As an important preprocessing procedure, dimensionality reduction for multi-label learning is an effective way to solve the challenge caused by high-dimensionality data. Most existing dimensionality reduction methods are mainly used to deal with single-label and multi-label data, which assumes each related label to the instance with the same important degree. However, there are different relatively important degrees for the related labels of each instance in many real applications. In this paper, a granular ball-based label enhancement algorithm is proposed to convert the logical label into label distribution for obtaining more supervision information. The granular ball can be regarded as local coarse grain to explore sample similarity based on neighborhood viewpoints. Then, the between-granular ball scatter and within-granular ball scatter measures are presented, which are utilized to construct a label distribution feature extraction algorithm. In addition, a two-stage mutual iterative learning framework is developed, label enhancement and dimensionality reduction are mutual interactive. Finally, Experiments are conducted with the six state-of-the-art methods on eleven multi-label data in terms of multiple representative evaluation measures. Experimental results show that the proposed method significantly outperforms other comparison methods by an average of 36.8% over six widely-used evaluation metrics.