高级检索
许俊鹏, 孟敏, 武继刚. 基于冗余消除的深度多视图聚类增强融合网络[J]. 计算机辅助设计与图形学学报. DOI: 10.3724/SP.J.1089.2023-00468
引用本文: 许俊鹏, 孟敏, 武继刚. 基于冗余消除的深度多视图聚类增强融合网络[J]. 计算机辅助设计与图形学学报. DOI: 10.3724/SP.J.1089.2023-00468
Junpeng Xu, Min Meng, Jigang Wu. Enhanced Fusion Network through Redundancy Elimination for Deep Multi-view Clustering[J]. Journal of Computer-Aided Design & Computer Graphics. DOI: 10.3724/SP.J.1089.2023-00468
Citation: Junpeng Xu, Min Meng, Jigang Wu. Enhanced Fusion Network through Redundancy Elimination for Deep Multi-view Clustering[J]. Journal of Computer-Aided Design & Computer Graphics. DOI: 10.3724/SP.J.1089.2023-00468

基于冗余消除的深度多视图聚类增强融合网络

Enhanced Fusion Network through Redundancy Elimination for Deep Multi-view Clustering

  • 摘要:  多视图聚类通过以无监督的方式探索多视图的语义信息而得到了广泛的关注. 然而, 现有的方法大多只考虑在多个视图中探索公共语义, 而忽略了表示的多样性学习. 因此, 它们不能利用跨视图的互补信息, 这可能会抑制多视图表示学习的能力. 同时, 他们忽略了视图之间冗余信息的影响. 为了解决上述问题, 提出了一种基于冗余消除的深度多视图聚类增强融合网络REMVC. 具体来说, 本文引入了特征聚合模块, 旨在学习样本之间的全局结构关系, 并利用全局特征生成一致的表示. 考虑到多视图冗余问题, 引入冗余消除策略减少潜在空间特征之间的信息相关性, 提高了嵌入表示的辨别能力. 然后, 为了探索多视图实例特征的互补信息, 引入公共表示的正则化约束模块, 它由维度上的方差和协方差两部分组成, 该模块最大化公共表示每个维度上的方差, 并最小化不同维度之间的协方差. 最后, 所提出的网络联合优化了表示学习、冗余消除和正则化约束模块, 以提高聚类性能. 在四个广泛使用的数据集上进行的实验表明, REMVC优于几种最先进的多视图聚类方法.

     

    Abstract: Multi-view clustering has received much attention by exploring the semantic information of multiple views in an unsupervised manner. However, most existing methods only consider exploring the common semantics across multiple views while neglecting to promote the diversity of representations. Thus, they cannot take advantage of complementary information across views, which may inhibit the capability of multi-view representation learning. At the same time, they ignore the impact of redundant information among views. To address these issues, a novel enhanced fusion network through redundancy elimination for deep multi-view clustering is proposed, called REMVC. Specifically, a feature aggregation module is introduced to learn the global structural relationships between samples and generate consistent representations using global features. To address the issue of redundancy in multiple views, a redundancy elimination strategy is employed to reduce the information correlation between latent space features, enhancing the discriminative capability of the embedding representation. Furthermore, to explore the complementary information of instance features across multiple views, a regularization constraint module for the common representation is proposed, which consists of variance and covariance on dimensions. This module maximized the variance of each dimension in the common representation while minimizing the covariance between different dimensions. Finally, the proposed network jointly optimizes the representation learning, redundancy elimination, and regularization constraint modules to enhance clustering performance. Experimental results on four widely used datasets demonstrate that REMVC outperforms several state-of-the-art multi-view clustering methods.

     

/

返回文章
返回