In multi-label learning, the curse of dimensionality is one of major challenges. Existing single-label feature selection methods cannot be directly applied to multi-label data, and multi-label feature selections have thus been widely studied. As an effective granular computing tool, rough set theory has been applied to multi-label feature selections for addressing various realistic applications. However, existing rough set based methods not only cannot effectively characterize the ability of features to distinguish multi-label sample pairs, but also usually own high time complexity. In this paper, we propose two novel multi-label feature selection methods from the perspective of discerning sample pairs. First, relative discernibility pair matrixes of features are defined in the framework of fuzzy rough set, where each element represents the degree of distinguishing the corresponding sample pair by features. On this basis, a novel evaluation measure of feature subsets is defined. Afterwards, a heuristic multi-label feature selection approach titled RDPM based on the proposed measure is put forward. Inspired by sampling and ensemble strategies, another efficient and robust multi-label feature selection approach titled RDPM_SE is proposed as well. Finally, extensive experiments on thirteen real-world multi-label data sets are conducted, and experimental results show that the proposed algorithms outperform seven state-of-the-art methods in terms of performances and the running time.