训练集容量对决策树分类错误率的影响研究  

Proportion of Train-Set's Influence on Error-Rate of Decision Tree

作　　者：孙微微[1] 刘才兴[1] 田绪红[1] 

机构地区：[1]华南农业大学信息学院计算机科学与工程系,广州510642

出　　处：《计算机工程与应用》2005年第10期159-161,共3页Computer Engineering and Applications

基　　金：国家自然科学基金项目(编号:60375005)

摘　　要：数据挖掘算法必须在实际数据集上进行验证,而数据集容量是有限的,训练集比例过低会导致训练不足,训练集比例过高会导致算法评价过于乐观。针对训练集容量对评价效果的影响问题,对25个UCI数据集的不同比例训练集运用决策树算法C4.5,得出不同训练集容量对决策树分类错误率的影响关系。实验结果表明,训练集比例至少为50%时才能使分类错误率达到相对平稳。Algorithm in Data Mining must be validated upon real dataset,but the amount of sample in any dataset is limited.Excessively low proportion of train-set will cause inadequate training,and excessively high proportion of train-set will cause optimistic evaluation.For proportion of train-set's influence on evaluation,C4.5 is used upon different proportion of train-sets from 25 UCI dataset,Proportion of Train-Set's Influence on Error-Rate of Decision Tree describing is found.Results show that proportion of train-set needs to be at least 50% for a comparatively stable error-rate.

关 键 词：可能近似正确模型PAC 训练集 错误率 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]