The　discretization　of　continuous　attributes　is　an　important　method　for　compressing　data　and　simplifying　analysis,　which　is　also　the　focuses　in　the　research　fields　of　pattern　recognition,　machine　learning　and　Rough　Set　analysis.　At　present,　there　have　existed　many　algorithms　of　discretization.　The　main　problem　of　them　is　that　the　choice　of　breakpoints　which　influence　the　effect　of　discretization　deeply　is　random.　The　optimal　discretization　has　been　proved　to　be　NP-hard.　A　great　majority　of　algorithms　of　discretization　adopt　heuristics　which　is　hard　to　get　satisfied　discrete　effect.　Based　on　the　theory　of　Rough　Set　this　thesis　firstly　discusses　the　problems　above　then　puts　forward　a　new　algorithm　based　on　clustering　of　mean　and　importance　of　attributes.