摘要
针对目前深度多视角子空间聚类算法因为缺少对自表达矩阵的低秩表示约束而导致的模型缺乏鲁棒性的问题,提出了深度低秩多视角子空间聚类算法。在深度多视角子空间聚类算法的基础上,通过矩阵分解将自表达层分解为多视角一致性自表达层和单视角特异性自表达层,得到具有低秩线型约束的双层自表达模块;强制所有视角的一致性自表达层的参数相同、特异性自表达层的参数各不相同,充分利用多视角数据的互补性;将自表达模块嵌入到每个视角的深度自编码器中,得到可以通过反向传播算法求解的深度低秩多视角子空间聚类模型;在深度模型训练中,一致性自表达层学习多视角数据的一致性信息,特异性自表达层学习单个视角的独特信息,双层自表达模块隐性地添加了低秩表示约束。6个公开数据集上的实验结果表明:与深度多视角子空间聚类算法相比,所提算法的聚类正确率平均提升了0.064,标准化互信息提升了0.064;所提算法的正确率和标准化互信息优于其他11种先进聚类算法的,聚类正确率最大提升了0.097,标准化互信息最大提升了0.103。
To deal with the problem that the current deep multi-view subspace clustering methods lack robustness due to the lack of low-rank representation constraints on the self-representation matrix,we propose a novel deep low-rank multi-view subspace clustering(DLRMSC)method.Based on the deep multi-view subspace clustering algorithm,the self-expression layer is split into two low-rank linear self-expression layers:the commonly shared self-expression layer and the view-specific self-expression layer,which becomes a double-layer self-expression module with low-rank constraint.To make full use of the complementarity of multi-view data,the parameters of the commonly shared self-expression layer are forced to be the same,but the parameters of specific self-expression layer are different.We embed the self-expression module in the middle of the deep auto-encoder of each view,so as to obtain a deep low-rank multi-view subspace clustering model which can be solved by backpropagation.In the training stage,the commonly shared self-expression layer extracts the common information of multi-view data,the view-specific self-expression layer extracts the specific information of each view,and the double-layer self-expression module guarantees low-rank representation constraint.The experimental results on six public datasets show that compared with the deep multi-view subspace clustering algorithm,the clustering accuracy and normalized mutual information(NMI)of this proposed method on six datasets are significantly improved,where the average accuracy is increased by 0.064 and the average NMI is increased by 0.064.The DLRMSC method substantially outperforms the eleven comparative state-of-the-art clustering methods,and its accuracy is increased by 0.097 and NMI is increased by 0.103 at most.
作者
闫金涛
李钟毓
唐启凡
周志豪
YAN Jintao;LI Zhongyu;TANG Qifan;ZHOU Zhihao(School of Software Engineering,Xi’an Jiaotong University,Xi’an 710049,China)
出处
《西安交通大学学报》
EI
CAS
CSCD
北大核心
2021年第11期125-135,共11页
Journal of Xi'an Jiaotong University
基金
国家自然科学基金资助项目(61902310)。
关键词
多视角聚类
子空间聚类
自编码器
矩阵分解
低秩表示
multi-view clustering
subspace clustering
autoencoder
matrix factorization
low-rank representation
