| With the rapid development of mobile payment,various apps have become the main payment way of offline consumption.In order to meet the needs of different Internet payment scenarios,aggregated payment technology came into being.Aggregate payment is an extension of third-party payment platform services and is an inclusive payment tool that integrates multiple payment channels and can be used across all platforms.Due to the low technical threshold and high profit,the demand for aggregated payment platforms in financial supervision and superior payment channels continuously increases.Hence,the risk control ability of the platform needs to be improved to cope with future opportunities and challenges.This thesis focuses on anomaly detection of the aggregated payment platform based on merchant transactions.We aim to detect abnormal transactions with illegal interests and locate abnormal merchants to make the platform more secure and compliant.According to the business characteristics and the consuming scenarios of the aggregated payment platform,a large amount of transaction data and merchant information are analyzed.We can extract the relevant features to describe the transaction and merchant characteristics by mining massive data.A transaction-level merchant anomaly detection model is proposed according to the platform data characteristics.The detection model can effectively improve prediction accuracy.In addition,to solve the problem of different transaction intentions and the mixture of high-risk(risk types are fraud,gambling,etc.)and low-risk(risk types are cash-out,order swiping,etc.)merchants,cluster-based and risk screening-based merchant anomaly detection models are proposed,greatly improve the recognition rate of anomalies.Main contributions of this thesis are as follows:1.Design a merchant anomaly detection model based on normal transaction representation according to the three modules of transaction context,transaction history in the same period,and merchant information:Considering that the amount of abnormal merchant data accumulated by the platform is small and the abnormal label is at the merchant level,it is impossible to determine whether the specific transaction is abnormal.We first extracts transaction features by combining business knowledge,then proposes the assumption that all transactions of normal merchants on the day are normal,and draws on the inspection methods of business experts on abnormal merchants.A merchant anomaly detection model based on normal transactions is developed.The model can learn other transaction information adjacent to time through the network structure between transactions,combine the transaction history information and merchant information of the same period to learn the normal transaction representation,and then judge whether the current transaction is abnormal according to the transaction representation.2.Propose a clustering-based merchant classification method to solve the problem of different transaction intentions of merchants:The transactions of merchants with different business patterns show different intentions.The model proposed in this thesis is based on one-class classification and aims to learn the pattern of normal transactions.Different transaction intentions will reduce the purity of transaction samples and affect the model’s performance.Therefore,in this thesis,we adopted the clustering method to classify merchants with different business patterns and learn the transaction characteristics of merchants with varying business patterns.3.Propose a method based on risk screening to solve the mixed problem of highrisk and low-risk merchants:The severe mixture of high-risk and low-risk merchants is a common problem in anomaly detection.This thesis leverages the distribution difference of merchants of several perspectives to separate low-risk merchants from the mixed data,enabling the model to identify more abnormal merchants accurately.The complete experimental results on the real data set of an aggregated payment platform show that the proposed method can identify more abnormal merchants for merchant anomaly detection effectively.Compared with the existing anomaly detection model,our model has higher recognition accuracy for abnormal merchants,which is of great significance in reducing the pressure on business personnel and improving the risk control ability of the payment platform. |
