Research On Bayesian Network Model Of Credit Evaluation From Missing Data

The Peer to Peer Lending(P2P)model has been welcomed in the financial fields since it came into being,and has been developed rapidly in recent years.Its characteristics of low threshold,high interest rate and convenience have brought some benefits to both lenders and borrowers.Especially for SMEs and those who do not meet the lending criteria of traditional financial institutions,which makes the financing difficulties in the past have been solved.It also has a certain impetus to the implementation of Inclusive Financial policy and the improvement of financial markets of our country.However,with the development of P2 P online lending,some drawbacks have gradually emerged.On the one hand,because of the low threshold and high interest rate on the P2 P platform,it is easy to increase the default risk of borrowers.On the other hand,many credit indicators that need to be filled in by borrowers are just a decoration,so the credit data of borrowers on the platform contain a lot of missing values,which will affect the later credit evaluation and reduce the accuracy of credit evaluation.Therefore,it is rather important to construct a stable and effective P2 P credit evaluation model for missing data,which is also the basic premise for the normal operation of P2 P lending platforms.In recent years,many scholars have studied many methods of modeling the P2 P credit evaluation model.However,in the existing work,there is few methods of constructing P2 P credit evaluation model from missing data.Therefore,to avoid the mentioned disadvantages,for the missing data,this paper address the P2 P credit evaluation model based on Bayesian network which not only is good at combining prior knowledge with data information,and also has unique advantages in eliminating the uncertainty.The main work that have been done are concluded the following three aspects.Firstly,this paper proposes a structure learning algorithm of Bayesian network for missing data,named PQISEM algorithm(Structure Learning of Bayesian network based on partial qualitative influences and Structure-Expectation Maximization algorithm).In PQISEM algorithm,in order to constrain the parameters and structure during Bayesian network structure learning,we introduced the easily obtained qualitative influence knowledge with good reliability and robustness into the learning algorithm,which makes the parameters and structure of Bayesian network are closer to the real network.The proposed algorithm outperforms other algorithms in terms of BIC score,KL divergence,and running time.Secondly,P2 P credit data are been preprocessed and then the P2 P credit evaluation Bayesian network model,named CEBN model,is been constructed.Specifically,we firstly preprocess the data,including discretizing the continuous variables by ChiMerge algorithm.And then the structure and parameters of CEBN model on P2 P platform are constructed using PQISEM algorithm from the preprocessed data,which not only improves the utilization rate of data,but also improves the accuracy of the model.Thirdly,we infer and predict the borrowers' credit based on the constructed CEBN model on P2 P lending platforms.Specifically,we use the inference principle of Bayesian network to infer and predict the borrower's repayment status and then we obtain the evaluation results are more excellent than other Bayesian network structure learning algorithms in terms of BIC score and prediction accuracy,and also are better than other classical classification algorithm such as Na?ve Bayes,C4.5,SVM,Adaboost and so on by accuracy,precision,F value and other evaluation indicators.Which verifies the better performance of the CEBN model.