Robust dimensionality reduction method based on relaxed energy and structure preserving embedding for multiview clustering

Existing multiview clustering methods often learn inferior quality affinity graphs from original data, which suffer from excessive dimensionality and contain noise and outliers. Therefore, dimensionality reduction must be performed on the original data, which leads to the following problems: 1) the learned projection matrix does not preserve the main dis-criminative information of the original data; 2) the local and global structures are not pre-served simultaneously; and 3) the negative effects of noise, outliers, and corruption cannot be overcome. To alleviate the aforementioned problems, a robust dimensionality reduction method based on a relaxed energy-preserving and structure-preserving embedding for multiview clustering is proposed in this study. In particular, the proposed architecture is a double-relaxed energy-preserving embedding model, which uses two projection matrices to extract the main discriminant features from the original data using two energy -preserving embedding routes, yielding clean data and a representation matrix in low -dimensional space. Subsequently, an affinity matrix that captures both local and global data structures is learned by performing locality-preserving projection and global low -rank-preserving embedding on the clean data and the representation matrix, respectively. Experimental results on benchmark datasets validate the effectiveness of the proposed algorithm.(c) 2022 Elsevier Inc. All rights reserved.

Automated summary

Existing multiview clustering methods often learn inferior quality affinity graphs from original data, which suffer from excessive dimensionality and contain noise and outliers. To address these problems, a robust dimensionality reduction method based on a relaxed energy-preserving and structure-preserving embedding for multiview clustering is proposed. The proposed architecture is a double-relaxed energy-preserving embedding model that extracts main discriminant features from original data using two energy-preserving embedding routes, resulting in clean data and a representation matrix in low-dimensional space. The learned affinity matrix captures both local and global data structures through locality-preserving projection and global low-rank-preserving embedding on the clean data and representation matrix, respectively. Experimental results on benchmark datasets validate the effectiveness of the proposed algorithm.