Comprehensive Multi-view Representation Learning

Abstract

Recently, Multi-view Representation Learning (MRL) has drawn immense attentions in the analysis of multi-source data and ubiquitously employed across different research fields. This important issue is designed to learn a feature representation with sufficient information from multiple views. In this paper, we propose a novel Comprehensive Multi-view Representation Learning (CMRL), which can fully explore available information contained in both the feature representations and subspace representations of multiple views. The desired feature representation learned in CMRL profits from the consistency and complementarity of multi-view data. Specifically, the complementary information is mined by applying the degeneration mapping model on multiple feature representations, the consensus information is explored by imposing a low-rank tensor constraint on multiple subspace representations. Further, the objective function of CMRL is optimized by an Augmented Lagrangian Multiplier (ALM) based algorithm. Finally, our CMRL is evaluated on seven benchmark multi-view datasets and compared with several state-of-the-art methods, experimental results illustrate the superiority and effectiveness of the proposed method. What is more, we find that the proposed method can also be successfully applied to multi-view subspace clustering and achieves promising clustering results.