基于非平衡数据的LDA-BPNN信用评分模型  

LDA-BPNN Credit Scoring Model Based on Unbalanced Data

作　　者：王静 王艳丽 孙士保 贾少勇 WANG Jing;WANG Yan-li;SUN Shi-bao;JIA Shao-yong(College of Information Engineering,Henan University of Science and Technology,Luoyang Henan 471023,Chinal;College of Software,Henan University of Science and Technology,Luoyang Henan 471023,China)

机构地区：[1]河南科技大学信息工程学院,河南洛阳471023 [2]河南科技大学软件学院,河南洛阳471023

出　　处：《计算机仿真》2023年第2期303-308,414,共7页Computer Simulation

基　　金：国家自然科学基金项目(51474095);河南省重点攻关项目(152102210277);河南省高校科技创新团队支持计划项目(17IRTSTHN010);河南科技大学科技创新团队项目(2015XTD011);河南科技大学重大产学研合作培育基金项目(2015ZDCXY03)。

摘　　要：随着互联网金融的蓬勃发展,信用评分已成为评判是否为贷款人放款的重要依据,传统的信用评分方法的单一性,不能有效地减少不良贷款带来的巨大损失。提出一种基于非平衡数据的LDA-BPNN方法构建信用评分模型,通过Borderline-SMOTE算法进行非平衡数据过采样,达到平衡样本分布,进一步构建LDA-BPNN模型,显著提升了模型的预测准确率。在不同数据集上,利用多类机器学习算法进行定性和定量对比,并通过pr曲线、AUC等评价指标进行验证,验证结果表明上述模型的分类效果更佳。With the booming development of Internet finance,credit score has become an important basis for judging whether the lender is lending money.The singleness of traditional credit score methods cannot effectively reduce the huge loss caused by non-performing loans.Therefore,an LDA-BPNN method based on unbalanced data was proposed to build a credit scoring model.The Borderline-SMOTE algorithm was used to oversample the unbalanced data to achieve balanced sample distribution.The further construction of the LDA-BPNN model significantly improved the prediction accuracy of the model.Qualitative and quantitative comparisons were made by using multiple machine learning algorithms on different data sets,and verified by pr curve,AUC and other evaluation indexes.The verification results show that the classification effect of this model is better.

关 键 词：信用评分 自适应合成抽样算法 线性判别分析 多层前馈神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]