<p>Multi-label feature selection is a hotspot in multi-label learning, aiming to tackle the curse of dimensionality. Recently, several embedded models based on sparsity regularization have emerged. Most of them focus on learning an optimal feature selection matrix by means of regression, in which the correlation of instances and labels is concerned. However, the redundancy between features and the discriminative structure of labels have not been involved in. To argue these issues, a novel approach named discriminative identification of redundant features for multi-label feature selection&#xa0;(DIRF) is explored. In the proposed model, a feature affinity graph is constructed to find potentially redundant features with the idea that high similarity between features implies redundancy. Meanwhile, discriminative label correlation is revealed in terms of both label similarity and dissimilarity. Two regularizers are thereby designed to penalize the weights of redundant features. The structural consistency between original labels and predicted labels is therefore maintained. Extensive experiments and analysis show that the proposed DIRF outperforms the state-of-the-art multi-label feature selection methods.</p>