互补与一致的多视角子空间聚类网络

Complementary and Consistent Multi-View Subspace Clustering Network

针对在非线性子空间中同时探索多视角数据的互补性与一致性的问题,提出了一种互补与一致的多视角子空间聚类网络C2MSCN。该网络由编码器、自表达层、谱聚类模块和解码器4部分组成。通过编码器将数据映射到潜在的非线性空间,探索多视角数据间复杂的非线性关系;利用自表达层学习多视角共享的自表达系数矩阵和单视角私有的自表达系数矩阵,同时挖掘多视角数据间的互补与一致信息;将系数矩阵供给谱聚类模块以获得聚类标签;通过解码器将自表达数据还原至原始空间。为充分挖掘多视角数据间的互补信息,针对私有自表达系数矩阵加入多样性规范化项;为利用聚类标签监督自表达系数的学习,在目标函数中加入自监督规范化项。实验结果表明:所提算法在4个评价指标下6个数据集的对比实验中取得了1.25的平均排名;在Yale数据集上,用准确率评价的聚类准确度较线性空间多视角聚类算法CSMSC、深度多模态子空间聚类算法DMSCN和自编码器算法Ae2-nets分别提高了8.3%、6.3%和11.7%;参数敏感度实验和消融性实验表明,所提算法能有效在非线性子空间中探索数据互补性与一致性,且在不同数据集上表现稳定。

A complementary and consistent multi-view subspace clustering network(C2MSCN)is proposed to better explore the complementarity and consistency of multi-view data in nonlinear subspace.The network consists of four parts:encoder,self-expression layer,spectral clustering module,and decoder.This network firstly maps the data to the potential nonlinear space through the encoder to explore the complex nonlinear relationship among multi-view data.Then the self-expression layer is used to simultaneously learn the self-expression coefficient matrix shared by multiple views and private self-expression coefficient matrixes of multiple views to simultaneously mine the complementary and consistent information in multi-view data.Then the coefficient matrixes are fed to the spectral clustering module to obtain the clustering label.Finally,the self-expressed data is restored to the original space through the decoder.A multiplicity normalized item for private self-expression coefficient matrixes is added to fully mine the complementary information among multi-view data.Meanwhile,a self-supervised normalized term is added to supervise the learning process of self-expression coefficient matrixes.Experimental results show that the proposed algorithm achieves an average ranking of 1.25 in six data sets under four evaluation indexes.On the Yale dataset,the clustering accuracy evaluated by ACC is 8.3%,6.3%,and 11.7%higher than that of the linear space multi-view clustering algorithm CSMSC,the deep multi-view subspace clustering algorithm DMSCN and the autoencoder algorithm Ae2-nets,respectively.Parameter sensitivity experiments and ablation experiments indicate that the proposed algorithm can effectively explore data complementarity and consistency in nonlinear subspace,and performs stably in different data sets.