Graph regularized virtual label regression for unsupervised feature selection

Abstract

Feature selection is an important technique to deal with high-dimensional unlabeled data. Over the past decades, various unsupervised feature selection methods have been presented. However, some of these methods fail to consider the discriminative ability due to lack of label information or ignoring the local structure. To address this issue, we propose a novel unsupervised feature selection approach, called graph regularized virtual label regression (GVLR). In GVLR, we learn a virtual label matrix to guide the selection of features, which can ensure the selected features be more discriminative. Besides, we utilize graph Laplacian to preserve the geometric structure of the feature space. To improve the robustness of our model, we further adopt an ℓ2,1-norm constraint on the feature selection matrix and loss function. Finally, we incorporate the label regression and graph Laplacian into the subspace learning based feature selection framework. Thus, our method can enhance the discriminative ability of selected features, and preserve the local structural information of features simultaneously. Extensive experimental results on several public datasets show that the proposed GVLR method can obtain superior performance in comparison with some state-of-the-art feature selection algorithms.