Application Research Of LightGBM Model Based On Bayesian Optimization In Credit Evaluation

Due to the impact of the 2020 epidemic,our country’s credit scale has grown rapidly;the increase in scale is followed by an increase in credit risk.Therefore,it is necessary to adopt new methods to study the problem of personal credit evaluation.This paper combines the actual business situation and adopts a new evaluation index F-beta value,that is,when the accuracy rate and the recall rate are fused,the recall rate is given more weight.The research content is as follows:(1)The commonly used method to solve the problem of sample imbalance is the SMOTE algorithm to construct a new sample to balance the sample;this article gives a new attempt,because there are less overdue data and more non-overdue data;we can convert the two classification problem into a single Classification problem,that is,outlier detection problem.Use negative data to train a classifier to get a bound;beyond the bounds are positive data(2)There are two methods for parameter selection: empirical method and ergodic method.These two methods have their own advantages and disadvantages in the tuning effect and running time;the empirical method can quickly obtain the specific values of the parameters based on past experience,but the empirical method has an obvious defect,that is,the optimal parameter combination cannot be obtained.The traversal method can get the optimal parameters,but its time cost is very high.A compromise is made between the advantages and disadvantages of the two methods,that is,the parameter tuning effect of this method is slightly worse than that of the traversal method,which reduces the time cost.Bayesian parameter optimization is designed based on the second.Through multiple comparisons,the performance comparison of a single model,the performance of the same model under different optimization algorithms,and the comparison with other machine learning algorithms,the final conclusion is reached.In addition,the importance of the characteristics of the variables is analyzed,and the top five variables that have the greatest impact on overdue are analyzed and recommendations are given.