决策树剪枝方法的比较_魏红宁.pdfVIP

40 1 Vol. 40 No. 1 2005 2 JOURNAL OF SOUTHWEST JIAOT ONG UNIVERSIT Y Feb. 2005 : 0258-2724( 2005) 01-0044-05 魏红宁 ( , 6 10031) : , 4 . PEP , MEP ; REP , ; , CCP REP . , REP, , PEP . : ; ; ; PEP; MEP; REP ; CCP : TP311 : A Comparison among Methods of Decision Tree Pruning WEI H ong-ning ( Administr tive Office, Southwest Ji otong University, Chengdu 610031, Chin ) Abstract: To select suit ble pruning method in decision tree pruning, four w el-l known pruning methods w ere comp red in terms of comput tion l complex ity, tr vers l str tegy , error estim tion nd theoretic l principle by t king cl ssific tion nd regression tree s n ex mple. Comp red with pessimistic error pruning ( PEP) , minimum error pruning ( MEP) is less ccur te nd produces l rger tree. Reduced error pruning ( REP) is one of the simplest pruning str tegies, but it h s the dis dv nt ge of requiring sep r te d t set for pruning. Cost-complexity pruning ( CCP) produces sm ller tree th n REP with simil r ccur cy. Pr ctic lly, if the tr ining d t is bund nt, REP is prefer ble; nd if the tr in d t is the expected ccur cy is high but w ith limited d t , PEP is good choice. Key words: d t mining; decision tree; post pruning; pessimistic error pruning ; minimum error pruning ; reduced error pruning; cost-complex ity pruning . . , . , , , , . , , [ 1] . , . [ 2] , , . , ; . , , . ( 1) CART [ 3] ( cl ssific tion nd regression trees) , , . 1 t