Research On Adaptive Cost-sensitive Multi-classification Model For P2P Lending Platform Fraud Detection

Recently,with the rapid development of Internet finance,there has emerged more financial business innovation patterns,which greatly promote the prosperity of the financial industry.As a new Internet financial innovation,P2 P lending can better meet the borrowing requirements of individuals and small enterprises,due to its micro-finance lending,independent choice,and with the freedom in borrowing.P2 P lending makes great significance in improving our existing financial system and relieves private investment demand.However,behind its advantages,P2 P lending has gradually exposed more and more risk and supervision issues.Fraudulent events,such as platforms collapse,run,and get into financial crisis,occur frequently and it brings huge hindrance to the healthy development of financial industry.Thus,this paper puts forward an adaptive cost-sensitive weighted ensemble algorithm to build multi-class decision tree integration model,which can realize P2 P lending platforms fraud detection and early warning and provide decision support for regulators and investors.Based on the review of characteristics and methods of P2 P lending fraud detection problem,this paper researches data-imbalanced and multi-class issue and put forward an ensemble algorithm for fraud detection.The major research works are listed as follows:(1)The research status of P2 P lending fraud detection are summarized,and the data imbalance and cost-sensitive characteristics are analyzed.Combining these characteristics,the major problems of existing methods are pointed out.(2)This paper dissertation researches the advantages and shortages of decision tree algorithms on fraud detection,and put forward the major problems of C5.0 algorithm,which is the base for problem solving.(3)Based on the major problems of existing method,the adaptive cost-sensitive weighted ensemble algorithm ACWEA is put forward.The major improvements include three parts:Firstly,the weighted information gain rate is put forward as the standard of attribute selection,which not only considers the amount of information,but also emphasizes more priority on detecting of fraud,reflecting the contribution differences on different classes.Secondly,based on C5.0 algorithm,the cost matrix is introduces to adjust the weighting factors of fraud,aiming to minimize the misclassification cost.For the static learning mechanism,the adaptive cost-sensitive learning is put forward with the cost matrix,the objective function,and the searching method to improve the generalization ability.Finally,to solve the multi-classification problem,the stepwise one-to-many method is used to convert the integration dataset into binary classification and construct the ensemble model,which can improve all categories’ classification performance.(4)This paper use 2890 real P2 P lending platform records from web portal to test the model.The large-scale data experiments compare the performance to popular machine learning methods,and the statistical analysis results verifying the significant advantages of ACWEA.