gbm_extract.log


R version 2.15.2 (2012-10-26) -- "Trick or Treat"
Copyright (C) 2012 The R Foundation for Statistical Computing
ISBN 3-900051-07-0
Platform: x86_64-w64-mingw32/x64 (64-bit)

R is free software and comes with ABSOLUTELY NO WARRANTY.
You are welcome to redistribute it under certain conditions.
Type 'license()' or 'licence()' for distribution details.

  Natural language support but running in an English locale

R is a collaborative project with many contributors.
Type 'contributors()' for more information and
'citation()' on how to cite R or R packages in publications.

Type 'demo()' for some demos, 'help()' for on-line help, or
'help.start()' for an HTML browser interface to help.
Type 'q()' to quit R.


This is a session spawned by NppToR.


[Previously saved workspace restored]

>
>
> require(earth) # for etitantic data
Loading required package: earth
Loading required package: leaps
Loading required package: plotmo
Loading required package: plotrix
> data(etitanic)
> mydata <- etitanic
>
> require(gbm)
Loading required package: gbm
Loading required package: survival
Loading required package: splines
Loading required package: lattice
Loaded gbm 1.6.3.2

> gbm1 <- gbm(survived ~ .,
+ data=mydata,
+ var.monotone=c(0,0,0,0,0),
+ distribution="bernoulli",
+ n.trees=2, # number of trees
+ shrinkage=0.005, interaction.depth=3, bag.fraction = 1, train.fraction = 1,
+ n.minobsinnode = 10, cv.folds = 5, keep.data=TRUE, verbose=FALSE)
>
>
> # Two full trees
> pretty.gbm.tree(gbm1, 1)
  SplitVar SplitCodePred LeftNode RightNode MissingNode ErrorReduction Weight    Prediction
0        1  0.000000e+00        1         5           9      73.139438   1046 -2.010714e-17
1        2  9.500000e+00        2         3           4       6.512112    658 -4.202694e-03
2       -1  3.584234e-03       -1        -1          -1       0.000000     43  3.584234e-03
3       -1 -4.747146e-03       -1        -1          -1       0.000000    615 -4.747146e-03
4       -1 -4.202694e-03       -1        -1          -1       0.000000    658 -4.202694e-03
5        0  1.000000e+00        6         7           8      19.437432    388  7.127249e-03
6       -1  1.354899e-03       -1        -1          -1       0.000000    152  1.354899e-03
7       -1  1.084503e-02       -1        -1          -1       0.000000    236  1.084503e-02
8       -1  7.127249e-03       -1        -1          -1       0.000000    388  7.127249e-03
9       -1 -2.010714e-17       -1        -1          -1       0.000000   1046 -2.010714e-17
> pretty.gbm.tree(gbm1, 2)
  SplitVar SplitCodePred LeftNode RightNode MissingNode ErrorReduction Weight    Prediction
0        1  2.000000e+00        1         5           9      72.409580   1046 -7.371885e-06
1        2  9.500000e+00        2         3           4       6.447161    658 -4.185715e-03
2       -1  3.563973e-03       -1        -1          -1       0.000000     43  3.563973e-03
3       -1 -4.727564e-03       -1        -1          -1       0.000000    615 -4.727564e-03
4       -1 -4.185715e-03       -1        -1          -1       0.000000    658 -4.185715e-03
5        0  3.000000e+00        6         7           8      19.243329    388  7.078582e-03
6       -1  1.347789e-03       -1        -1          -1       0.000000    152  1.347789e-03
7       -1  1.076960e-02       -1        -1          -1       0.000000    236  1.076960e-02
8       -1  7.078582e-03       -1        -1          -1       0.000000    388  7.078582e-03
9       -1 -7.371885e-06       -1        -1          -1       0.000000   1046 -7.371885e-06
>
> # Which variable is used for the first splitting rule (row 0)?
> pretty.gbm.tree(gbm1, 1)[0,]
[1] SplitVar       SplitCodePred  LeftNode       RightNode      MissingNode    ErrorReduction Weight         Prediction
<0 rows> (or 0-length row.names)
> pretty.gbm.tree(gbm1, 2)[0,]
[1] SplitVar       SplitCodePred  LeftNode       RightNode      MissingNode    ErrorReduction Weight         Prediction
<0 rows> (or 0-length row.names)
>
> # SplitVar is 1 in both trees, but it is zero based, so add 1.
> # So the splitting variable is 'sex'.
> attr(gbm1$Terms,'term.labels')[1+1]
[1] "sex"
>
> # Sex has two levels.
> (var.levels <- gbm1$var.levels[[2]])
[1] "female" "male"
>
> # The first rule of the first tree splits on SplitCodePred=0.
> # This must be zero based, so it must be female.
> var.levels[0+1]
[1] "female"
>
> # The first rule of the second tree splits on SplidCodePred=2.
> # Using the same "zero based" logic, now there is an out of bounds error.
> var.levels[2+1]
[1] NA
>
> # The splitting rules do not make sense, so try working backwards.
> # The first node is female
> mydata[1,'sex']
[1] female
Levels: female male
>
> # Predict the outcome of the first row.
> predict(gbm1, newdata=mydata[1,], n.trees=1, single=T)
[1] 0.01084503
> predict(gbm1, newdata=mydata[1,], n.trees=2, single=T)
[1] 0.0107696
>
> # These predictions correspond to these terminal nodes.
> pretty.gbm.tree(gbm1, 1)[8,]
  SplitVar SplitCodePred LeftNode RightNode MissingNode ErrorReduction Weight Prediction
7       -1    0.01084503       -1        -1          -1              0    236 0.01084503
> pretty.gbm.tree(gbm1, 2)[8,]
  SplitVar SplitCodePred LeftNode RightNode MissingNode ErrorReduction Weight Prediction
7       -1     0.0107696       -1        -1          -1              0    236  0.0107696
>
> # In both trees the terminal nodes trace back to row 0,
> # which has different SplitVar.  I am confused.
>