Multiset Canonical Correlations Analysis With Global Structure Preservation

Abstract

This paper considers unsupervised dimensionality reduction of multi-view data, where locality preserving canonical correlation analysis and a new locality-preserving canonical correlation analysis are two typical effective methods. However, they ignore the global structure while considering the local structure of data, and are sensitive to noises because of the relationship of neighbors based on the Euclidean distance. In this paper, we propose a novel multi-view dimensionality reduction method: multiset canonical correlations analysis based on low-rank representation. Our model introduces the cross-view similarity matrix to consider the correlation of all different points in cross views, which makes it not only preserve the local structure but also the global structure of data. And the cross-view similarity matrix is constructed by using low-rank representation, which can make the model more robust. In addition, a parameter $\beta $ is introduced to adjust the importance of the correlation of different sample points, enhancing the generalization ability for different datasets. Experiments on four multi-view datasets show our proposed method has better performance than the related methods.