基于半监督学习和信息融合的港口客户信用风险评价系统

发布时间：2017-12-30 20:31

本文关键词：基于半监督学习和信息融合的港口客户信用风险评价系统　出处：《北京交通大学》2017年博士论文　论文类型：学位论文

【摘要】：伴随着经济一体化、全球化的发展趋势,我国国民经济的发展和对外贸易迅速增加,造就了水路运输的快速增长,也推动了港口的发展。为了维系和争取更多的客户,港口扩大了信用结算的适用范围。随着信用结算政策的改变,客户信用风险问题开始不断困扰着港口的管理者。客户延期还款甚至恶意拖欠,严重影响了港口正常的经营。而传统依靠人工的客户信用风险评价方法,已经难以满足港口日常经营管理的需求。因此,如何利用现有资源,增强港口信息化建设和应用水平,对港口客户未来一段时间内的信用风险水平进行评价,从而降低或规避因客户信用风险给港口带来的损失,提高港口的应变能力,是港口目前亟待解决的问题。本文面向港口客户信用风险评价,以广东省教育部产学研项目《广州港集团生产业务管理系统及通用软件产业》(2008B090500244)、《基于RFID的港口汽车滚装管理系统应用示范工程》(2009B090300467)和国家自然科学基金重点项目《物流资源整合与调度优化研究》(71132008)等为支持,深入分析了港口客户信用风险的成因,综合应用半监督学习、主动学习、信息融合、神经网络和遗传算法等理论和方法,设计并构建了基于半监督学习和信息融合的港口客户信用风险评价系统,主要研究内容和成果如下:(1)提出了面向港口的客户信用风险评价体系本文在对港口客户信用风险成因深入分析的基础上,比较了港口客户信用评价体系和现有主要信用评价体系之间的差异。针对目前相关研究不足、套用现有客户信用风险评价体系难以满足实际需求等问题,明确了构成评价体系的指标,引入外部影响因素,构建了港口客户信用风险评价体系。(2)构建了基于Tri-Training和标签传递算法的半监督文本倾向分类框架本文对有标签样本不足情况下的文本倾向分类进行了研究。针对实际应用中有标签样本不足,影响文本倾向分类性能的问题,引入了半监督学习算法。在深入了解半监督学习的基础上,针对标签传递算法无法直接处理样本外数据和Tri-Training算法易受初期噪音干扰的问题,提出了一种结合Tri-Training与标签传递算法的半监督文本倾向分类框架(Label-propagation Improved Tri-TrainingFramework,LIT2)。(3)提出了一种基于主动学习的半监督文本倾向分类优化策略针对LIT2在训练过程中出现的学习能力瓶颈,有针对性地提出了一种基于主动学习的优化策略。采用成员查询式的主动学习优化策略,通过构建训练信息较丰富的有标签样本,帮助LIT2克服训练前期的学习能力瓶颈;通过基于池的主动学习,选取具有较高训练价值的样本,从而令LIT2克服训练后期的学习能力瓶颈,综合提高LIT2的训练效率和分类性能。(4)提出了基于内外部信息融合的港口客户信用风险评价模型对港口客户信用风险评价的过程,也是融合港口企业内外部信息的过程。结合信息融合模型和BP神经网络,本文提出了基于内外部信息融合的港口客户信用风险评价模型(Internal and External Information Fusion based Port Customer Credit Evaluation Model,IEPCCM)。针对BP神经网络中的不足,提出了一种基于多项改进的 BP 神经网络构建方法(Mutil-Improved BP-NN Model Construction Method,M2C),并应用 M2C 构建了IEPCCM。(5)实现了 SIPCC的原型化开发在前文研究的基础上,根据基于内外部信息的港口客户信用风险评价系统(Semi-supervised Learning and Information Fusion based Port Customer Credit Evaluation System,SIPCC)的需求分析和体系框架,应用Java EE平台,Spring-Hibernate联合框架,结合Nutch等关键插件,完成了 SIPCC系统原型系统的开发,并实现了内部数据管理、外部信息抽取、文本倾向分析和客户信用评价等核心功能。通过多种技术框架体系和关键技术的联合应用,进一步提高了系统的高效性、可扩展性、实用性和安全性。
[Abstract]:Along with the economic integration, the development trend of globalization, the development of China's national economy and foreign trade increased rapidly, creating a rapid growth of water transport, but also promote the development of the port. In order to maintain and attract more customers, expand the scope of port credit settlement. With the credit settlement policy changes, the problem of customer credit the risk began to haunt the port managers. Customers deferred repayment even malicious default, seriously affected the normal operation of ports. While the traditional credit risk evaluation methods rely on artificial, has been difficult to meet the needs of the port daily management. Therefore, how to use existing resources, strengthen port information construction and application level, evaluation the credit risk level of port customers over a period of time, so as to reduce or avoid the credit risk to the port loss, improve The strain capacity of port, port is the urgent problem to be solved. In this paper, for the port customer credit risk assessment in Guangdong Province, the Ministry of education and research project "Guangzhou port group business management system and general software industry (2008B090500244) >, < RFID port ro ro management system application demonstration project > based on (2009B090300467) and key" the project of National Natural Science Foundation of logistics resource integration and scheduling optimization research "(71132008) as the support, in-depth analysis of the causes of the credit risk of the port, the comprehensive application of semi supervised learning, active learning, information fusion, neural network and genetic algorithm theory and method, the design and construction of the port customer credit risk evaluation system supervised learning and