Graph-based multi-label feature selection with dynamic graph constraints and latent representation learning

Currently, multi-label feature selection with joint manifold learning and linear mapping has received much attention. However, the low-quality graph matrix used by existing methods leads to model limitations. Traditional linear mapping cannot learn the coupling relationship between different outputs. In addition, existing approaches ignore latent supervisory information in label correlation. To this end, we obtain a dynamic graph matrix with Laplace rank constraints by the \(L_{1}\) $L_1$ norm with a conventional graph matrix. We also mine more reliable supervised information from label correlations by introducing latent representation learning. Moreover, we integrate all the above terms into a linear mapping learning framework based on improved matrix decomposition, and design a simple and effective scheme based on alternating iterations to optimize this framework. Numerous experimental results validate the competitive advantage of the proposed method over existing state-of-the-art methods.