基于类别非平衡时序增量数据批的多SVM动态集成企业信用评估建模

To further improve the effectiveness and applicability of enterprise credit scoring models, this project attempts to implement theoretical and empirical studies on the credit scoring modelling methods based on multi-SVM (support vector machine) dynamic ensemble from the new view of class-imbalance and time series incremental data batches. On the basis of exploring the basic theory of class-imbalance and time series incremental data batches in enterprise credit scoring modelling, resampling techniques and multi-SVM ensemble methods are integrated to propose the multi-SVM ensemble approaches based on a single imbalanced data set for credit scoring modelling. In addition, the sliding data batch windowing method and the data batch weighted sampling method are put forward and combined with multi-SVM ensemble methods, to bring forward the multi-SVM dynamic ensemble approaches based on time series incremental data batches for credit scoring modelling. Furthermore, the above key techniques are integrated through the two-stage integration mechanism and the embedding integration mechanism to construct the multi-SVM dynamic ensemble approaches based on class-imbalanced time series incremental data batches for credit scoring modelling. Furthermore, comparative empirical studies are to be carried out with the sample data collected from Chinese listed companies and UCI Machine Learning Repository to validate the effectiveness of these approaches. Theoretically, this research helps promote a further innovation by integrating the class-imbalance treating methods and the time series incremental dynamic learning algorithms. On the other hand, it helps practitioners construct enterprise credit scoring models that are better suitable for the current credit scoring characteristic trend and have better recognization capability.

为了提高企业信用评估的有效性和可用性，本课题尝试从类别非平衡时序增量数据批的新视角来研究多SVM动态集成企业信用评估建模方法体系，并展开实证检验。在探索企业信用评估建模中的类别非平衡时序增量数据批理论基础的前提下，将重抽样技术和多SVM集成方法融合来研究单个类别非平衡数据集上的多SVM集成建模方法，提出移动数据批窗口与数据批赋权抽样思想并与多SVM集成方法融合来研究基于类别平衡时序增量数据批的多SVM动态集成建模方法，进而通过两阶段融合机制和嵌入式融合机制整合前两者的关键技术来构建基于类别非平衡时序增量数据批的多SVM动态集成企业信用评估建模方法。基于上市公司数据和UCI数据开展实证比较研究，以检验方法的有效性。该研究将在理论上促进类别非平衡处理方法和时序增量动态学习方法的融合拓展创新，同时在实务方面有助于建立更加符合当前时序特征变化趋势并能有效识别的企业信用评估模型。

基于企业信用评估中违约样本远远少于正常样本的类别非平衡现象，以及时间推移过程中企业信用评估呈现的概念漂移现象，为了提高企业信用评估的有效性和可用性，本项目从类别非平衡时序增量数据批的新视角来研究多SVM动态集成企业信用评估建模方法体系，并展开实证检验。本项目首先对企业信用评估建模中的类别非平衡时序增量数据批理论基础展开研究，分析了考虑类别非平衡时序增量数据批的企业信用评估特征因素，构建了基于类别非平衡时序增量数据批的企业信用评估内涵与框架。其次，对基于类别非平衡时序增量数据批的多SVM动态集成企业信用评估建模的方法体系展开研究，具体包括单个类别非平衡数据集上的多SVM集成企业信用评估建模方法、基于类别平衡时序增量数据批的多SVM动态集成企业信用评估建模方法、基于类别非平衡时序增量数据批的多SVM动态集成企业信用评估建模方法。将重抽样技术、多SVM集成方法以及滑动时间窗口与数据批时序赋权思想有机融合，构建了体系相对完整的基于类别非平衡时序增量数据批的企业信用评估多SVM动态建模方法理论模型。最后，基于中国上市公司数据开展实证比较研究，并对各个方法的性能指标展开统计检验，验证了所提方法的有效性。该研究将在理论上促进类别非平衡处理方法和时序增量动态学习方法的融合拓展创新，同时在实务方面有助于建立更加符合当前时序特征变化趋势并能有效识别的企业信用评估模型。