Fraud Detection Model Of Third-party Payment Based On Anomaly Detection

At present,with the rapid development of Internet finance business,the black industry is blooming,and has become one of the world’s fastest growing "industries".At the same time,various forms of online transaction fraud methods emerge in an endless stream and update quickly.How to detect fraud online transactions and give early warning is crucial for third-party payment platforms,which not only has great significance for protecting the interests of consumers and merchants,but also promotes the healthy development of third-party payment.Third-party payment platforms are constantly exploring anti-fraud detection technologies based on machine learning,mainly using supervised learning algorithms which highly rely on labeled data.On the one hand,it requires a lot of manpower to label online transactions,and on the other hand,it is plagued by the problem of unbalanced sample,so the model is not much effective.The anti-fraud work of the third-party payment platform is still facing severe challenges.In view of the above problems,this paper carries out an in-depth study.This paper takes the online transaction of third-party payment as the research object and studies the detection of fraud risk in online transaction.This paper proposes a research framework for fraud detection:(1)By studying the current theories and literatures of third-party payment at home and abroad,understand the research direction of relevant content,define the concept of third-party payment and its risks,especially fraud risks,and locate the difficulties and pain points of relevant research;(2)Based on the difficulty of current fraud detection research,namely the problem of unbalanced sample,this paper proposes corresponding solutions from the dimensions of data,feature selection and model.(3)Collect data to verify the effectiveness of the proposed method from the perspective of practical application,and compare with other commonly used methods;(4)Summarize the research conclusions,reflect on the shortcomings,and put forward ideas for improvement.Based on the online transaction data of third-party payment platforms,this paper draws the following conclusions through empirical analysis:(1)At the data level,the fraud transaction sampling based on user granularity and the normal sample purification scheme based on PU-learning have increased the proportion of fraud samples from 1.79% to 2.98%,and the model effect after sampling has been verified on the general model.(2)At the feature selection level,from the comparison of feature selection methods,this paper adopts the permutation method to extract the important features of the black and white samples of the One-Class SVM model for modeling,which is better than other methods that use the feature importance calculation function of the model and the feature selection method based on IV value commonly used in practice.The characteristics of fraud transactions and normal transactions can be extracted more effectively,which is more effective for improving the detection effect of the final model.(3)At the model level,under the same feature scheme,compared with the model under other unbalanced sample processing methods,the model proposed in this paper has a more significant effect on the identification of fraud transactions,with the accuracy,recall and AUC increased by at least 8.9%,7.6% and 6.6%.(4)Compared with the common fraud detection models constructed by various feature selection schemes and unbalanced data processing schemes,the fraud detection model proposed in this paper also has more significant effects,with three indicators improving at least 3.2%,4.2%and 1.0%.The research in this paper has certain reference significance for improving the detection of fraud risks in online transactions of third-party payment platforms and improving the anti-fraud system of third-party payment platforms.