多类别非平衡企业信用评估的多SVM集成建模研究

Current research on enterprise credit evaluation is limited to the situation of binary classification, and can not satisfy the practical demand for multi-class enterprise credit evaluation. This project attempts to explore the theoretical and methodological system based on multi-SVM ensemble for enterprise credit evaluation from the new view of multi-class and class imbalance. This project firstly explores the relevant connotation, multi-dimensional indicators and theoretical framework for multi-class imbalanced enterprise credit evaluation. Then it designs the decomposition and aggregation approach for multi-class enterprise credit evaluation. In addition, it integrates the data-level balancing approaches with multi-SVM ensemble approaches to construct the multi-SVM ensemble models for imbalanced enterprise credit evaluation. Furthermore, it combines the above methods with the nearest neighbors method for sample types analysis to propose the multi-SVM ensemble modeling methodology for multi-class imbalanced enterprise credit evaluation. Finally, this project carries out comparative empirical studies to test the effectiveness of the above enterprise credit evaluation modeling methods in situation of multi-class and class imbalance, based on the multi-dimensional data collected from listed companies and the modified performance metrics. This research can promotes enterprise credit evaluation modeling to move from binary classification and prediction to multi-class classification and prediction. Therefore, it can help improve the theoretical and methodological system for enterprise credit evaluation and provide the important decision support tools for the practice of enterprise credit evaluation.

当前的企业信用评估研究局限于二分类情形，不能满足多类别企业信用评估的现实需求。本项目从多类别非平衡的全新视角探索企业信用评估的多SVM集成建模理论与方法体系，并展开实证检验。在探索多类别非平衡企业信用评估相关内涵、多维指标体系和理论框架的基础上，设计多类别企业信用评估的分解融合方法，并将数据层面的平衡化处理方法与多SVM集成方法相结合来构建非平衡企业信用评估的多SVM集成模型，进而融合样本类型分析的近邻法来构建多类别非平衡企业信用评估的多SVM集成建模方法体系。基于多维度上市公司数据和改进的模型性能评价指标，开展多类别非平衡情形下企业信用评估建模方法的实证比较研究，以检验方法的有效性。该研究将企业信用评估建模从二分类预测推向多类别分类预测研究，有利于促进企业信用评估建模理论和方法体系的完善，并为企业信用评估实务提供重要的决策支持工具。

以往基于机器学习的企业信用评估研究主要局限于二分类情形，忽略了多类别企业信用评估建模及实证研究。为了克服理论研究的不足，以满足企业信用评估更加精细化的现实需求，本项目从多类别非平衡的视角探索企业信用评估的多分类器集成建模理论与方法体系，并展开实证检验。首先，对多类别非平衡企业信用评估的理论基础展开研究，将上市公司信用状态分为信用真良好、信用伪良好、信用危机三类，并构建了包含财务指标与非财务指标的多类别企业信用评估指标体系，在此基础上提出了多类别非平衡企业信用评估的基本思路框架。其次，对多类别非平衡企业信用评估的多分类器集成建模方法体系展开探索，在分析多类别企业信用评估的分解融合方法与非平衡企业信用评估的多分类器集成方法的基础上，提出了适合于多类别非平衡企业信用评估的基于分解融合策略的SMOTE-Bagging集成模型与基于分解融合策略的SMOTE-Adaboost集成模型。最后，基于上市公司样本数据和改进的模型性能评价指标，开展多类别非平衡企业信用评估的实证比较研究，发现本项目提出的多类别企业信用评估模型具有较好的测试效果。该研究将企业信用评估建模从二分类预测推向多类别分类预测研究，有利于促进企业信用评估建模理论和方法体系的完善，并为企业信用评估实务提供重要的决策支持工具。