问题
I generated an (unpruned) classification tree on R with the following code:
fit <- rpart(train.set$line ~ CountryCode + OrderType + Bon + SupportCode + prev_AnLP + prev_TXLP + prev_ProfLP + prev_EVProfLP + prev_SplLP + Age + Sex + Unknown.Position + Inc + Can + Pre + Mol, data=train.set, control=rpart.control(minsplit=5, cp=0.001), method="class")
printcp(fit) shows:
Variables actually used in tree construction:
Age
CountryCode
SupportCode
OrderType
prev_AnLP
prev_EVProfLP
prev_ProfLP
prev_TXLP
prev_SplLP
Those are the same variables I can see at each node in the classification tree, so they are correct. What I do not understand is the result of summary(fit):
Variable importance:
29 prev_EVProfLP
19 prev_AnLP
16 prev_TXLP
15 prev_SplLP
9 prev_ProfLP
7 CountryCode
2 OrderType
1 Pre
1 Mol
From summary(fit) results it seems that variables Pre and Mol are more important than SupportCode and Age, but in the tree Pre and Mol are not used to split the data, while SupportCode and Age are used (just before two leafs, actually... but still used!). Why?
回答1:
The importance of an attribute is based on the sum of the improvements in all nodes in which the attribute appears as a splitter (weighted by the fraction of the training data in each node split). Surrogates are also included in the importance calculations, which means that even a variable that never splits a node may be assigned a large importance score. This allows the variable importance rankings to reveal variable masking and nonlinear correlation among the attributes. Importance scores may optionally be confined to splitters; comparing the splitters-only and the full (splitters and surrogates) importance rankings is a useful diagnostic.
Also see chapter 10 of book 'The Top Ten Algorithms in Data Mining' for more information https://www.researchgate.net/profile/Dan_Steinberg2/publication/265031802_Chapter_10_CART_Classification_and_Regression_Trees/links/567dcf8408ae051f9ae493fe/Chapter-10-CART-Classification-and-Regression-Trees.pdf.
来源：https://stackoverflow.com/questions/49217820/why-variable-importance-is-not-reflected-in-variable-actually-used-in-tree-const