平均一依赖估测算法在个人信用评估中的研究

发布时间：2018-07-20 15:02

【摘要】：随着中国市场经济的高速发展,以个人的信用作为担保向银行贷款正在成为一种趋势。信用贷款数量以及贷款金额的不断增加,使得银行面临客户违约所造成的风险也在逐年增加,因此,在向客户发放信用贷款前,银行必须根据客户的真实信息客观准确地对其信用进行评估,根据信用情况发放贷款,从而降低由于客户违约而给银行带来的损失。本文分析了样本数据集和平均一依赖估测(Averaged One-Dependence Estimators,AODE)模型结构特点,针对样本数据集中的连续属性,首先提出了一种离散化方法,连续属性经过离散化处理后,能够有效地提高AODE模型的分类精度;其次根据粗糙集理论的属性约简规则,在保持信息系统分类能力不变的条件下,删除其中不相关或不重要的指标属性,从而筛选出能表示样本的最小指标集;最后将Adaboost与AODE模型相结合构建集成AODE分类器,构建个人信用评估模型。本文主要工作如下:(1)为了解决样本数据集中的连续属性离散化问题,提出一种改进的的离散粒子群优化算法。将连续属性的断点集合作为离散粒子群,通过粒子间的相互作用最小化断点子集,同时引入模拟退火算法作为局部搜索策略,提高了粒子群的多样性和寻找全局最优解的能力;利用粗糙集理论中决策属性对条件属性的依赖度衡量决策表的一致性,从而达到连续属性离散化的目的。(2)针对样本数据集中大部分指标属性存在冗余且不具备同等重要性,不利于在数据分析中做出简明的决策,对样本数据集的指标属性进行约简是信用评估的重要步骤。本文提出一种基于禁忌离散粒子群优化的属性约简算法对个人信用评估指标进行选取。由于禁忌搜索算法对初始解有较强的依赖性,而离散粒子群算法在迭代时容易陷入局部最优解,因此在指标选取过程中,采用离散粒子群算法在全局进行搜索,禁忌搜索算法在局部进行寻优,在不影响分类质量的前提下,删除冗余属性,简化知识库,构建个人信用评估指标集合。(3)AODE模型在进行分类时,组成它的每一个超父独依赖估测模型(Super Parent One-Dependence Estimator,SPODE)对分类的贡献程度是一样的,然而每一个SPODE模型的结构不同,对最终分类结果的影响也不同。本文针对平均一依赖估测算法的结构弱点提出了相应的改进。首先从构成它的每一个超父独依赖估测模型中,采用随机抽样法,选取一定数量的SPODE模型组成平均一依赖估测模型;然后采用Adaboost算法构建集成AODE分类模型;最后将集成AODE分类模型用于个人信用评估。仿真实验结果表明,集成AODE评估模型能够有效地提高个人信用评估的预测准确率。
[Abstract]:With the rapid development of China's market economy, it is becoming a trend to take personal credit as guarantee to lend to banks. The number of credit loans and the amount of loans are increasing, which makes banks face the risk of customers defaulting. Therefore, before issuing credit loans to customers, The bank must evaluate its credit objectively and accurately according to the customers' true information, and make loans according to the credit situation, so as to reduce the losses caused by the customers' default. In this paper, the structural characteristics of sample dataset and Averaged One-Dependence estimation (Aode) model are analyzed. For the continuous attributes of the sample dataset, a discretization method is proposed, and the continuous attributes are discretized. It can effectively improve the classification accuracy of AODE model. Secondly, according to the attribute reduction rules of rough set theory, the irrelevant or unimportant index attributes can be deleted under the condition that the classification ability of information system remains unchanged. Finally, the Adaboost and AODE model are combined to construct the integrated Aode classifier, and the personal credit evaluation model is constructed. The main work of this paper is as follows: (1) an improved discrete particle swarm optimization algorithm is proposed to solve the continuous attribute discretization problem in the sample data set. The breakpoint set of continuous attributes is regarded as discrete particle swarm, and the breakpoint subset is minimized by the interaction between particles. At the same time, simulated annealing algorithm is introduced as a local search strategy, which improves the diversity of particle swarm and the ability of finding global optimal solution. The consistency of decision table is measured by the dependence of decision attributes on conditional attributes in rough set theory, so as to achieve the purpose of discretization of continuous attributes. (2) aiming at the redundancy of most index attributes in the sample data set and the lack of equal importance, most of the index attributes in the sample data set are redundant and do not have the same importance. It is unfavorable to make simple decision in data analysis. Reducing index attribute of sample data set is an important step of credit evaluation. In this paper, an attribute reduction algorithm based on Tabu discrete Particle Swarm Optimization (DPSO) is proposed to select individual credit evaluation indexes. Because Tabu search algorithm has strong dependence on initial solution, and discrete particle swarm optimization algorithm is easy to fall into local optimal solution during iteration, discrete particle swarm optimization algorithm is used to search globally in the process of index selection. Tabu search algorithm searches locally, removes redundant attributes, simplifies knowledge base, and constructs individual credit evaluation index set without affecting classification quality. (3) AODE model is used for classification. Each superparent One-Dependence estimation model (SPODE) has the same contribution to the classification, but each SPODE model has different structure and different influence on the final classification results. In this paper, we propose a corresponding improvement to the structural weakness of the average-dependence estimation algorithm. First of all, a certain number of SPODE models are selected to form an average dependency estimation model from each of the super-parent sole dependence estimation models, and then the integrated AODE classification model is constructed by using Adaboost algorithm. Finally, the integrated AODE classification model is applied to personal credit assessment. The simulation results show that the integrated AODE evaluation model can effectively improve the prediction accuracy of personal credit assessment.
【学位授予单位】：兰州交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP18

【参考文献】