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

Multi-view clustering has received much attention by exploring the semantic information of multiple views in an unsupervised manner. To address the shortcomings of existing methods that overlook the learning of representation diversity and redundant information between views, a novel enhanced fusion network through redundancy elimination for deep multi view clustering (REMVC) is proposed. Firstly, a feature aggregation module is introduced to learn the global structural relationships between samples for generating consistent representations. Secondly, a redundancy elimination module is introduced to reduce the information correlation between potential spatial features and improve the discriminative ability of embedded representations. Finally, a regularization constraint module for common representation is introduced to explore the complementary information of multi-view instance features by maximizing the variance on each dimension of the common representation and minimizing the covariance between different dimensions. The proposed network jointly optimizes the representation learning, redundancy elimination, and regularization constraint modules to improve clustering performance. The results of experiments on 7 widely used datasets show that REMVC outperforms the 7 multi-view clustering methods involved in comparison in terms of accuracy, normalized mutual information, and purity.