A Research On Anti-fraud Of Bank Card Transactions Based On Deep Learning And Ensemble Learning

With the ever-changing methods of bankcard transaction fraud,the rule-based anti-fraud systems have been insufficient to adapt to the current fraud situation,and neural network with selflearning capabilities has come into view,as an important tool for anti-fraud.However,the traditional neural network has only one hidden layer node,which has limited ability to generalize complex and variable fraud features.It is usually get into trouble in local optimum and gradient diffusion.On the basis of traditional neural networks,deep learning increases the number of hidden layers,and also overcomes the problems in traditional neural networks by layer-by-layer training,thus enabling the model to have more analytical and reasoning capabilities.Therefore,this paper attempts to apply Deep Belief Network(DBN)to anti-fraud research on bank card transactions.In order to further improve the anti-fraud ability of DBN model,this paper proposes an ensemble learning model based on DBN and Support Vector Machines(SVM).The main works are as follows:1.This paper analyses the fraud situation and characteristics of bankcard transactions currently facing domestic banks,compares the anti-fraud systems based on rules and neural networks.2.A stacking ensemble learning model using DBN and SVM as the base classifier and Logistic regression model as the metaclassifier is proposed.The SVM model is used to express feature associations with high frequency under anti-fraud scenarios.The DBN model is used to mine unknown fraud feature associations,and the fraud probabilities derived from the two models are passed along with the actual fraud tags to the Logistic regression model.Train again to get the final results of fraud.3.In the process of constructing DBN and SVM model,optimizing the performance of every single model by processing data and choosing the model parameters.And the differential ensemble is used to reduce the risk of overfitting.4.Using actual bank card transaction data to evaluate the model performance.The experiments prove that the ensemble model has achieved better prediction results than any single model,reflecting the application value of the ensemble model in anti-fraud research during bank card transactions.