"Integrating Data Envelopment Analysis, Partial Least Squares and Artificial Intelligence Approaches for Risk Management in Financial Decision Domains"

在财务决策领域的风险管理方面已经进行了大量的研究。其巨大的经济意义使风险管理研究成为最具挑战性的研究课题之一。本研究的短期目标是整合数据包络分析（DEA）、偏最小二乘（PLS）和人工智能方法来改进现有的风险管理模型。从长远来看，本研究的目标是提供一种创新和独特的方法来解决应用DEA， PLS和人工智能方法进行风险管理的关键问题。它旨在为金融决策领域的风险管理开辟新的可能性。本研究亦旨在为金融机构提供一套实用的工具，以提升其在信贷评估、贷款组合优化及安全管理等关键业务领域的风险管理能力。本研究的意义可以概括为：首先，为特征选择提供了一个通用的分析框架，为评估特征选择的有效性提供了一个客观的方法。此外，将递归特征消除策略引入到基于pls的特征选择方法中，可以显著提高预测过程，产生具有很高预测精度的模型。其次，本研究将尝试利用基于PLS的特征选择来为DEA模型选择最相关的输入和输出。这将克服DEA方法的一个主要局限性，为DEA的应用开辟许多新的可能性。第三，数据分类不平衡是现实中普遍存在的问题。问题的本质在于一个类在数据集中占主导地位，而不是其他类，这使得分类器为了追求准确性而将所有内容分类到主导类中。本研究提出结合重采样和集成学习进行信息融合，以解决数据分类不平衡的问题。最终，行业客户分析平台将为金融市场，包括机构和监管机构带来显著的效益。

Considerable research efforts have been spent on risk management in financial decision domains. Its large economic significance makes risk management research one of the most challenging research topics. The short term objective of this proposed research is to integrate Data Envelopment Analysis (DEA), Partial Least Squares (PLS) and artificial intelligence approaches to improve the current risk management models. In the long run, the objective of the proposed research is to provide an innovative and unique approach to address key issues in applying DEA, PLS and artificial intelligence approaches to risk management. It aims at opening up new possibilities for risk management in financial decision domains. It is also the target of this research to provide financial organizations with a set of practical tools to enhance their risk management capabilities in critical business areas such as credit assessment, loan portfolio optimization and security management. The significance of this research can be summarized as below: Firstly, a general analytical framework for feature selection provides an objective way to evaluate the effectiveness of feature selection. Furthermore, recursive feature elimination strategy will be introduced to PLS-based feature selection method, which can significantly enhance the prediction process and produce models with very high prediction accuracy. Secondly, the proposed research will attempt to utilize PLS based feature selection to select the most relevant inputs and outputs for DEA models. This will overcome one of the major limitations of DEA method and open up many new possibilities for DEA applications. Thirdly, imbalanced data classification is a prevalent problem in reality. The nature of the problem lies in that one class dominates the data set against other classes, which tricks the classifiers to classify everything into the dominating class in sheer pursuit of accuracy. This research proposes to integrate re-sampling and ensemble learning for information fusion to address imbalanced data classification. In the end, the industrial customer analysis platform will bring significant benefits to the financial market, including institutions and regulators.