Research Of Intelligence Anti-Fraud Methods In The Field Of Consumer Lending

With the rapid popularization of inclusive finance in China and the rapid development of Internet finance,consumer finance business has experienced a spurt of development and has become the focus of retail business development of various commercial banks.However,the decline in consumer finance customer base and the lack of sufficient new customer base Credit data,the consumer finance industry is facing pressure from the failure of traditional risk prevention and control methods.At the research level,fraud risks mainly include individual fraud and gang fraud,and individual fraud is gradually evolving to organized and large-scale group fraud,and the behavior pattern of fraudulent customers is changing with each passing day.Traditional theories and methods have been used to identify fraud.overwhelmed.In order to cope with the complexity and diversity of fraud,mainstream consumer financial fraud detection methods have gradually changed from the original black-and-white list and rule system based on expert systems to machine learning-based detection methods.On the one hand,fraud behaviors gradually show gang and professionalism.On the other hand,the existing machine learning fraud detection methods only consider from the perspective of users' own characteristics,and do not consider the interaction between users.Aiming at the above two aspects,this article starts from the real consumer finance company data set and uses the call relationship between users to build an association network between users.Aiming at the user's associated network,semi-supervised and supervised learning methods are used to introduce the user's network characteristics into the machine learning anti-fraud process.Through experimental comparison,better fraud detection results are obtained.First of all,the gradual gang formation of fraud behaviors is reflected in the association network with users as nodes.There is a "homogeneity effect" between nodes,that is,the relationship between fraud users and normal users is sparse,but the relationship between fraud users is sparse.close.Based on this situation,we propose a collaborative classification method based on label propagation for fraud detection.The collaborative classification problem is defined as given a network and the label information of some nodes,how to infer the label information of unknown nodes based on the network information,and improved the label propagation algorithm avoids performance degradation caused by the uneven distribution of fraud samples and normal samples in the associated network.Secondly,based on the shortcomings of the current machine learning fraud detection scheme,which only uses the inherent characteristics of users,it is proposed to use network representation learning methods and network statistical indicators to extract the network features of user nodes,and introduce the network features into the feature system of fraud detection.And use multiple machine learning models to train predictions,and use model fusion on the basis of single models to improve detection results.The experimental results prove that the added network features improve the final model fraud detection performance,and the use of network representation learning to extract network features requires less feature engineering and better results than extracting inherent features.