Thanks Terry!
I managed to figure that out shortly after posting (as is the way!) Adding an
additional covariate that splits below one of the x branches but not the other
and means the class proportion to go over 0.5 means the x split is retained.
However, I now have another conundrum, this time with rpart in anova mode...
library(rpart)
test_split <- function(offset) {
y <- c(rep(0,10),rep(0.5,2)) + offset
x <- c(rep(0,10),rep(1,2))
if (is.null(rpart(y ~ x, minsplit=1, cp=0, xval=0)$splits)) 0 else 1
}
sum(replicate(1000, test_split(0))) # 1000, i.e. always splits
sum(replicate(1000, test_split(0.5))) # 2-12, i.e. splits only sometimes...
Adding a constant to y and getting different trees is a bit strange,
particularly stochastically.
Will see if I can track down a copy of the CART book.
Jonathan
________________________________________
From: Therneau, Terry M., Ph.D. [therneau at mayo.edu]
Sent: 16 May 2017 00:43
To: r-help at r-project.org; Marshall, Jonathan
Subject: Re: Odd results from rpart classification tree
You are mixing up two of the steps in rpart. 1: how to find the best candidate
split and
2: evaluation of that split.
With the "class" method we use the information or Gini criteria for
step 1. The code
finds a worthwhile candidate split at 0.5 using exactly the calculations you
outline. For
step 2 the criteria is the "decision theory" loss. In your data the
estimated rate is 0
for the left node and 15/45 = .333 for the right node. As a decision rule both
predict
y=0 (since both are < 1/2). The split predicts 0 on the left and 0 on the
right, so does
nothing.
The CART book (Brieman, Freidman, Olshen and Stone) on which rpart is based
highlights the
difference between odds-regression (for which the final prediction is a percent,
and error
is Gini) and classification. For the former treat y as continuous.
Terry T.
On 05/15/2017 05:00 AM, r-help-request at r-project.org
wrote:> The following code produces a tree with only a root. However, clearly the
tree with a split at x=0.5 is better. rpart doesn't seem to want to produce
it.
>
> Running the following produces a tree with only root.
>
> y <- c(rep(0,65),rep(1,15),rep(0,20))
> x <- c(rep(0,70),rep(1,30))
> f <- rpart(y ~ x, method='class', minsplit=1, cp=0.0001,
parms=list(split='gini'))
>
> Computing the improvement for a split at x=0.5 manually:
>
> obs_L <- y[x<.5]
> obs_R <- y[x>.5]
> n_L <- sum(x<.5)
> n_R <- sum(x>.5)
> gini <- function(p) {sum(p*(1-p))}
> impurity_root <- gini(prop.table(table(y)))
> impurity_L <- gini(prop.table(table(obs_L)))
> impurity_R <- gini(prop.table(table(obs_R)))
> impurity <- impurity_root * n - (n_L*impurity_L + n_R*impurity_R) #
2.880952
>
> Thus, an improvement of 2.88 should result in a split. It does not.
>
> Why?
>
> Jonathan
>
>