Research On Credit Risk Assessment Based On TabNet

With the vigorous development of the world economy and the gradual deepening of China’s reform and opening up,credit card payment has become one of the most commonly used payment methods in modern society.As a growing service,how to evaluate whether the applicant can repay the loan plays a key role in correctly supporting financial institutions to define their policies and strategies.In the past decades,people have proposed many methods to evaluate personal credit risk,but there are still many problems to be solved to accurately evaluate credit risk.One of the problems is how to make full use of the existing unlabeled data,so as to reduce the labeling cost.The key technology to solve this problem is semi supervised learning.Another problem is how to improve the accuracy of the model based on deep learning.The key to solve the problem is to introduce the idea of tree model.The mainstream credit evaluation method in the credit scenario is the scorecard,and a small number of businesses are evaluated by machine learning method.The model accuracy and the progressiveness of the technology of credit evaluation have great room for improvement.In the real scene,the labeling of samples needs to consume a lot of resources,so a large number of samples without class labels are wasted,which leads to only a small number of samples with class labels in the process of model training.Semi supervised learning can make full use of unlabeled data to reduce the cost of labeling,enrich the training set of the model,and improve the accuracy of the model.Tabnet is emerging in the field of tabular data analysis.It combines the characteristics of tree model and NN model to guide the assessment of credit risk.Through the research,this paper mainly does the following work:First,combined with relevant data,the tabnet deep learning algorithm is used to establish a credit risk assessment model,which is optimized in combination with the actual scenario.This paper compares the credit risk evaluation model constructed by tabnet model with tree model and CNN model.The experimental results show that the preliminarily optimized tabnet model performs better in accuracy than the control model;Secondly,the problems existing in the characteristics and parameter selection of TabNet are studied and analyzed respectively.Genetic algorithm(GA)is used to optimize the attention transformer automatic feature selection module of TabNet,and particle swarm optimization(PSO)is used to optimize the super parameter selection of TabNet to realize the automatic search of parameters.Analyze and verify several major reasons for the "premature" problem of genetic algorithm,and further optimize the genetic algorithm according to the experimental results;Finally,semi supervised training is introduced to improve the sample data.In order to solve the common problems of data imbalance and missing data labels in credit scenarios,combined with unsupervised pre training and pseudo label technology,an iterative weighted semi supervised training method is designed,which dynamically generates pseudo labels for unlabeled data during the training process,expands the training set data,and improves the effect of the algorithm.To sum up,this paper introduces the tabnet model to assess personal credit risk,studies the characteristics of the model and the problems in the parameter selection module,gives the optimization direction of the model,combines the semi supervised training method,and finally constructs a semi supervised credit risk assessment model based on TabNet.