R help - Apr 2003 - Classification problem - rpart

I am performing a binary classification using a classification tree.
Ironically, the data themselves are 2483 tree (real biological ones)
locations as described by a suite of environmental variables (slope, soil
moisture, radiation load, etc). I want to separate them from an equal number
of random points. Doing eda on the data shows that there is substantial
difference between the tree and random classes, e.g., box and whisker plots
for slope show separation.

The data frame is thus:

curvegrid,dir2tl,dist2tl,slope,tasp,tci10,class
-0.000244141,266,1852.701,2.382412,0.2124468,131,random
0.3005371,246,1146.342,10.45694,0.8045813,63,random
.
.
.
.
-0.3000488,90,10,20.25561,-0.1293357,62,tree
-0.5,90,10,18.68057,-0.05228489,61,tree
-0.6994629,0,0,18.30121,0.0320744,66,tree

I've run rpart on similar data without an issue but when I try it on this
data as follows:

tree <- rpart(class ~ curvegrid + slope + tci10, method="class")

I get the following output:
> tree
n= 4966 

node), split, n, loss, yval, (yprob)
      * denotes terminal node

1) root 4966 2483 dw (0.500000000 0.500000000)  
  2) slope=0.3206026,0.5159777,0.679302,0.7163697,1.1324.......... 2574   94
dw (0.963480963 0.036519037) *
  3) slope=0,0.1011371,0.1013844,0.2027681,0.2267014,0.32......... MISSING
2392    3 random (0.001254181 0.998745819) *


This is not like other trees I have run!

And:

summary(tree)
> summary(tree)
Call:
rpart(formula = class ~ curvegrid + slope + tci10)
  n= 4966 

         CP nsplit  rel error    xerror       xstd
1 0.9609344      0 1.00000000 1.0322191 0.01418310
2 0.0100000      1 0.03906565 0.7635924 0.01378822

Node number 1: 4966 observations,    complexity param=0.9609344
  predicted class=dw      expected loss=0.5
    class counts:  2483  2483
   probabilities: 0.500 0.500 
  left son=2 (2574 obs) right son=3 (2392 obs)
  Primary splits:
     slope     splits as  RRRRRRLRRRLRRRRLLRRRRRRR.......
     tci10     splits as  RRRRRRRRRRLLRLLRLLRLLRLL.......

etc.

Node number 2: 2574 observations
  predicted class=dw      expected loss=0.03651904
    class counts:  2480    94
   probabilities: 0.963 0.037 

Node number 3: 2392 observations
  predicted class=random  expected loss=0.001254181
    class counts:     3  2389
   probabilities: 0.001 0.999

I'm assuming that I have to adjust something in rpart.control. I am also
hesitant at posting prematurely but am in fetters.

Thanks in advance, Andy

You have slope and tci10 (at least) coded as factors.  Run summary() on 
your data frame.  My bet is that you have MISSING in there, and did not
declare that when using read.table (or whatever)

Do *look* at your data at least cursorily.

On Thu, 10 Apr 2003, Andy Bunn wrote:
> I am performing a binary classification using a classification tree.
> Ironically, the data themselves are 2483 tree (real biological ones)
> locations as described by a suite of environmental variables (slope, soil
> moisture, radiation load, etc). I want to separate them from an equal
number
> of random points. Doing eda on the data shows that there is substantial
> difference between the tree and random classes, e.g., box and whisker plots
> for slope show separation.
> 
> The data frame is thus:
> 
> curvegrid,dir2tl,dist2tl,slope,tasp,tci10,class
> -0.000244141,266,1852.701,2.382412,0.2124468,131,random
> 0.3005371,246,1146.342,10.45694,0.8045813,63,random
> .
> .
> .
> .
> -0.3000488,90,10,20.25561,-0.1293357,62,tree
> -0.5,90,10,18.68057,-0.05228489,61,tree
> -0.6994629,0,0,18.30121,0.0320744,66,tree
> 
> I've run rpart on similar data without an issue but when I try it on
this
> data as follows:
> 
> tree <- rpart(class ~ curvegrid + slope + tci10,
method="class")
> 
> I get the following output:
> 
> > tree
> n= 4966 
> 
> node), split, n, loss, yval, (yprob)
>       * denotes terminal node
> 
> 1) root 4966 2483 dw (0.500000000 0.500000000)  
>   2) slope=0.3206026,0.5159777,0.679302,0.7163697,1.1324.......... 2574  
94
> dw (0.963480963 0.036519037) *
>   3) slope=0,0.1011371,0.1013844,0.2027681,0.2267014,0.32......... MISSING
> 2392    3 random (0.001254181 0.998745819) *
> 
> 
> This is not like other trees I have run!
> 
> And:
> 
> summary(tree)
> > summary(tree)
> Call:
> rpart(formula = class ~ curvegrid + slope + tci10)
>   n= 4966 
> 
>          CP nsplit  rel error    xerror       xstd
> 1 0.9609344      0 1.00000000 1.0322191 0.01418310
> 2 0.0100000      1 0.03906565 0.7635924 0.01378822
> 
> Node number 1: 4966 observations,    complexity param=0.9609344
>   predicted class=dw      expected loss=0.5
>     class counts:  2483  2483
>    probabilities: 0.500 0.500 
>   left son=2 (2574 obs) right son=3 (2392 obs)
>   Primary splits:
>      slope     splits as  RRRRRRLRRRLRRRRLLRRRRRRR.......
>      tci10     splits as  RRRRRRRRRRLLRLLRLLRLLRLL.......
> 
> etc.
> 
> Node number 2: 2574 observations
>   predicted class=dw      expected loss=0.03651904
>     class counts:  2480    94
>    probabilities: 0.963 0.037 
> 
> Node number 3: 2392 observations
>   predicted class=random  expected loss=0.001254181
>     class counts:     3  2389
>    probabilities: 0.001 0.999
> 
> I'm assuming that I have to adjust something in rpart.control. I am
also
> hesitant at posting prematurely but am in fetters.
It's not rpart, it is your data manipulation: `pilot error'.

-- 
Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595