Unsupervised feature selection based on matrix factorization with redundancy minimization

Abstract

Unsupervised feature selection (UFS) based on matrix factorization (MF) is an efficient technique for dimensionality reduction in image processing task. Most MF based methods learn cluster indicator matrix and bases matrix, and exploit the bases matrix for feature selection via ranking weights. Since correlated features tend to have similar rankings and these methods select the top ranked features with large correlations, the selected features might contain redundant information. Toward this end, we propose a novel matrix factorization with redundancy minimization method, in which performing MF and removing redundant features are incorporated into a coherent model. To reduce the redundancy, we define a regularization to penalize the high-correlated features. The effective (formula presented) imposed on bases matrix is suitable for feature selection. Experimental results on image datasets validate the effectiveness of the proposed approach.