基于多源域知识迁移学习的小微企业信用评分

Credit Scoring of Small and Micro Enterprises Using MultisourceInformation Transfer Learning

针对新业务,新场景下金融机构目标数据集“高维小样本”的问题,本文提出了基于多源域知识迁移学习的小微企业信用风险测度方法,其能够迁移学习其它数据源(源域)的知识以提升目标域模型的预测效果.该方法通过对来自多个源域的多种源域知识进行归纳提取,进而将其纳入目标域模型的构建中,可以充分利用源域知识,提升目标域模型的估计精度.另外,模型无需获取各源域的原始数据,因此很大程度上降低了数据传输中隐私泄露的风险.模拟实验和企业信用评分的实例数据验证了所提方法的可行性及其在变量选择,系数估计和分类预测上的良好效果.该方法能够在隐私限制的背景下有效迁移源域知识以克服信用评分中目标数据集信息量不足,而导致估计效果较差的问题.

When building credit scoring models for new products or businesses,financial institutions often encounter the“high-dimensional,small samples”problem which results in unsatisfactory model performance.We propose a credit scoring method for small and micro enterprises based on a multi-source transfer learning technique.This method can transfer knowledge from other data sources(source domains)to improve the prediction performance of the credit scoring model of the target domain.Specifically,we first extract the multi-form knowledge from each source domain and then incorporate the information into the building process of the target domain model.The proposed method can take full advantage of the knowledge from source domains,and improve the prediction accuracy of the target domain model.In addition,throughout the modeling process,there is no need to obtain the original data of each source domain,which greatly reduces the risk of privacy leakage during data transmission.Simulation studies and real data analysis illustrate the superior performance of the method on variable selection,estimation,and prediction aspects.This method can effectively transfer the source domain information under privacy-preserving constraints to overcome the”high-dimensional,small samples”problem in the target data set.