Deep Multi-View Subspace Clustering Based on Intact Space Learning

Abstract

Since single-view data is not enough to describe objects comprehensively, how to make full use of multi-view data has become a popular research direction in the field of machine learning and pattern recognition. In this paper, we mainly focus on multi-view subspace clustering. We propose a deep multi-view subspace clustering based on intact space learning model (DMVSC-ISL). Our model contains an intact space learning module and a subspace clustering module. The intact space learning module utilizes an auto-encoder networks to reduce the redundant and noise information to obtain the latent representation of each view and utilizes a degradation networks to obtain the intact space representation based on these different latent representations. Comparing single-view, the obtained intact space representation contains richer multi-view information. The subspace clustering module utilizes a self-expression layer to calculate the reconstruction coefficient matrix of the intact space representation. Through calculating intact space representation, we integrate the intact space learning and subspace clustering into one train network. Through considering their interactions, our model has good clustering property. By comparing the existing clustering methods on four multi-view data sets, experimental results show the superiority of DMVSC-ISL clustering performance.