In　a　variety　of　text　classification　algorithm,　KNN　is　a　competitive　one　with　simple　implementation　and　high　efficiency.　However,　with　the　expansion　of　the　size　of　the　text,　the　runtime　of　KNN　will　grow　rapidly　that　cannot　be　afford.　In　this　paper,　we　improve　the　KNN　by　introducing　the　kd-tree　storage　structure　and　reducing　the　sample　space　through　the　sample　clestering　methods.　And　experiment　shows　that　the　runtime　of　improved　KNN　algorithm　reduce　apparently.　©2010　IEEE.