Tongxin xuebao (Dec 2020)
Multi-label feature selection based on dynamic graph Laplacian
Abstract
In view of the problems that graph-based multi-label feature selection methods ignore the dynamic change of graph Laplacian matrix, as well as such methods employ logical-value labels to guide feature selection process and loses label information, a multi-label feature selection method based on both dynamic graph Laplacian matrix and real-value labels was proposed.The robust low-dimensional space of feature matrix was used to construct a dynamic graph Laplacian matrix, and the robust low-dimensional space was used as the real-value label space.Furthermore, manifold and non-negative constraints were adopted to transform logical labels into real-valued labels to address the issues mentioned above.The proposed method was compared to three multi-label feature selection methods on nine multi-label benchmark data sets in experiments.The experimental results demonstrate that the proposed multi-label feature selection method can obtain the higher quality feature subset and achieve good classification performance.