深度多网络嵌入聚类  

Deep Multi-network Embedded Clustering

作　　者：陈锐 唐永强 张彩霞[1,3] 张文生 郝志峰[1] CHEN Rui;TANG Yongqiang;ZHANG Caixia;ZHANG Wensheng;HAO Zhifeng(School of Mechatronic Engineering and Automation,Foshan University,Foshan 528225;Research Center of Precision Sensing and Control,Institute of Automation,Chinese Academy of Sciences,Beijing 100190;Guangdong Province Smart City Infrastructure Health Monitoring and Evaluation Engineering Technology Research Center,Guangdong Engineering Technology Research Center,Foshan 528000)

机构地区：[1]佛山科学技术学院机电工程与自动化学院,佛山528225 [2]中国科学院自动化研究所精密感知与控制研究中心,北京100190 [3]广东省工程技术研究中心广东省智慧城市基础设施健康监测与评估工程技术研究中心,佛山528000

出　　处：《模式识别与人工智能》2021年第1期14-24,共11页Pattern Recognition and Artificial Intelligence

基　　金：国家自然科学基金青年基金项目(No.61803087);广东省教育厅特色创新项目(No.2019KTSCX192);广东省基础与应用基础研究基金粤港澳应用数学中心项目(No.2020B1515310003);佛山核心技术攻关项目(No.1920001001367)资助。

摘　　要：现有的深度无监督聚类方法通常局限于单网络结构设计,无法充分利用多种异构网络提取特征中蕴含的互补信息,制约深度聚类方法性能的进一步提升.为此,文中提出深度多网络嵌入聚类算法(DMNEC).首先,以端到端的方式预训练多个异构网络分支,获取各网络的初始化参数.在此基础上,定义多网络软分配,借助多网络辅助目标分布建立面向聚类的KL散度损失.与此同时,利用样本重建损失对预训练阶段的解码网络进行微调,保留数据的局部结构性质,避免特征空间发生扭曲.通过随机梯度下降与反向传播优化重建损失与聚类损失的加权和,联合学习多网络表征及其簇分配.在4个公开图像数据集上的实验验证文中算法的优越性.Existing deep unsupervised clustering methods cannot make full use of the complementary information between the extracted features of different network structures due to the single network structure in them,and thus the clustering performance is restricted.A deep multi-network embedded clustering(DMNEC)algorithm is proposed to solve this problem.Firstly,the initialization parameters of each network are obtained by pretraining multi-network branches in an end-to-end manner.On this basis,the multi-network soft assignment is defined,then the clustering-oriented Kullback-Leibler divergence loss is established with the help of the multi-network auxiliary target distribution.The decoding network in the pretraining stage is finetuned via reconstruction loss to preserve the local structure and avoid the distortion of feature space.The weighted sum of reconstruction loss and clustering loss is optimized by stochastic gradient descent(SGD)and back propagation(BP)to jointly learn multi-network representation and cluster assignment.Experiments on four public image datasets demonstrate the superiority of the proposed algorithm.

关 键 词：深度无监督聚类 数据表征 多网络分支 互补信息 局部结构保留 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]