Multi-label learning is widely used in fields such as image annotation, text classification, and biological function prediction, but it faces challenges such as the curse of dimensionality in high-dimensional feature spaces, insufficient modeling of label correlations, and noise sensitivity. This paper proposes a hierarchical multi-label attribute reduction method based on granular-ball. Through a dynamic local density adjustment mechanism, a hierarchical information granulation method, and a progressive hierarchical reduction strategy, it addresses three limitations of traditional fixed-granularity strategies: inability to adapt to local density changes, insufficient label correlation modeling, and excessive noise sensitivity. The theoretical analysis reveals the error propagation limit and the noise robustness mechanism, confirming the effectiveness of the method in balancing attribute reduction efficiency and information granularity accuracy. Experimental results on six multi-label datasets show that the method improves the average accuracy by 9.2% on the emotions dataset, reduces hamming loss by 3.2% on the scene dataset, and only degrades performance by 2.17% under 25% label noise. This study provides a new solution for attribute reduction and information granulation in multi-label learning, with broad application prospects.