Abstract: Attribute reduction is one of key problem in the theoretical research of Rough set. This paper improves HORAFA algorithm based on Frequency of attribute appears in the discernibility matrix. It takes the core as foundation, and joins the most important attributes in the set of attributes until cannot join. In order to find optimum reduction of information system, this paper adds the converse eliminate action until cannot delete. We give a demonstration at last section of the paper, and it verify improved method validity.