一种全连接粒神经网络分类方法  

A Classification Method of Fully Connected Granular Neural Network

作　　者：傅兴宇 陈颖悦[2] 陈玉明 江海亮 黄涛 FU Xingyu;CHEN Yingyue;CHEN Yuming;JIANG Hailiang;HUANG Tao(College of Computer and Information Engineering,Xiamen University of Technology,Xiamen 361024,China;Center of Economics and Management,Xiamen University of Technology,Xiamen 361024,China)

机构地区：[1]厦门理工学院计算机与信息工程学院,福建厦门361024 [2]厦门理工学院经济与管理实验中心,福建厦门361024

出　　处：《山西大学学报（自然科学版）》2023年第1期91-100,共10页Journal of Shanxi University(Natural Science Edition)

基　　金：国家自然科学基金(61976183);福建省自然科学基金(2019J01850);厦门市科技计划项目(3502Z20193069);厦门市产学研项目(2022CXY04028)。

摘　　要：全连接神经网络需要大量的数据支持,才能训练好一个分类网络,往往现实中没有提供大量的数据供给网络训练。针对全连接神经网络缺少数据训练会使网络分类效果不佳这个问题,研究粒计算理论,从不同角度增广数据并进行粒化,提出一种全连接粒神经网络的分类方法。首先,该网络对所有样本进行单特征参照样本相似度粒化,形成参照样本粒子。同时引入邻域判别函数进行邻域粒化,形成邻域粒子。一个样本上的多个特征粒子构成一个粒向量,将构造的粒向量输入到该网络进行分类,进而提出了全连接粒神经网络。在多个UCI数据集上实验,用全连接粒神经网络和不同的分类算法进行比较,其结果表明了所提出的全连接粒神经网络分类方法的正确性与有效性。Fully connected neural networks need a lot of data to train a classification network. In reality, normally there isn’t large amount of data for network training. Aiming at the problem that the lack of data training for fully connected neural network will make the network classification effect poor, this paper studies the granular computing theory, expands and granulates the data from different angles, and puts forward a classification method of fully connected granular neural network. Firstly, the network granulates the similarity of single feature reference samples for all samples to form reference sample particles. At the same time, the neighborhood discriminant function is introduced to granulate the neighborhood to form neighborhood particles. Multiple characteristic particles on a sample form a particle vector. The constructed particle vector is input into the network for classification, and then a fully connected particle neural network is proposed. Experiments on several UCI data sets are carried out to compare different classification algorithms with fully connected granular neural network. The results show that the proposed classification method of fully connected granular neural network is correct and effective.

关 键 词：全连接 粒神经网络 邻域粒化 相似度粒化 分类网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]