数据特征驱动下基于半监督学习的小微企业贷款信用风险预测研究

Small and micro enterprises are important force to promote the rapid development of China's economy. In view of the fact that the promotion and perfection of loan credit risk evaluation system plays an important role in improving the financing ability of small and micro enterprises, this project intends to carry out the research on the credit rating methods of small and micro enterprises by using semi-supervised learning methods based on the idea of data feature-driven modeling to improve the forecasting performance of credit risk. First of all, the data characteristics are systematically combed, and a semi-supervised kernel-free support vector machine model combined with the data of historical loan-applying enterprises and the enterprises to be applied for loans is constructed to reveal the effectiveness and robustness of semi-supervised learning methods for the credit rating of small and micro enterprises. On this basis, 1) the semi-supervised kernel-free support vector machine models based on probability are constructed for solving the problem of data noise and feature loss; 2) the semi-supervised support vector machine models combined with deep learning are constructed in view of the diversity of credit rating categories and the characteristics of many data features; 3) in view of the characteristics of data type diversity, a new data feature quantization method is proposed, and then the semi-supervised support matrix machine models are constructed. Finally, distributed algorithms for large-scale data are designed to solve these models. In this project, some data feature-driven semi-supervised learning methods are proposed, which provide new ideas for data mining and forecasting theories and methods. This project also provides decision support for domestic financial institutions to achieve accurate forecasting of small and micro enterprise loan credit risk.

小微企业是推动我国经济快速发展的重要力量。鉴于贷款信用风险评价体系的推进和完善对提高小微企业融资能力有重要作用，本项目拟基于数据特征驱动建模的思想，开展基于半监督学习的小微企业信用评级方法研究，以提高信用风险的预测精度。首先，系统梳理数据特征，构建结合历史申贷企业和拟申贷企业数据的半监督无核支持向量机模型，揭示半监督学习在小微企业信用评级中的有效性和鲁棒性。在此基础上，1）针对数据存在噪音和特征缺失的问题，构建基于概率的半监督无核支持向量机模型；2）针对信用评级类别多样性和数据特征多的特点，构建半监督深度支持向量机模型；3）针对数据类型多样性的特点，提出新的数据特征量化模式，构建半监督支持矩阵机模型。最后，设计面向大规模数据的分布式算法实现模型的求解。本项目提出数据特征驱动的半监督学习方法，为数据挖掘和预测理论与方法提供新思路，为国内金融机构实现小微企业贷款信用风险精准预测提供决策支持。

近年来，社会信用体系建设成为推动国家治理体系和治理能力现代化的重要环节，信用风险评估作为其关键任务受到广泛关注。考虑到小微企业信用风险评估数据通常存在噪声点、异常点、类别不平衡等问题，构建数据特征驱动的分类模型并设计面向大规模数据的快速求解算法是提升分类性能的重要因素。本项目以支持向量机方法作为基础，研究数据特征驱动的信用评级全监督和半监督分类方法，进而提高信用风险预测能力。首先，针对数据存在噪声点和异常点的问题，一方面，提出基于Universum先验信息的带有模糊隶属度函数的无核支持向量机分类方法和加权的无核支持向量机分类方法，另一方面，考虑数据集非线性可分情形，提出基于数据特征映射的深度支持向量机分类方法。其次，针对数据集存在类别不平衡的问题，建立二分类和多分类模型，具体包括提出考虑多种度量目标的无核最小最大概率机二分类方法和基于模糊隶属度函数的一次性多分类方法。最后，为进一步提高分类效率，提出无核拉普拉斯双子支持向量机分类方法。另外，研究信用评级相关领域中实际问题，为探究数据特征构造提供支持。项目执行期间共发表SCI收录学术论文8篇，培养研究生9名，参加学术会议12次，会议报告3次。