaction fraud detection.The HA AN model consists of three components.Firstly,we introduce feature-level attention to better extract each node’s multi-dimensional information.Secondly,neighbors of each node may play different roles,so we use node-level attention to aggregate the neighbors when fusing neighboring information,enabling adaptive matching of weights to different neighbors.Finally,a series of meta-paths are defined in terms of both internal and cross-views,and the importance of different meta-paths is learnt through the meta-path attention.This three-layer attention structure not only enables learning by assigning different weights to various aspects of information,but also has comparable interpretability to models such as logistic regression,GBDT,and xgboost,which improves the confidence of the model for fraudulent transaction cases in the real environment.In this paper,two datasets of different scales are constructed using real data on the JingDong platform and extensive experiments are carried out.The experimental results show that the F1-Scores of the models are 0.9117 and 0.8862 respectively,with an improvement of 0.003 and 0.01 respectively compared to the best base model.This verifies the rationality and effectiveness of HAAN model,and provides a new effective identification method for the study of fraudulent transaction detection.