Abstract: Aiming at the  problems that ignored by the existing embedded methods: the influence of the latent representation of instance
 correlation on pseudo-label learning, and the calculation error was caused by the fixed graph matrix, which increased with the deepening of iterations. We proposed a multi-label feature selection method with latent representation and dynamic graph constraints. Firstly, the proposed method used the latent representation of instance correlation to construct the pseudo-label matrix, and combined it with linear mapping and minimizing the Friedman norm distance between the pseudo-label and the ground-truth label to  ensure a high similarity between pseudo-labels and the ground-truth labels. Secondly, the dynamic graph was constructed by using the low-dimensional manifold structure of pseudo-labels to alleviate the problem of increasing calculation error with iteration depth caused by a fixed graph matrix.  The comparative experimental results with seven advance methods on 12 datasets show that the overall classification performance of the proposed method is superior to  the existing advanced methods, and it  can better deal with multi-label feature selection problems.

Key words: multi-label learning, feature selection, latent representation, dynamic graph, manifold learning