Multi-label Feature Selection via Dual Graph and Non-convex Sparse Regression

Abstract

Feature selection has been an important data preprocessing in multi-label learning and thus has been developed rapidly in recent years. Among these, the methods based on manifold regularization have received more attention. However, most existing methods only consider the geometric structure of the data manifold and ignore the local information in the feature space, which leads to the learned manifold information being incomplete. To tackle this problem, we propose a dual graph regularization for multi-label feature selection, which considers the geometric structure of the data manifold and feature manifold simultaneously. What is more, a new non-convex constraint consisting of the difference between $l_{2,1}$ -norm and Frobenius norm, denoted as $l_{2,1-2}$ -norm, is introduced to obtain more row-sparse solution in order to delete a more redundant features efficiently. An alternating minimization strategy is designed to optimize the proposed objective function. Comprehensive experiments are conducted on eight multi-label data sets, and the results demonstrate that our method is superior to the compared methods.