Latent shared representation for multi-view subspace clustering

Abstract

Currently, multi-view subspace clustering has obtained wide attention in the field of machine learning and pattern recognition. Since different views contain different view-specific information of the same object, how to utilize these views to recover a latent shared representation is important for subsequent clustering. In this paper, we propose a latent shared representation model for multi-view subspace clustering. Our model uses the view-specific generated matrices to recover a latent shared representation from the multi-view data. To consider the correlation information of multi-view data, we adopt an empirical Hilbert-Schmidt independence criterion to constrain these view-specific generated matrices. Based on the inexact augmented Lagrangian method, we also develop an alternating optimization algorithm to solve our model. Experimental results on the real-world multi-view data sets have validated the effectiveness of our model.