This paper presents a fast reduction algorithm for hierarchical classification, aiming to address the high time cost of calculation in neighborhood granulation and the limitation of existing hierarchical classification models that only consider empirical error. A median-based neighborhood granulation method is proposed, which enhances the efficiency of neighborhood granulation and is less affected by noise. The lower approximation quality in the hierarchical decision system (HieDS) is redefined, taking into account the generalization error. An attribute reduction algorithm under HieDS is also put forward. Finally, through experimental comparisons, our method has significantly improved both in reduction efficiency and classification accuracy.