Sparse multi-label feature selection via dynamic graph manifold regularization

Yao Zhang, Yingcang Ma

Abstract: Multi-label feature selection is a hot topic in multi-label high-dimensional data processing. However, some multi-label feature selection models use manifold graphs. Due to its fixed graph matrix, the model performance is poor, and learning a better fundamental graph matrix is also an urgent problem. Therefore, a sparse multi-label feature selection method is proposed via dynamic graph manifold learning (DMMFS). In this method, the sample space is mapped to the pseudo-label space with a real-label base manifold structure through linear mapping. Then, the Frobenius norm constructed the dynamic graph matrix, and the mutual constraint between the weight matrix and dynamic graph matrix is realized by feature manifold. Finally, experimental comparisons are made on eight multi-label reference data sets with seven of the latest methods. The experimental results prove the superiority of DMMFS.