based on information fusion, the main research contents and results are as follows: (1) proposed for the port customer credit risk evaluation system in Hong Kong Based on in-depth analysis of export credit risk causes, the difference between the port customer credit appraisal system and the existing credit evaluation system. Aiming at the problems related to research and apply the existing credit risk evaluation system to meet the practical needs and other issues, the structure of the evaluation system of indicators, external influence factors are introduced, and build a system port customer credit risk assessment. (2) constructed the semi supervised text label propagation algorithm Tri-Training and the tendency of the classification framework with a text label samples in case of insufficient in research of classification based on the labeled samples. Aiming at the shortage in application, effect of text tendency classification performance, introduces the semi supervised learning algorithm. Based on in-depth understanding of semi supervised learning, according to the label propagation algorithm cannot deal with the sample data directly and Tri-Training The algorithm is easily affected by the initial noise problem, we propose a combination of Tri-Training and label propagation algorithm of semi supervised text tendency classification framework (Label-propagation Improved Tri-TrainingFramework, LIT2). (3) proposed a semi supervised learning bottleneck text active learning tendency classification optimization strategy for LIT2 in the training process based on appearance and puts forward an optimization strategy based on active learning. The members of the active learning query optimization strategy, through the construction of training information is abundant labeled samples, help LIT2 overcome the bottleneck learning ability training early; active learning through pool based on the selection of higher value of training sample, so that LIT2 learning ability training to overcome bottleneck stage, improve LIT2 training efficiency and classification performance. (4) proposed and based on information fusion The process model of port customer credit risk evaluation of port customer credit risk assessment process, but also the integration of internal and external information of the port enterprise. Combined with information fusion model and BP neural network, this paper puts forward the internal and external information fusion of port customer credit risk evaluation model based on (Internal and External Information Fusion based Port Customer Credit Evaluation Model, IEPCCM) the BP neural network, this paper proposes a method to construct a BP neural network based on improved (Mutil-Improved BP-NN Model Construction Method, M2C), and the application of M2C to construct the IEPCCM. (5) implementation of the SIPCC prototype developed on the basis of previous studies, according to the port credit risk assessment system of internal and external customers based on the information of Learning (Semi-supervised and Information Fusion based Port Customer Credit Evaluation System, SIPCC) the demand analysis and the system framework, the application of Java EE platform, Spring-Hibernate framework, combined with Nutch and other key plug-in developed SIPCC prototype system, and realizes the internal external data management, information extraction, text orientation analysis and customer credit evaluation and other core functions. Through a combination of framework and key technology, to further improve the efficiency of the system, expansibility, practicality and safety.

【学位授予单位】：北京交通大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：F552.3

【参考文献】