With regard to the attribute values in decision table, which are described with hybrid data, a new algorithm of attribute reduction based on rough set theory is proposed. First, the similarity relations among objects with hybrid data are defined. In order to obtain reasonable soft partitions among objects, the optimization model for threshold accounting is presented. Then, based on the upper and lower approximation concept from rough set theory, the covering upper and lower similar partitions among objects are obtained. In succession, through descriptions of the upper and lower similar distribution matrixes found on condition attributes and decision attribute, the two attribute reduction results of different viewpoints can be retrieved intuitively, based on the max-distribution matrixes. Finally, the experiment results prove that this algorithm is effective and feasible.