A framework for multi-view feature selection via embedding space

Abstract

Multi-view learning has drawn much attention in the past years to reveal the correlated and complemental information between different views. Feature selection for multi-view data is still a challenge in dimension reduction. Most of the multi-view feature selection methods simply concatenate all views together without capturing the information between different views. In this paper, we propose an embedding framework for multi-view feature selection, Embedding Space based Multi-view Feature Selection (ESMFS). ESMFS comes up with a new concept called mapping consensus to embed all views of data to a unified space. By preserving the manifold information, ESMFS captures the fusing views’ information. ESMFS is suitable for both supervised and unsupervised feature selection. For practical purpose, we propose two methods ES-LRFS and ES-MAFS to illustrate ESMFS framework. Experiments show that ES-LRFS and ES-MAFS are of inclusiveness and efficiency for multi-view feature selection, thus proving the feasibility of ESMFS.