节点文献
基于知识管理的智能型贷款风险分类研究
【作者】 朱明；
【作者基本信息】 东华大学 ， 控制理论与控制工程， 2002， 博士
【摘要】 信贷风险管理是商业银行的永恒主题，尤其在东南亚经济危机之后，受到各国广泛关注。我国政府也十分重视，把提高信贷资产质量、防范信贷风险作为金融控制的目标，并倡导采用现代化信息技术手段进行风险防范控制研究。为了进一步加强金融监管，我国于1998年开始提出了与国际通行做法协调的贷款风险五级分类方法。本文正是适应这种需要，并结合国家自然科学基金项目《基于神经网络与专家系统结合的银行贷款风险管理研究》的研究任务，提出以知识管理为基础、人工智能技术为手段、面向贷款风险分类，进行信贷风险管理和控制研究。 本文通过对知识描述形式的研究，表述了经验符号性知识、模型化数量性知识和实例样本性知识三种类型，通过对其行为特征及转换、集成的研究，将这三种异构知识引入贷款风险分类分析；采用面向对象技术和人工智能的ES、ANN两种形式及其结合，实现了贷款风险五级分类。 本论文的研究内容及具体工作成果如下： 1．在认识论层次上统一了信息和知识的定义，提出了知识管理概念。按人类知识结构化程度将知识分为三种描述形式，这样的知识描述形式之间互相转换、互相渗透，能够实现知识互补性，强化了知识的复用能力，并将基于这三种知识描述形式用于贷款风险分类模型。 2．在贷款风险分类的财务因素分析时，银行有着较充足的各行各业的财务报表的样本实例，可以通过计算的各种比率结果及与同行业的比较而对借款人作出正确的判断，因此它具有模型数量性知识和实例样本性知识。由于神经网络具有容错性、健壮性和自学习的特点，可以通过学习来获得隐含在样本中的知识，为此，提出了基于BP神经网络的企业还款能力分析方法。我们用VC++实现了神经网络模型，并将该模型实际应用到还款能力的财务分析中。研究了财务指标的预处理问题、网络隐含层神经元数选取问题、网络连接权值的初值选取问题等，给出了17×11×5的网络拓扑结构，结果表明这种方法进行财务分析比常规方法有效，提高系统的应用效果。 3．贷款风险分类中非财务因素属于典型的符号性知识。由于基于知识系统是利用知识和推理过程来解决需要人类专家才能解决的问题的计算程序，它具有一般数据处理程序所不具备的优点，可以处理符号知识，利用启发式知识降低搜索复杂性，并具有提供良好的解释和吸收新知识的能力。我们通过面向对象方法对非财务因素的分析，给出了系统各个决策要素，编制了系统关联图，设计和组建了相应的知识库，采用面向对象的VC++语言实施了基于知识(KBS)的非财务因素原型系统。 4.通过ANN与ES的结合，实现了实例样本性知识与经验符号性知识的转换及知识求精。对于知识转换，提出了先通过ANN接受样本训练学习，然后抽取规则，转换为符号性知识的方法。这样既克服了单一专家系统知识获取的“瓶颈”，又避免了神经网络的“黑箱问题”;对于知识求精，先把已经得到的规则转换为网络拓扑结构，利用神经网络对知识进行求精，这样可解决初始知识库存在的知识不完全、知识之间不一致、有的知识不正确等问题，改善管理决策的智能水平和增强了系统的运行效能。 5.将贷款风险分类中五个因素所归结的三种异性知识，提出通过ES集成进行综合分析的方法，得出了五类分类。通过贷款风险分类的实践，可以实现对知识的获取、转换、求精、集成，从而达到知识的有效管理，促进整个银行的知识能够广泛的共享和适度的应用。籍此，可以进一步提高贷款风险管理的质量和金融机构信贷管理水平。 本文主要的新见解与创新之处表现在以下:首先把知识管理概念从企业管理层面脱离出来，通过运用知识管理，使知识的表达更丰富、精炼，从而提高了系统的智能化水平，促进了人工智能的研究从理论转向更深入的应用，这样的系统十分适合于经济管理、控制与决策。本文首次提出了将知识管理的思想、人工智能手段应用于贷款风险五级分类。通过运用人工智能手段及贷款风险分类中知识管理，所得的分析结果比以往传统的常规分析更有效，结果更准确，且使银行贷款风险管理具有智能化特征。本文中采用的三类知识的表示、知识转换、求精与集成，除用在银行风险管理之外，还可适合于复杂系统的深层知识表达。 鉴于有关贷款风险五级分类，今年中央又强调2002年1月开始正式在中国银行业全面推行贷款风险分类管理，所以，本文“基于知识管理的智能型贷款风险分类研究”的研究有其实际意义和一定应用前景。
【Abstract】 The credit risk management is a permanent topic in commercial bank to which has been paid attention by many foreign governments especially after the Asia Finance crisis’ broken out. Our government also has clearly asked the commercial bank to improve the quality of credit assets and prevent credit risks. Meanwhile our government has advocated carrying out the study of loan risk management by modern information technology. To strengthen the financial supervision, "The principle of loans risk classification" was published in 1998 and the real risk management is carried out in china. According to the requirement, the paper combined with the task-"the research of loans risk management based on the integration of ANN (Artificial Neural Network) and ES (Expert System)" which is supported by National Science Foundation of China. Set forth using AI (Artificial Intelligence) technology on the basis of Knowledge Management to research the loans risk management and control.Through the study of the style of knowledge description, the paper elaborates the numerical model knowledge, the symbolic