Unsupervised feature selection for visual classification via feature-representation property

Feature selection is designed to select a subset of features for avoiding the issue of 'curse of dimensionality'. In this paper, we propose a new feature-level self-representation framework for unsupervised feature selection. Specifically, the proposed method first uses a feature-level self-representation loss function to sparsely represent each feature by other features, and then employs an ℓ2,p-norm regularization term to yield row-sparsity on the coefficient matrix for conducting feature selection. Experimental results on benchmark databases showed that the proposed method effectively selected the most relevant features than the state-of-the-art methods.