Abstract: Dimensionality reduction of data is a significant and challenging task under multi-label learning,and feature selection is a valid technology to reduce the dimension of vector.In this paper,a multi-label-specific feature selection method based on neighborhood rough set theory was proposed.This method ensures theoretically that there exists a strong correlation between the obtained label-specific features and the corresponding labels,and then reduction efficiency can be improved well.Firstly,a reduction algorithm of rough set theory is applied to reduce redundant attributes,and the label-specific features are obtained while keeping the classification ability unchanged.Then,the concepts of neighborhood accuracy and neighborhood roughness are introduced,the calculation approaches to dependence and attribute significance based on neighborhood rough set are redefined,and the related properties of this model are discussed.Finally,a multi-label-specific feature selection model based on neighborhood rough set is presented,and the corresponding feature selection algorithm for multi-label classification task is designed.The experimental results under some public datasets demonstrate the effectiveness of the proposed multi-label-specific feature selection method.

Key words: Multi-label learning,Neighborhood rough set,Label-specific feature,Feature selection