Joint learning of graph and latent representation for unsupervised feature selection

Data samples in real-world applications are not only related to high-dimensional features, but also related to each other. To fully exploit the interconnection between data samples, some recent methods embed latent representation learning into unsupervised feature selection and are proven effective. Despite superior performance, we observe that existing methods first predefine a similarity graph, and then perform latent representation learning based feature selection with this graph. Since fixed graph is obtained from the original feature space containing noisy features and the graph construction process is independent of the feature selection task, this makes the prefixed graph unreliable and ultimately hinders the efficiency of feature selection. To solve this problem, we propose joint learning of graph and latent representation for unsupervised feature selection (JGLUFS). Different from previous methods, we integrate adaptive graph construction into a feature selection method based on the latent representation learning, which not only reduces the impact of external conditions on the quality of graph but also enhances the connection between graph learning and latent representation learning for benefiting the feature selection task. These three basic tasks, including graph learning, latent representation learning and feature selection, cooperate with each other and lead to a better solution. An efficient algorithm with guaranteed convergence is carefully designed to solve the optimization problem of the algorithm. Extensive clustering experiments verify the competitiveness of JGLUFS compared to several state-of-the-art algorithms.