[R] party tree coordinates

[Achim Zeileis]

On Wed, 31 Oct 2012, aelmore wrote:

> I'm hoping that folks out there with expertise in working with the party
> package can help me out here.  My team is trying to convert party tree
> output into a text file format that can be read by our image processing
> software.  We are running into difficulties because the way the two
> different programs identify their nodes is different.
>
> R numbers it's nodes 1, 2, 3, 4, 5, 6 etc. down the "yes" answers in the 
> tree until a terminal node is reached, at which point it backs up to the 
> parent of that terminal node, gives its "no" child the next number 
> available, follows any additional yes children down from there until it 
> hits a terminal node, and repeats necessary.

And that's essentially also how the visualization works. A suitably sized 
grid is set up first and then grid viewports are recursively created and 
traversed.

> Our image processing system uses a coordinate system that labels each 
> node according to its location in the tree (e.g. row 4, column 43). 
> For instance, the first split in the tree is at location (1,1), the two 
> daughter nodes are (2,1) and (2,2), and their daughters are (3,1), 
> (3,2), (3,3), (3,4), etc. etc. on down the tree.  It assigns these 
> location coordinates referencing all possible node locations.  In other 
> words, if node (2,1) in the above scenario happened to be a terminal 
> node, the daughters of node 2,2 would still be labeled (3,3) and (3,4), 
> even though nodes (3,1) and (3,2) would not actually exist. They are 
> still, from a labeling/ID perspective, retained, as it were, as phantom 
> place holders.
>
> Given that the plot function in the party package is able to produce 
> output that shows these column and row relationships in graphical 
> format, it seems to me that there ought to be a way to extract this 
> positional information from the package, but I haven't found a way to do 
> it yet.

I think we haven't got anything in the package that delivers this out of 
the box. (Also note that the tree is often sparser than all binary splits 
that could occur.)

> The plot.tree info in the manual says it creates "an (invisible) list 
> with components x and y giving the coordinates of the tree nodes," and 
> I'm wondering if it's talking about "tree coordinates," as used by our 
> image processor, or if it's refering to the x and y position on the 
> printed plot. Either way, I don't know how to access those x/y numbers.

That's not from the "party" package anyways. plot.tree is from the "tree" 
package whose code bases are not related.

> At any rate, I'm pretty stumped.  It seems to me that since the 
> positional relationships are produced in the plots, there ought to be a 
> way to get at them, programmatically.
>
> Any suggestions out there?

I would suggest that you take a loot at the "partykit" package. It 
contains a reimplementation of ctree() based on acleaner design of the 
recursive tree structure. It also offers more functions for user 
interaction. It comes with a vignette that is still somewhat rough but 
should give you a good idea how things work.

hth,
Z