Semi-Supervised Multi-Label Feature Selection based on Sparsity Regularization and Dependence Maximization

Abstract

Multi-label learning deals with data samples that are simultaneously related to multiple labels. Likewise traditional single label learning, it is also difficult for multi-label learning to deal with high dimensional data. Feature selection is an effective solution for this difficulty via selecting a smaller subset from original features. Supervised multi-label feature selection techniques ask for sufficient labeled training data, however, the labeled data is often scarce. Several semi-supervise techniques which utilize labeled data and unlabeled data simultaneously have also been proposed. These semi-supervised solutions mainly use sparsity regularization and manifold assumption, as such they fail to take full advantage of the relation between features and labels. To make better use of the limited labels information to effectively select features, we introduced a semi-supervised multi-label Feature Selection algorithm based on sparsity regularization and dependence maximization (FSSRDM in short). FSSRDM utilizes the Hilbert-Schmidt Independence Criterion (HSIC) to capture the dependence between the features and labels, and adopt the $\ell_{2,1}$ sparsity term to obtain the regression coefficient sparse matrix. Experimental results show that FSSRDM is more effective than some state-of-art feature selection methods.