Research On Personal Credit Risk Assessment Based On Interpretable Machine Learning Mode

In the current context of financial and economic globalization,with the continuous liberalization of financial markets in various countries,the assessment of credit risk is particularly important.There are currently two main methods for credit risk assessment,namely traditional methods represented by logistic regression and machine learning methods represented by various machine learning models.However,using traditional methods to assess customer credit risk results in weak model fit and difficulty in accurately assessing customer risk.Therefore,some scholars have introduced machine learning models into the field of credit risk assessment.However,two problems also arose: firstly,due to the high dimensionality of the data,the training of the model is difficult and the training cost is high.Secondly,machine learning models themselves have high complexity,making it difficult for people to understand the discrimination mechanism and prediction basis of the model,which leads to trust issues in the model.Therefore,in order to solve the problem of high data dimensions,this article uses the GAMI Net model to select features based on Lending Club personal credit data and a domestic commercial bank data,respectively,in order to reduce data dimensions and compare them with LASSO and VIF methods.Introduce the selected features using various methods into the credit risk assessment model,compare the prediction accuracy of the model,and verify the effectiveness of GAMI Net feature selection.For the interpretability of the model,this paper uses the PDP-ICE curve(partial dependence graph-individual conditional expectation graph),ALE(cumulative local effect graph),SHAP value and LIME model(local surrogate model)to analyze the interpretability of the model,so as to explore the discrimination mechanism of machine learning model and improve the transparency of the model.The empirical research results indicate that the features selected using the GAMI Net model perform better than other feature selection methods in each model,indicating that it can effectively reduce data dimensionality and improve the predictive accuracy of the model.From this,it can be seen that this method can be used for feature selection on high-dimensional data.Secondly,the various interpretable machine learning models used in this article are used for interpretability analysis of credit risk assessment models.The conclusions obtained from the model analysis using various interpretability methods are consistent and consistent with the actual situation,indicating that interpretability methods can be trusted to analyze the results of the model.In practice,the GAMI Net model can be used for feature selection of credit data to reduce data dimensions;At the same time,machine learning interpretability methods can be used to analyze machine learning models and trust the analysis results.