[R] CP for rpart

[Uwe Ligges]

weidong zhang wrote:

> Hi All,
> 
> I have some questions on using library rpart. Given my data below, the 
> plotcp gives me increasing 'xerrors' across different cp's with huge 
> xstd (plot attached). What causes the problem or it's not a problem at 
> all? I am thinking 'xerror's should be decreasing when 'cp' gets 
> smaller. 

No. Why? BTW: It's calculated by cross validation.


 > Also what the 'xstd' really tells us? If the error bars for
> each xerror overlap for different cp's, does that mean we don't have 
> significant improvement for misclassification rate when we split the tree?
 >
> 
> My data have are two classes with 138 observations and 129 attributes. 

Your problem is called overfitting, I guess.

Are you sure the class variable can be "explained" by the other 
variables in a way? Do you think the linear and othogonal separation of 
classes is appropriate for your problem?

You might want to look into a good book on Classification Theory.

Uwe Ligges


> Here is what I did:
> 
>> dim(man.dat[,c(1,8:136)])
> 
> [1] 138 130
> 
>> man.dt1 <- rpart(Target~.,data=man.dat[,c(1,8:136)], 
>> method='class',cp=1e-5, parms=list(split='information'))
> 
> 
>> plotcp(man.dt1)
> 
> 
>> printcp(man.dt1)
> 
> 
> Classification tree:
> rpart(formula = Target ~ ., data = man.dat[, c(1, 8:136)], method = 
> "class",
>    parms = list(split = "information"), cp = 1e-05)
> 
> Variables actually used in tree construction:
> [1] CHX.V  CYN.Cu SPF.Bi
> 
> Root node error: 25/138 = 0.18116
> 
> n= 138
> 
>       CP nsplit rel error xerror    xstd
> 1 0.18667      0      1.00   1.00 0.18098
> 2 0.00001      3      0.44   1.12 0.18897
> 
> 
> I would appreciate your help on this,
 >
> Weidong
> 
> _________________________________________________________________
> Instant message with integrated webcam using MSN Messenger 6.0. Try it 
> now FREE!  http://msnmessenger-download.com
> 
> 
> ------------------------------------------------------------------------
> 
> ______________________________________________
> R-help at stat.math.ethz.ch mailing list
> https://www.stat.math.ethz.ch/mailman/listinfo/r-help