关于融合GRASP算法的选择性集成学习方法研究

发布时间：2018-07-05 17:37

  本文选题：信用评估 + 集成剪枝　； 参考：《南京航空航天大学》2016年硕士论文

【摘要】：近年来,由于评估的准确性严重影响到金融机构的损失大小,信用评估问题得到了金融机构越来越多的关注。目前学者们已经提出很多方法用于解决信用评估问题。这些方法概括起来主要分为两大类:基于统计学的方法和基于非统计学的方法,前者主要包括线性判别分析、logit分析和probit分析等,后者主要包括支持向量机(support vector machine,SVM)、人工神经网络(artificial neural network,ANN)和决策树(decision tree,DT)等。尽管研究表明基于非统计学的方法取得了比较好的分类性能,但是单一模型用于解决信用评估问题存在着一定的局限性,后来学者们又提出了用集成学习(ensemble learning)解决该问题。然而,集成学习需要多个基分类器,这样增加了时间和空间复杂性,同时泛化能力差的基分类器也影响着集成系统最终的分类效果。然而在原始的集成系统中选择一个子集用于构建集成系统可以很好的解决这个问题,这种方法被命名为选择性集成(ensemble selection),或者也可以称之为集成剪枝(ensemble pruning)。本文提出了ELMsGraspEnS算法用于解决信用评估问题。该算法用ELM作为生成集成系统的基学习机,GraspEnS作为集成剪枝算法在原始的集成系统中选择一部分子集,因此ELMsGraspEnS继承了ELM和GraspEnS的优点。ELM算法的学习速度非常快,具有优越的泛化性能,并且可以有效的解决局部最优和过拟合问题。GraspEn S算法是GRASP算法在集成剪枝方面的应用,是一种组合优化的启发式算法,不仅具有了贪婪集成剪枝算法的优点,并且可以避免贪婪集成剪枝算法所具有的局部最优问题,另外,该算法还可以实现多点开始搜索。实验部分也表明了新提出的ELMsGraspEn S算法具有很好的分类效果。然而GRASP算法是一个无记忆算法,即在GRASP算法的迭代过程中无法利用前面迭代的信息,Path-Relinking算法是一个加强算法,融合GRASP和Path-Relinking可以避免GRASP算法中所存在的问题。鉴于此,本文提出了另一种PRelinkGrasp EnS算法用于解决信用风险评估问题,该算法也是用ELM算法作为基学习机,所不同的是在生成原始的集成系统时,用了Bagging技术,这样增加了基分类器的多样性,该算法用融合了GRASP和Path-Relinking用于选择性集成,这样不仅具有GRASP算法的优点,也结合了Path-Relinking的优势,使得PRelinkGraspEnS算法是一个有记忆的算法,实验结果表明新提出的PRelinkGraspEn S算法不仅具有优越的泛化性能,还可以加快收敛速度。
[Abstract]:In recent years, due to the accuracy of the evaluation of financial institutions seriously affected the size of losses, credit evaluation has been more and more concerned by financial institutions. At present, scholars have put forward many methods to solve the credit evaluation problem. These methods are divided into two main categories: statistical based method and non-statistical method. The former includes linear discriminant analysis (LDA) logit analysis and probit analysis. The latter includes support vector machine (support vector machine), artificial neural network (artificial neural network Ann) and decision tree (decision tree). Although the research shows that the non-statistical method achieves good classification performance, the single model has some limitations in solving the credit evaluation problem. Later, scholars proposed to use integrated learning (ensemble learning) to solve the problem. However, ensemble learning requires multiple base classifiers, which increase the complexity of time and space, and the poor generalization of base classifiers also affects the final classification effect of the ensemble system. However, choosing a subset from the original integration system to build the integration system is a good solution to this problem, which is called selective integration (ensemble selection), or integration pruning (ensemble pruning). In this paper, ELMsGraspEns algorithm is proposed to solve the credit evaluation problem. The algorithm uses ELM as the basic learning machine for generating integration system GraspEnS as the integrated pruning algorithm to select a subset in the original integration system, so ELMsGraspEnS inherits the advantages of ELM and GraspEnS. The learning speed of ELM algorithm is very fast. GraspEn S algorithm is an application of grasp algorithm in integrated pruning, and it is a heuristic algorithm of combinatorial optimization, which has excellent generalization performance and can effectively solve the problem of local optimality and over-fitting .GraspEn S algorithm is the application of grasp algorithm in integrated pruning, and it is a kind of heuristic algorithm of combinatorial optimization. It not only has the advantages of greedy integration pruning algorithm, but also avoids the local optimal problem of greedy integration pruning algorithm. In addition, the algorithm can also realize multi-point search. The experimental results also show that the new ELMsGraspEn S algorithm has a good classification effect. However, grasp algorithm is a memoryless algorithm. In the iterative process of grasp algorithm, the information of the previous iteration can not be used in the process of the Path-Relation algorithm is an enhanced algorithm, the fusion of grip and Path-Relationing algorithm can avoid the problems in grip algorithm. In view of this, this paper proposes another PRelinkGrasp Ens algorithm to solve the credit risk assessment problem. The algorithm also uses ELM algorithm as the basic learning machine. The difference is that bagging technique is used to generate the original integrated system. In this way, the diversity of base classifiers is increased. The algorithm combines grasp and Path-Relationing for selective integration, which not only has the advantages of grasp algorithm, but also combines the advantages of Path-Relationing, which makes PRelinkGraspEns algorithm a memorized algorithm. The experimental results show that the proposed PRelinkGraspEn S algorithm not only has excellent generalization performance, but also can accelerate the convergence speed.
【学位授予单位】：南京航空航天大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP181
，

本文编号：2101105