Manifold learning with structured subspace for multi-label feature selection

•The manifold learning is introduced to avoid too rigid fitting manner between input feature space and corresponding label space.•A latent subspace is constructed to captures the correlations among instances, which learns a more accurate geometry structure of data.•The label correlations are exploited in manifold framework, which ensures the global and local structural consistency of labels.•An efficient algorithm is summarized to solve the optimization problem of the proposed method.•Experiments are conducted on various of datasets, and the results of multiple metrics demonstrate the effectiveness of the proposed algorithm. Nowadays, multi-label learning is ubiquitous in practical applications, in which multi-label data is always confronted with the curse of high-dimensional features. Feature selection has been shown to effectively improve learning performance by selecting discriminative features. Conventional multi-label feature selection only focuses on associating input features with corresponding labels while neglecting the potential structural information, i.e., instance correlations and label correlations. To tackle this problem, we propose manifold learning with structured subspace for multi-label feature selection. Specifically, we first uncover a latent subspace for a more compact and accurate data representation, and take advantage of the subspace to explore the correlations among instances. Then, we explore label correlations in manifold learning to guarantee the global and local structural consistency of labels. Besides, l2,1-norm is introduced into loss function and sparse regularization to facilitate feature selection process. A detail optimization algorithm is presented to solve the objective function of the proposed method. Extensive experiments on real-world data show the superiority of the proposed method under various metrics.