In multi-label learning, objects are essentially related to multiple semantic meanings, and the type of data is confronted with the impact of high feature dimensionality simultaneously, such as the bioinformatics and text mining applications. To tackle the learning problem, the key technology, i.e., feature selection, is developed to reduce dimensionality, whereas most of the previous methods for multi-label feature selection are either directly transformed from traditional single-label feature selection methods or half-baked in the label information exploitation, and thus causing the redundant or irrelevant features involved in the selected feature subset. Aimed to seek discriminative features across multiple class labels, we propose an embedded multi-label feature selection method with manifold regularization. To be specific, a low-dimensional embedding is constructed based on the original feature space to fit the label distribution for capturing the label correlations locally, which is also constrained using the label information in consideration of the co-occurrence relationships of label pairs. Following this principle, we design an optimization objective function involving l2,1-norm regularization to achieve multi-label feature selection, and the convergence is guaranteed. Empirical studies on various multi-label data sets reveal that the proposed method can obtain highly competitive performance against some state-of-the-art multi-label feature selection methods.