Transactional-Behavior-of-Users-oriented Research On Fraud Identification Method

The popularity of electronic transactions not only greatly facilitates people’s life,but also gives fraudsters an opportunity to take advantage of it.In the identification of electronic transaction fraud,it plays an important role to accurately analyze the behavioral characteristics of users’ transactional behaviors.Due to the existence of camouflage,diversity and dynamics in users’ transactional behaviors,it is very challenging to accurately identify fraudulent transactions.First,the camouflage of users’ transactional behavior is that fraudsters generally disguise themselves as legitimate users to conduct transactions,resulting in the concealment of fraudulent transactional behaviors in legitimate transactional behaviors,which seriously interferes with the learning of fraud identification model and leads to the problem of transactional behavioral overlap;Second,the diversity of users’ transactional behavior is that users have both long-term and short-term trading habits,which increases the difficulty of identifying electronic transaction fraud and causes the problem of diversified transactional behavioral discrimination;Third,the dynamics of users’ transactional behavior is that users’ transactional behavior changes continuously with the passage of time,which degrades the performance of fraud identification model and causes the problem of transactional behavioral drift.In view of the above problems and challenges,this paper conducts the following research on fraud identification methods based on user transactional behavior:1.For the problem of transactional behavioral overlap,a transactional-featureextraction-based fraud identification method is proposed.Existing fraud identification methods fail to fully consider the importance of transactional features and ignores the impact of key transactional features on fraud identification,which makes it difficult for the fraud identification model to accurately identify the fraudulent transactions hidden in the legitimate transactions.Therefore,a set of transactional rules is proposed based on the characteristics of frauds,and new transactional features are constructed according to theses rules;Then,based on the Gated Recurrent Unit and attention mechanism,two feature extraction methods are designed to extract the importance between transactional features and the relationship between users’ historical transactional behaviors and current transaction,which effectively improves the importance of key transaction features on fraud identification.Compared with the existing methods on real-world transaction data sets,the experimental results show that the proposed method effectively extracts key transaction features and significantly improves the performance of the fraud identification.2.For the problem of diversified transactional behavioral discrimination,a behavior-aware memory network-based fraud identification method is proposed.Existing studies generally take the users’ long-term trading habits as his/her trading behavior,while ignoring the impact of the users’ short-term trading habits,thus the fraud identification model mistakenly identifies some short-term trading habits as fraudulent transactional behaviors,resulting in a large number of misjudgments of the identification model.Therefore,this study improves the memory structure of the recurrent neural unit,introduces long-term and short-term memory states to make the model store users’ long-term and short-term transactional behaviors respectively,and these two transaction behaviors are extracted from the perspective of transactional representation to form new representations of transactions,so as to improve the performance of the fraud identification model and reduce the misjudgment of the model.The comparative experimental results show that compared with existing methods,the proposed method can not only accurately extract users’ long-term and short-term transactional behaviors,but also can more accurately identify fraudulent transactions.3.For the problem of transactional behavioral drift,a drift-aware attention network-based fraud identification method is proposed.Existing studies generally believe that the changes of users’ transactional behavior caused by different transaction time intervals are the same,ignoring the relationship between the changes of users’ transactional behaviors and the transaction time interval,which makes it difficult for the fraud identification model to accurately describe the dynamics of users’ transactional behavior.Therefore,this paper improves the gating mechanism of recurrent neural unit.By designing a drift-aware gate and a behavioral attention module in the recurrent neural unit,this paper accurately perceives and extracts the different changes of users’ transactional behavior caused by different transaction time intervals,so as to improve the performance of fraud identification model.The comparative experimental results show that compared with the existing methods,the proposed method can deal with the problem of transaction behavioral drift more effectively and can identify frauds better.4.For comprehensive analysis of users’ transactional behavioral characteristics,a transactional behavioral characteristics-oriented fraud identification ensemble method is proposed.Users’ transactional behaviors present a variety of characteristics,such as the camouflage,diversity and drift of users’ transactional behaviors.An efficient fraud identification model has to comprehensively consider these behavioral characteristics,so as to identify frauds more accurately and reasonably.In this paper,the camouflage,diversity and drift of the user’s transaction behavior are comprehensively considered,and the models proposed for the above three aspects are integrated into the fraud identification ensemble method for transactional behavior.The comparative experimental results on real-world transaction data sets show that compared with existing fraud identification methods,the proposed method can comprehensively consider these behavioral characteristics and can effectively improve the performance of fraud identification.