E-commerce Fraud Detection Method Based On Address Embedding

In recent years,with the rapid development of e-commerce,a large number of credit and t,hird-party financial fraud incidents have occurred in e-commerce transactions,causing huge losses to users and third-party payment companies.Therefore,the e-commerce fraud detection problem has aroused widespread concern in industry and academia.The machine learning and data mining methods have been widely used in e-commerce fraud detection.The existing e-commerce fraud detection methods mainly train the machine learning model based on feature engineering.Under the support of the e-commerce transaction data set,it achieves the purpose of e-commerce fraud detection by manually extracting relevant features,including user features,transaction features,payment featuresetc.In practical applications,such methods often face the following challenges:(1)Most machine learning models can only accept input in vector form,while e-commerce transactions often have some complex form data,such as IP address and shipping addres s.These addresses information can't be directly represented by vector data.(2)Features in e-commerce data are highly different,such as timing information,geographic location information,and amount infor-mation.Existing fraud detection models are difficult to achieve effective integration and utilization of these features.Aiming at the above challenges,we propose an address representation learning method based on e-commerce transaction information embe dding.The low-dimensional vector repre-sentation of addresses learned from massive e-commerce transaction data,not only effectively maintains the geographical proximity of addresses,but also better reflect the population flow between cities;On this basis,we use the XGBoost model to effectively integrate various types of differentiated features in e-commerce transaction data and improve the effectiveness of fraud detection.Finally,we collected more than 30 million e-commerce transaction data in the actual business scenarios of third-party payment companies.By querying the user's report informa-tion and reviewing by the business personnel,we added two classes of labels:normal trans-action and fraudulent transaction for the transaction record in the data set.And based on the data set,we design the address embedding visualization experiment and fraud detection experiment.The results show that our proposed model can achieve significant improvement compared with the comparison algorithm.