Multi-label learning of label-specific features via landmark selection and correlation Information

The existing algorithms in multi-label learning typically rely solely on the feature space to directly predict all labels. However, this approach may not be the most optimal strategy in practical applications. In this paper, we propose a label-specific multi-label learning algorithm that incorporates landmark selection and correlation information (LSCI). The representative labels are selected by a landmark selection matrix, while the label correlation is achieved by exploring the correlation between landmarks and other labels. The instance correlation constraints are designed to preserve local structural attributes at the same time. By acquiring the correlation information, we accomplish proficient extraction of features specific to the label. Experiments demonstrating that our proposed algorithm based on landmark selection and correlation information competes favorably with state-of-the-art algorithms.