个人信贷违约预测机器学习模型的解释性方法研究  

Research on interpretative methods of machine learning models for personal credit default prediction

作　　者：陈玉沂 刘高勇[1] 蔡焕仪 Chen Yuyi;Liu Gaoyong;Cai Huanyi(School of Management,Guangdong University of Technology,Guangzhou 510520,China)

机构地区：[1]广东工业大学管理学院,广州510520

出　　处：《现代计算机》2024年第13期68-72,共5页Modern Computer

基　　金：广东省大学生创新创业计划项目(S202311845168)。

摘　　要：近年来个人信贷业务需求量激增,金融机构利用机器学习模型对客户进行信贷违约预测,预测结果的可解释性影响着金融机构的决策。首先基于机器学习模型LightGBM、XGBoost、CatBoost构建个人信贷违约预测模型,然后通过超参数优化和Voting投票融合方法提升了模型的性能,最后采用置换特征重要性、LIME、SHAP和反事实解释四种解释方法,从全局和局部层面对模型预测结果进行解释性分析,提高了模型的可信度和实用性。In recent years,as the demand for personal credit business has surged,financial institutions predict credit default with machine learning models.The interpretability of the prediction results is so important that influences the decisionmaking of financial institutions.Firstly,a personal credit default prediction model is constructed based on machine learning models LightGBM,XGBoost and CatBoost.Then,the model is experimentally optimized by using hyperparameter optimization algorithms and Voting fusion methods.Finally,the model prediction results are interpretively analyzed globally and locally through four interpretative methods,including Permutation Feature Importance,LIME,SHAP and Counterfactual Interpretation,which greatly enhance the reliability and practicability of the model.

关 键 词：金融信贷 机器学习 可解释性 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程] F830.5[自动化与计算机技术—控制科学与工程]