Research On Early Warning Of Personal Consumption Credit Default Based On Cross Attention Mechanism

With the steady development of China’s economy and the continuous improvement of social consumption level,Internet consumer credit has become more and more popular with consumers by virtue of its advantage of paying first and repaying later.At the same time,it also exposed problems such as imperfect risk management system and incorrect default warning.In order to solve these problems,some scholars use statistical analysis methods and machine learning models for modeling and analysis.However,large scale,poor user characteristics and unbalanced credit data lead to poor early warning effect of existing methods.In this regard,this work uses cross focus mechanism and better loss function to study the early warning of consumer credit risk.The use of reduced and dimension features not only makes the process of selecting model features very inefficient,but also reduces the measurement accuracy of the model and increases the training cost.In order to solve the above problems,this paper proposes an algorithm and model based on cross attention mechanism.Firstly,this paper proposes a cross attention algorithm with variable weight based on cross attention algorithm.The variable weight cross focus algorithm uses the credit business indicator system and the correlation matrix of credit features to group individual credit features,uses the attention mechanism to increase the calculation of feature correlation within the group,and uses the method of merging variable weight features to strengthen the calculation of data correlation for each group,thus overcoming the inefficiency of model feature selection.Secondly,based on the variable weight cross focus algorithm and Transformer model,a financial cross focus(CAFI)model is proposed.CAFI model design adopts the minimum parameter quantity and lighter structure,which overcomes the problem of reduced prediction accuracy of the model,improves the overall training efficiency and reduces the training cost.Numerical experiments show that CAFI model has higher test set precision(80.31%)and verification set precision(80.01%)than traditional machine learning models on the public consumption credit data set,which shows that variable weight cross focus mechanism and CAFI model are more effective in the classification of personal consumption credit.The imbalance of consumer credit data means that the model cannot accurately predict potential default users,thus causing financial losses to credit institutions.Based on the existing data imbalance loss function,this paper proposes a weight loss function with variable adjustment parameters.Weight loss introduces penalty parameters and gradient terms of loss function and real value as adjustment parameters to increase the contribution of some types of samples to the loss function.Compared with other loss functions,it makes the model pay more attention to the default users in the training process,effectively avoids the second type of error that confuses the default users with normal users,reduces the impact of data imbalance on the model training results,and improves the effectiveness of classification results.The numerical experiment results of unbalanced personal consumption credit data show that: 1)Compared with other loss functions,weight loss can significantly improve the AUC value and F1 score of CAFI training results.The research shows that weight reduction can effectively solve the problem of data imbalance and improve the effectiveness of CAFI model.2)Compared with other machine learning models,CAFI weight loss method combining CAFI model and weight loss function has higher AUC value(0.60)and F1 score value(0.3633),indicating that CAFI weight loss method works well on unbalanced credit data and can effectively reduce losses caused by false predictions of credit institutions.In a word,the research results of this paper can provide feasible methods and research programs for Internet consumer credit early warning to support the application and development of Internet credit risk management.