Unsupervised Feature Selection with Local Structure Learning

Conventional graph-based unsupervised feature selection approaches carry out the feature selection requiring two stages: first, constructing the data similarity matrix and next performing feature selection. In this way, the similarity matrix is invariably kept unchanged, totally separated from the process of feature selection and the performance of feature selection highly depends on the initially constructed similarity matrix. In order to address this problem, a novel unsupervised feature selection method is proposed in this paper where constructing similarity matrix and performing feature selection are together incorporated into a coherent model. Besides, the constructed similarity matrix has k connected components (k is the number of data clusters). At last, five state-of-the-art unsupervised feature selection methods are compared to validate the effectiveness of the proposed method.