Robust multi-label feature selection with shared coupled and dynamic graph regularization

Lingzhi Wang, Hongmei Chen, Bo Peng, Tianrui Li, Tengyu Yin

Abstract: The graph-based multi-label feature selection (MFS) method plays a pivotal role in big data era with the exponential growth of multi-label data. As the activity of the multi-label field increases, it also exposes some multi-label problems. Traditional graph-based methods adopt the original spaces to construct Laplacian graphs, which increases redundancy and noise. In addition, when exploring the deeper subtle connections between features and labels of multi-label samples, and extracting spatial correlations, there is a lack of consideration for shared connection information between feature space and label space. To track these tricky problems, this study adopts matrix factorization method to decompose the original matrix space and extract the low dimensional matrix. On the one hand, dynamic graph regularization preserves the spatial geometry and the interference of redundant features are reduced while extracting correlations. On the other hand, the decomposed low dimensional matrix obtains the shared connection information from the dual space of feature and label, which is beneficial for mining information from data. Furthermore, l2,1/2-norm is applied to the feature weight matrix to enforce row-sparsity and robustness. So this study proposes a robust MFS with Shared Coupled and Dynamic graph Regularization (SCDRMFS). An iterative method for solving the objective function is proposed, and its convergence is proved from two aspects, theoretically and experimentally. Moreover, experiments on nine real benchmark datasets are performed to verify the effectiveness of the proposed method. SCDRMFS is contrasted with six latest algorithms. It is concluded from the experimental results that the proposed algorithm SCDRMFS can improve the classification performance for multi-label datasets.