【作者】 赵春；

【导师】 朱群雄；

【作者基本信息】 北京化工大学 ， 控制理论与控制工程， 2012， 博士

【摘要】 随着信息技术的快速发展和管理理论研究取得重大进展，信息技术在企业管理决策领域中的应用受到越来越多的关注。面对残酷的市场竞争环境，企业对风险管理的要求日益提高，如何客观评价企业管理过程中存在的财务风险，并对其进行及时预警是企业管理层始终追求的目标。传统的企业财务风险分析与预警研究方法主要包括统计分析和人工智能模型。随着企业规模的扩大和信息披露越来越频繁，统计分析模型已经不能适应海量数据分析的要求，人工智能模型没有考虑到财务数据的时间延续性。另外，企业财务风险分析与预警研究受企业内外部多种因素影响，不确定性非常高，而数据挖掘技术在不确定性理论研究中的优秀表现让两者紧密联系起来。因此，针对传统方法无法解决的问题，本文深入研究关联规则数据挖掘方法，提出了三种新的关联规则改进型算法，极大提高了数据挖掘的效率；同时，将这些算法应用于企业财务风险分析与危机预警的研究，提出了企业财务风险概念层次树模型和时间序列动态维护的财务危机预警模型。主要研究内容如下：1.基于Hash结构的关联规则交互挖掘算法HIUA现有的关联规则挖掘算法主要基于支持度-置信度框架，同一数据库在不同的支持度和置信度阈值下，算法产生的频繁项集和关联规则的数目是不同的。由于用户事先无法确定合适的支持度和置信度阈值，需要不断尝试不同的阈值才能得到理想的频繁项集和关联规则。本文针对支持度阈值变化时的关联规则维护问题，即当用户调整阈值时存在多次遍历数据库和重复计算问题，提出了基于Hash结构的关联规则交互挖掘算法HIUA，该算法改进了原始IUA算法的剪枝过程，并通过Hash结构快速存取算法执行过程中得到的支持度计数，提高算法运行效率。2.基于部分支持度树的关联规则增量式更新算法IUPS_Miner关联规则的挖掘算法通常假定数据库是静态的，在阈值固定的情况下，如果数据库发生变化，算法需要通过重新进行数据库扫描和计算来得到新的规则。本文针对阈值不变而数据库发生变化时的关联规则维护问题，提出了基于部分支持度树PS_Tree结构的关联规则增量式更新算法IUPS_Miner，该算法只需对新增数据库进行挖掘，通过合并已有的和新增的部分支持度树生成新的部分支持度树，来减少对原数据库的扫描和重复计算，有效地维护了已挖掘的关联规则，提高算法的效率。3.关联规则的动态维护算法ARDM关联规则的动态维护是指当数据库和支持度阈值同时发生变化的情况下，关联规则的维护与更新问题。现有的挖掘方法普遍存在多次扫描数据库或重复遍历复杂数据结构的问题。本文针对数据库与支持度阈值同时发生变化的情况，提出一种基于关联规则交互挖掘和增量挖掘的动态维护算法ARDM，该算法利用已有的挖掘结果进行交互挖掘和增量挖掘，即在原来的数据库中使用新的支持度阈值进行交互挖掘；然后在新增加的数据库中使用新的支持度阈值进行增量挖掘，并通过Hash结构与模式增长方法进行优化，进一步提高算法的效率。4.关联规则交互挖掘算法在企业财务风险分析中的应用企业财务风险分析的研究是通过建立财务风险指标体系，寻找指标体系中具有信任度高的规则，为企业的管理决策提供帮助。传统的财务风险分析方法通常采用统计分析模型，存在的缺点是假设条件多，无法处理海量数据。本文针对上述问题，提出了关联规则交互挖掘的方法，更加广泛的选择多个方面的财务指标，通过挖掘所有财务指标之间的规则，最终确定选择更具有代表性的财务风险指标。首先，构建财务风险分析指标体系，通过变量相关性分析进行指标筛选；然后，提出企业财务风险概念层次树模型，并采用递减支持度阈值的交互挖掘策略，寻找财务风险指标之间的规则；最后，选择国内上市企业中的ST公司进行企业财务风险分析的实证研究，提出了影响企业财务风险的十个关键指标和防范财务风险的建议。5.时间序列动态维护挖掘算法在企业财务危机预警中的应用企业财务危机预警的研究主要是跟踪财务指标波动和变化趋势，当指标波动超出一定的范围，系统就应该发出危机预警。现有的企业财务危机预警的方法主要是基于人工智能数据挖掘模型，存在的缺点是没有考虑到财务指标数据的时间延续性。本文针对财务指标数据具有时间序列特征，提出了基于时间序列动态维护的企业财务危机预警模型。首先，构建时间序列财务数据挖掘模型；然后，采用时间序列增量挖掘和交互挖掘策略，进行关联规则的动态维护挖掘，寻找财务指标之间的规则和预测危机企业的发展趋势；最后，选择国内上市企业中的ST公司进行财务危机预警的实证研究，提出了定性和定量进行财务危机预警的方法，并给出了危机企业不同阶段的关键指标。更多还原

