基于混淆矩阵的分类器选择集成方法研究

发布时间：2018-04-29 23:12

本文选题：多分类器系统 + 选择性集成　；参考：《河南理工大学》2016年硕士论文

【摘要】：集成学习是机器学习领域的重要研究方向,它通过训练多个个体分类器并把它们组合起来形成多分类器系统,以此来提高分类性能。但是随着计算机技术的发展,数据量的增大,参与集成的分类器也就越来越多,这样一方面计算量迅速增长,另一方面,分类器之间的差异度也变小,影响了集成的准确性,而一个有效的集成系统需要参与集成的分类器具有比较高的准确性和差异性。研究表明,从训练产生的基分类器中选择一部分来集成,这种方法可能比使用全部的基分类器来进行集成效果更好。因此,从大量的基分类器中选择出具有高差异性的分类器作为代表参与集成,已成为集成学习的一个研究趋势,需要进行更加深入的研究。本文在集成学习的基础上,首先介绍了多分类器集成的国内外研究背景和意义,总结了集成学习的国内外研究现状。其次介绍了集成学习的概念和两种经典的集成算法Bagging和Boosting算法,接着,分别列举了乘积规则、求和规则等6种集成规则。然后从差异性度量公式的角度出发,介绍了差异性度量的概念,以及常用的度量公式。最后提出了一种新的多分类器选择性方法,具体方法是构造所有基分类器的混淆矩阵作为聚类算法的数据对象,依据各聚类中样本的分布情况,选择出一定数量的分类器作为代表,构成新的待集成分类器集合,然后把这个方法应用于Bagging的训练过程中。为了验证本文方法的可行性,在UCI数据集上进行实验,将本文的方法应用于Bagging算法的训练过程得到的实验结果与使用原始的Bagging算法得到的结果进行比较,表明,该方法可以有效提高集成系统的准确率。并选择不同的集成规则进行集成,对结果进行分析。
[Abstract]:Ensemble learning is an important research direction in the field of machine learning. It improves classification performance by training multiple individual classifiers and combining them to form multiple classifiers. However, with the development of computer technology and the increase of data volume, more and more classifiers are involved in the integration. On the one hand, the amount of computation increases rapidly, on the other hand, the difference between the classifiers becomes smaller, which affects the accuracy of integration. An effective integration system needs to participate in the integration of classifiers with high accuracy and difference. It is shown that this method is more effective than using all the base classifiers to integrate a part of the basic classifiers generated by the training. Therefore, it has become a research trend of ensemble learning to select classifiers with high diversity from a large number of base classifiers as representatives to participate in integration, which need to be further studied. On the basis of integrated learning, this paper first introduces the research background and significance of multi-classifier integration at home and abroad, and summarizes the current research situation of integrated learning at home and abroad. Secondly, the concept of integration learning and two classical integration algorithms, Bagging and Boosting, are introduced. Then, six kinds of integration rules, including product rule and summation rule, are listed respectively. Then, from the point of view of the difference measurement formula, the concept of the difference measure and the commonly used measurement formula are introduced. Finally, a new multi-classifier selective method is proposed. The method is to construct the confusion matrix of all base classifiers as the data object of the clustering algorithm, according to the distribution of samples in each cluster. A certain number of classifiers are selected as representatives to form a new ensemble of classifiers, and then this method is applied to the training process of Bagging. In order to verify the feasibility of this method, the experiment is carried out on the UCI dataset. The experimental results obtained from the training process of the Bagging algorithm are compared with the results obtained by using the original Bagging algorithm. This method can effectively improve the accuracy of the integrated system. Different integration rules are selected and the results are analyzed.
【学位授予单位】：河南理工大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP181;TP311.13

【相似文献】