Abstract: Multi-label classification is a kind of complex decision making tasks, where one object may be assigned with more than one decision label. This kind of tasks widely exist in text categorization, image recognition, gene function analysis. Multi-label classification is usually described with high-dimensional vectors, and some of the features are superfluous and irrelevant. A great number of feature selection algorithms have been developed for single-label classification to conquer the curse of dimensionality. However, as to multi-label classification, fewer researches have been reported for designing feature selection algorithms. In this work, we introduce rough sets to multi-label classification for constructing a feature selection algorithm. We redefine the lower approximation and dependency, and discuss the properties of the model. After that, we design a neighborhood rough sets based feature selection algorithm for multi-label classification. Experimental results show the effectiveness of the proposed algorithm.