【Abstract】 With the rapid development of Information Technology and significantimprovement of management theory research, Information Technology hasbeen paid more and more attention in the area of enterprise managementdecision. Confronting the fierce competition in the market, the enterprisemakes increasing requirements for risk management, therefore how toevaluate the existing financial risks objectively and timely forewarning in theenterprise management process becomes the goal the enterprise is alwaysseeking for. The traditional methods for financial risk analysis andforewarning include statistical analysis method and neural network model,however with the enlargement of the enterprise in scale and more frequentinformation disclosure, the traditional statistical analysis methods have beenunable to meet the requirement of massive data analysis; also the neuralnetwork model doesn’t consider the time continuity of financial data. Inaddition, the enterprise financial risk analysis and forewarning research isinfluenced by a variety of internal and external factors with high uncertainty,however, the excellent performance of data mining technology in uncertaintytheoretical study makes them closely linked. In order to solve the above problems existed in the traditional methods, after further research on theassociation rule mining methods, we present3improved algorithms based onnew association rules, which improved the mining efficiency greatly. In themeanwhile, we applied them to the enterprise financial risk analysis and crisisforewarning research and proposed both conception hierarchical tree modelfor enterprise financial risks and financial crisis forecasting model based ontime series. The paper is organized as follows.1. Hash based association rule interactive mining algorithm (HIUA)The present association rules mining algorithms are mainly based on thesupport and confidence framework. For the same database, different supportand confidence threshold will generate different frequent item sets anddifferent numbers of association rule with the same algorithm. Since the userscannot in advance know which support-confidence threshold is appropriate,they need to constantly test different thresholds to get the ideal frequent itemsets and association rule. The new algorithm is aimed to improve theassociation rule maintenance issues when the support threshold is changed. Inother words, the previous algorithm involves scanning the database multipletimes and repeated calculation issues while the users adjusted the threshold, soHash based association rule interactive mining algorithm HIUA is proposed,which improves the pruning process of the original IUA algorithm and useHash structure to quickly access to the support counting during the execution.By this way, the algorithm efficiency is improved. 2. Association rule incremental updating algorithm based on PS_Tree(IUPS_Miner)In general, association rule mining algorithm assumes that the database isstatic, in the condition of specifying fixed threshold; it needs re-scan thedatabase to compute new association rules once the database has been updated.Towards the above-mentioned association rule maintenance issues, we presentan efficient association rule incremental updating algorithm based on PS_Tree(IUPS_Miner) that only needs mining the new database. By merging theretained PS_Tree with the new PS_Tree to reduce the cost of scanning theoriginal database and repeated calculation of associations, it efficientlymaintained the previously discovered association rules and improved thealgorithm efficiency.3. Association rule dynamic maintenance algorithm (ARDM)The dynamic maintenance of association rules refers to the maintenanceand update issues of association rules when both the database and supportthreshold are changed at the same time. The present mining methods usuallyhave the problems that involve multiple scanning of database or repeatedlytraversing the complex structure. In this paper, for the situations the databaseand support threshold are changed simultaneously, we present an associationrule dynamic maintenance algorithm based on interactive mining andincremental mining, which uses the already generated associations forincremental mining and interactive mining. Basically the algorithm applies interactive mining to the original databases and then applies incrementalmining to the new added databases using the new support threshold;furthermore, the efficiency is optimized and improved with Hash structure andpattern growth methods.4The application of association rule interactive mining algorithms inenterprise financial risk analysisThe objective of the enterprise financial risk research is to constructingfinancial risk index system, and then determines the high support patterns inthe index system to help for the enterprise management decision. Thetraditional method for enterprise financial risk is usually based on statisticalanalysis model, which drawback is many assumptions so that it cannot processmass data. Aim at the above-mentioned problems, association rule interactivemining is proposed in this paper, which choose multiple wider ranges offinancial indexes first, and then ultimately determine the most representativefinancial risk indicator by mining the rules between all financial indexes. Thedetailed steps can refer to the following: firstly establish financial risk indexsystem in which the selection of financial index is based on variablecorrelation analysis; then build a risk conception hierarchical tree to find therules between the finance risk indicators with interactive mining strategies ofdecreasing support threshold; finally, select the ST companies in the domesticlisted companies for the empirical research of enterprise financial risk analysis,and propose10key indicators that influence the enterprise financial risk and suggestions for avoiding financial risks.5. The Application of association rule dynamic maintenance mining algorithmin financial crisis forecasting.The research of financial crisis forecasting mainly focuses on tracking thefinancial index fluctuations and trends, and the system is supposed to providewarning alert once the financial index fluctuates beyond a certain range. Theexisting methods for financial crisis forecasting are mainly based on artificialintelligence data mining models, which own the drawback that doesn’t takethe time continuity of the financial index data into account. In this paper,considering the time series characteristic for the financial index, we present adynamic-maintained enterprise financial crisis forecasting model based ontime series. The concrete steps refer to the following: firstly construct thefinancial data mining model based on time series; then based on time seriesincremental mining and interactive mining mechanism, find the rules betweenthe financial index and predict the development trend for the crisis enterprisewith association rule dynamic maintenance mining; finally, we select the STcompanies in the domestic listed companies for the empirical research ofenterprise financial crisis forecasting, and determine the key indicator todefine different phase of crisis enterprise.更多还原