experience knowledge and the instantial swatch knowledge. These three kinds of different knowledge have been used in loans risk classification in the paper. Just as the using of the Object Oriented technique and the AI (include ES, ANN, and the integrated of ANN and ES) technique, can the paper realize the five-grade loans risk classification.In this paper, the study content and the concretely achievement are as follows:1. Unifying the definitions of information and knowledge on the layer of epistemology, we put forward the concept of knowledge management. According to human’s knowledge structure, we divide knowledge by three kind of description. Such interchange, permeation among the descriptions can make the knowledge to be shared together and enforce the reuse ability of knowledge. These knowledge descriptions can be used in the model of loans risk classification.2. On the analysis of financial affairs in classifying loans risk, bank has large of finance report forms samples. It can make correct judge by calculating various rates and comparing them with other enterprise. So the bank has the numerical model knowledge and the instantial swatch knowledge. As neural network has the features of fault-toleration, robust and self-learning, it can acquire the knowledge hidden in the samples through learning. Thus, we provide a kind of method realized by VC++ to analyze the ability of paying back loans of enterprises based on BP neural network. In the method, we have discussed the pretreatment of financial indexes, the choice of the number of hidden-unit, the initial weight and so on, and got a 17 × 11 × 5 topology structure. The results have proved that this method is more effective than usual financial analysis.3. In the loans risk classification, non-financial affairs belong to typical symbolic knowledge. As Knowledge-Based System (KBS) is such a calculating procedure utilizing knowledge and reckon to resolve the questions only can be resolved by human specialists that it is has advantages other procedures do not have to treat symbolic knowledge. It can reduce searching complexity by enlightening knowledge. We have analyzed the structure of non-financial affairs, established system dependency diagram and designed the Knowledge Base. At last we have built up the non-financial affairs prototype system based on KBS by VC++.4. Integrating ANN with ES, we have brought about the change between the symbolic knowledge and the instantial knowledge. The importance of rule extraction from trained ANN is that of using the ANN for the ’Learning’ of impliedly rules within swatch knowledge, and then expressing with rules. The goal in rule refinement is to use combination of ANN learning and rule extraction techniques to produce a ’better’ set of symbolic-rules which can then be applied back in the original problem domain. In the rule refinement process, the initial rule base is inserted into an ANN by programming some of the