带质心的K最近邻增强模糊最小最大神经网络的集成方法  

Ensemble Method of K-Nearest Neighbor Enhancement Fuzzy Minimax Neural Networks with Centroid

作　　者：陈鹏[1] 赵建成 余肖生[1] CHEN Peng;ZHAO Jiancheng;YU Xiaosheng(College of Computer and Information, Three Gorges University, Yichang 443002, China)

机构地区：[1]三峡大学计算机与信息学院,湖北宜昌443002

出　　处：《重庆理工大学学报（自然科学）》2021年第9期116-129,共14页Journal of Chongqing University of Technology：Natural Science

基　　金：国家重点研究发展计划项目(2016YFC0802500)。

摘　　要：在分类任务中,传统的模糊最小最大神经网络及其变体在训练网络的时候没有考虑超盒内部训练数据的分布情况,并且考虑扩展系数的问题也不是很充分,导致每次训练新数据集都要重新选择最优的扩展系数。因此,提出了一种带质心的K最近邻增强模糊最小最大神经网络的集成方法。在该方法中,一方面,每个超盒都带有质心,用来描述之前训练的样本在该超盒的大体分布情况,并且在扩展规则及收缩方面考虑了样本距离超盒质心的因素;另一方面,使用5个带质心的K最近邻增强模糊最小最大神经网络作为弱分类器,每个分类器设置不同的扩展系数,当该方法训练完后,将得出的离散属性值作为随机森林的训练集,最后,使用测试样本验证网络的分类性能。实验结果表明:提出的方法在准确率、精准率、召回率以及F-score等方面大部分的结果高于传统的FMMN及其变体的结果,该方法有效地克服了FMMN的准确性过于依赖扩展系数的问题。In the classification,the traditional Fuzzy Minimum and Maximum Neural Network(FMMN)as well as its variants did not consider the distribution of the training data in the hyperbox,and the expansion coefficient was not fully considered,resulting in that each training of new data sets need re-selecting the optimal expansion coefficient.Therefore,the authors propose an integrated method of K-nearest neighbor enhancement fuzzy min-max neural network with centroid.In the method,on the one hand,each hyperbox contains a centroid,which is used to describe the distribution of the previously trained sample,and the problem of distance from sample to the centroid of the hyperbox is considered in terms of expansion rules and contraction;On the other hand,five K-nearest neighbor enhanced fuzzy min-max neural networks are used as weak classifiers,and each classifier is set with different expansion coefficients.After the training,the discrete attribute values are used as the training set of random forest.Finally,the test samples are utilized to verify the classification performance of the network.The experimental results show that the method proposed in the paper outperform the traditional FMMN and its variants in accuracy,precision,recall,and F-score,and this method effectively overcomes the problem that the accuracy of FMMN depends too much on the expansion coefficient.

关 键 词：E-CFMM 集成方法 超盒收缩 弱分类器 质心 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]