Research On Transaction Fraud Detection Based On Rule Attention Machine

With the economic globalization and the rapid development of big data,artificial intelligence and other technologies,the prosperity of digital financial technology has gradually changed people's consumption habits and the development mode of traditional financial industry.Recent years,China's Internet consumer finance industry has a good trend,which has promoted the development of relevant industries and thirdparty mobile payment platforms such as PayPal,Alipay and WeChat.A large number of e-commerce platforms are accepted and used by the public,and the card free online transaction gradually replaces the traditional bank card and cash transaction mode.Especially in 2020,the outbreak of Covid-19 led to a surge in the number of online transactions,which not only brought great convenience to people's lives,but also provided more opportunities for criminals,promoted the frequent occurrence of online transaction fraud cases,damaged the rights and interests of normal users,merchants and platforms,and caused huge economic losses.Accurate detection of fraudulent transactions has become one of the tasks that can not be ignored in the anti fraud in the financial field.However,the complexity of transaction fraud detection task makes it face some challenges.Firstly,in real trading activities,the proportion of illegal transaction data in all transactions is very small,and the unbalanced category distribution poses a challenge to prediction modeling.In addition,the behavior of transaction fraudsters may evolve with the improvement of the detection system,so the data may have conceptual drift.Some work in the field of fraud detection has achieved good accuracy,but the model is similar to black box,which is difficult to explain the internal transaction fraud model.In view of the above challenges,this paper proposes a transaction fraud detection model based on rule attention mechanism to identify illegal transactions more accurately.The model mainly consists of two parts:the transaction decision rule generation module based on tree and the dual attention transaction perception module,which improves the prediction effect of the model.The method based on tree and integration is used to generate transaction fraud decision rules,improve the effectiveness of the model,give the model interpretability through the advantages of tree structure,and deal with the imbalance of transaction data.The rules are embedded and represented,and the high-dimensional sparse rules are processed to improve the flexibility of subsequent model updating and adapting to different data.Two attention mechanisms are added to calculate the attention score of rule interaction and rules under specific conditions,which is helpful to model training and assign different scores to decision rules,which is helpful to the visibility and interpretability of the pattern of fraudulent transactions.The comparative experiments were carried out on real data sets.It is proved that the proposed method has better performance than the existing common methods.And verify the effectiveness of the decision rule generation module and attention perception module proposed by the model for the task.