Abstract: To solve the problem of feature redundancy, a multi-label learning and feature selection method based on uncorrelated regression and adaptive spectrum is proposed. The regression model with uncorrelated constraints was used to generate low-redundant but differentiated feature subsets, so that manifold learning and feature selection could be carried out simultaneously. The spectral term based on information entropy was introduced into the manifold framework to ensure the local geometric structure of the data in the subsequent learning process. Comprehensive experiments on multiple common multiple label data sets show that the proposed method can achieve high-dimensional data feature selection efficiently and accurately.