2

I'm currently using decision trees (CART) in R with packages rpart and rattle for classification.

After training my CART tree, I found that some rules conflict with each other. Consider the following tree, with the conflicting rules indicated by the red circle.

enter image description here

In the parent node the split is CHWC.VLV >= 15; if this is true you go left in the tree and if it is false you go right in the tree. To the left, we find that the child node's rule is CHWC.VLV < 15. However based on the splitting rule in the parent node, I wouldn't expect any of the observations in this part of the tree to have values CHWC.VLV < 15.

Does anybody know the cause of this apparent conflict?

MichaelChirico
  • 33,841
  • 14
  • 113
  • 198
Sharp Yan
  • 345
  • 1
  • 9
  • 1
    One of the aspects of prediction that regression is not good with are NA's. Tree methods on the other hand have no difficulty with those. Are you sure this is not one of those splits? (Need data to do anything other than speculate.) – IRTFM Jan 28 '16 at 01:55
  • Also thanks the contribution from 42. – Sharp Yan Jan 28 '16 at 05:05

1 Answers1

3

This sort of issue generally comes from simply not outputting using enough digits of precision when outputting your CART tree. As a simple example, let's consider the following dataset:

CHWC.VLV <- seq(14, 16, length.out=10000)
outcome <- ifelse(CHWC.VLV >= 14.97, ifelse(CHWC.VLV <= 15.34, 1, 2), 3)

We can train and plot our CART model with:

library(rpart)
mod <- rpart(outcome~CHWC.VLV)
library(rpart.plot)
prp(mod)

enter image description here

This appears to be a contradiction, because the left subtree from the root node should have all values CHWC.VLV >= 15, but the next split is CHWC.VLV < 15. However, upon plotting with more digits of accuracy we see that this is, in fact, not a contradiction:

prp(mod, digits=4)

enter image description here

josliber
  • 43,891
  • 12
  • 98
  • 133