利用证据神经网络的多分类器系统构造  

Design of Multiclassifier Systems Based on an Evidential Neural Network

作　　者：和红顺 韩德强[1,2] 杨艺 HE Hongshun;HAN Deqiang;YANG Yi(School of Electronic and lnformation Engineering,Xi＇an Jiaotong University,Xi＇an 710049,China;CETC Key Laboratory of Aerospace Information Applications,Shijiazhuang 050004,China;State Key Laboratory for Strength and Vibration of Mechanical Structures,Xi＇an Jiaotong University,Xi＇an 710049,China)

机构地区：[1]西安交通大学电子与信息工程学院,西安710049 [2]中国电子科技集团公司航天信息应用技术重点实验室,石家庄050004 [3]西安交通大学机械结构强度与振动国家重点实验室,西安710049

出　　处：《西安交通大学学报》2018年第11期93-99,141,共8页Journal of Xi'an Jiaotong University

基　　金：国家自然科学基金资助项目(61573275;61671370);国家"973计划"资助项目(2013CB329405);中国电子科技集团公司航天信息应用技术重点实验室高校合作课题(KX172600034)

摘　　要：为了充分利用数据信息进而提高分类正确率,提出一种证据神经网络的分类器,并据此构造了多分类器系统。首先将训练数据中的含混数据视为新类别——混合类,将原始的训练数据重组成含有混合类的训练数据,然后使用证据神经网络分类器系统用重组后含混合类的训练数据进行训练,对分类输出进行证据建模,并使用多种不同的证据组合规则实现多分类器融合。采用人工数据集和UCI数据集进行对比实验,结果表明:与其他采用神经网络的多分类器系统相比,采用证据神经网络的多分类器系统能有效提高分类正确率;在数据集Magic 04和Waveform2上,采用提出的多分类器系统比采用投票法的神经网络多分类器系统的分类正确率分别提高了6%和10%左右。A new evidential neural network classifier is proposed to make full use of the data information and to improve the classification performance and an implementation of multiple classifier systems based on the new evidential neural network classifier is presented. Firstly, the ambiguous data contained in the training data are considered as a new category compound class and the original training data are reconstructed into a new data set with compound classes. Then the evidential neural network classifiers are trained using the new data set, and classified outputs are evidentially modeled with the belief function. A variety of rules of evidence combination is used to realize multiclassifiers fusion. Experimental results on the artificial data sets and some UCI data sets and comparisons with some other existing multiclassifier systems based on neural networks show that the proposed multiclassifier systems effectively improve the classification accuracy. Especially, comparisons with the neural network multiclassifier systems based on the voting law on the data sets Magic 04 and Waveform2 show that the proposed multiclassifier systems have about 6% and 10 %increases in classification accuracy, respectively.

关 键 词：神经网络 多分类器系统 证据理论 信度函数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]