R partykit calculate classification probabilities on sub stree

Question

I have trained a partykit package ctree classification decision tree and I need to calculate classification probabilities for sub tree (not only for leaf nodes). So for example if a sub tree consists of 3 leaf nodes with the following probabilities: leaf 1 (120 observations) : 0.45 leaf 2 (160 observations) : 0.49 leaf 3 (190 observations) : 0.83

for this hypothetical sub tree the weighted average probability would be 120*0.42 + 160*0.49 + 190*0.83 / (120+160+190) = 0.507

and so on I need to traverse on the ctree object and calculate all weighted probabilities for each node recursively.

I have this code:

data(airquality)
airq <- subset(airquality, !is.na(Ozone))
airct <- ctree(Ozone ~ ., data = airq,
                 controls = ctree_control(maxsurrogate = 3))
traverse <- function(treenode){
    if(treenode$terminal){
      bas=paste("Current node is terminal node with",treenode$nodeID,'prediction',treenode$prediction)
      print(bas)
      return(0)
    } else {
      bas=paste("Current node",treenode$nodeID,"Split var. ID:",treenode$psplit$variableName,"split value:",treenode$psplit$splitpoint,'prediction',treenode$prediction)
      print(bas)
    }
    traverse(treenode$left)
    traverse(treenode$right)
  }

which traverse on the tree does not work on partykit object. On the other hand I have this code, which lists all porbabilities for leaf nodes only :

preds.ls <- list(predict(airct , type = "prob"))[1]
pred.probs.df <- unique(as.data.frame((preds.ls[[1]])))

Any suggestions to combine these 2 snippets to a code that will traverse on a PARTYKIT object and calculate this weighted average are appreciated

It's not quite clear to me what exactly want to do as the code as posted contains some errors. However, I think that this answer will help you do what you want (or ask a more precise question): http://stackoverflow.com/questions/41968910/r-extracting-inner-node-information-and-splits-from-ctree-partykit/41976697#41976697 — Achim Zeileis, Mar 22 '17 at 11:37

Zelazny7 · Accepted Answer · 2017-03-21T14:12:36.837

0

I'm not familiar with partykit but this simple function walks a ctree and extracts the probability for every internal and terminal node:

   library(party)

    set.seed(100)
    dt <- ctree(factor(mpg > 20)~., data = mtcars,
                control = ctree_control(minsplit=2, minbucket=1, mincriterion=0))

    traverse <- function(node) {
      if (node$terminal) {
        return(node$prediction[2])
      }
      return(c(node$prediction[2],
               traverse(node$left), traverse(node$right)))
    }

Calling the function produces the following vector of probabilities:

> traverse(dt@tree)
[1] 0.4375000 1.0000000 0.1428571 0.4285714 0.7500000 0.0000000 0.0000000

The left most value is the population value verified by the following:

> mean(mtcars$mpg > 20)
[1] 0.4375

The rest of the values are going to be in order from left to right. You can see that the 1s and 0s line up where expected.

edited Mar 21 '17 at 14:12

answered Mar 21 '17 at 14:02

Zelazny7

39,946
18
70
84

Does this implementation take under consideration the number of observations for each leaf/terminal node ? – NRG Mar 21 '17 at 15:15
Yes, the internal node probabilities are the probabilities of the entire subtree. – Zelazny7 Mar 21 '17 at 15:20
Note that this solution pertains to the old `party` implementation but does not work for the newer `partykit` implementation. The representation of the tree changed significantly. – Achim Zeileis Mar 22 '17 at 11:36
@ Achim Zeileis - Do you have any reference to a partykit solution to calculate sub tree probability ? – NRG Mar 23 '17 at 08:28

R partykit calculate classification probabilities on sub stree

1 Answers1