The discretization of real values is always one of the key problems to be solved in the domain of machine learning for its great contribution to speeding up the followed learning algorithms, cutting down the real demand of algorithms on running space and time, and improving the clustering capability of the ultimate learning results. The basic characteristics and framework of discretization approaches based on rough set model are analyzed at first, then the different measurements of the importance of candidate cuts are discussed and researched. Two new heuristic algorithms are put forward to finally select the useful cuts from a candidate set. The selected cuts of the two algorithms will adequately maintain the discernible relation of information systems for their full considering the specialty of rough set, which perfectly embodies the advantages of this theory. Moreover, excellent discretization results may be expected through these heuristic algorithms.