Attribute reduction (also called feature subset selection) plays an important role in rough set theory. Different from the classical attribute reduction algorithms, the methods of attribute reduction based on covering rough sets appear to be suitable for numerical data. However, it is time-consuming in dealing with the large-scale data. In this paper, we study the problem of attribute reduction of covering decision systems based on graph theory. First, we translate this problem into a graph model and show that finding the attribute reduction of a covering decision system is equivalent to finding the minimal vertex cover of a derivative hypergraph. Then, based on the proposed model, a thm for covering decision systems is presented. Experiments show that the new proposed method is more effective to handle the large-scale data.