Extremely randomised trees in caret using ranger

[hadjipantelis]

Hello,

Since version 0.6.4 the package ranger supports Extremely randomised trees. Strictly-speaking the only "new thing" would be to actually change the argument splitrule = "extratrees", nevertheless I have the feeling that people would miss out using extr. random trees if they are not careful. There is a package, which is included in caret already, extraTrees that implements the same algorithm. That implementation though does not do multi-core computation out-of-the-box (and its code-base hasn't change in the last 2 years either, which worries me too) so I think there is a case to use ranger directly.

Should one:

1. Let things as is and hope/trust the user to find the e.r.t. functionality themselves or
2. Add a method like: rangerExtreme so users can directly invoke extremely randomise trees?

Let me note that 0.6.4 is not available on CRAN, so you might want to nap the issue of adding the e.r.t. functionality until it is included in the CRAN release?

All best,
Pantelis

[topepo]

My thought is that we should have a different method and update caret when the newer ranger is on CRAN. I don't have a good way of checking for those package dependencies in the package generally, so a check could be written into the method (and it can be removed after a year so some suitable time).

[hadjipantelis]

Perfect; I think what you suggested is very reasonable.

[topepo]

I'm coding this up but it looks like it might need to wait until the next release. Now that I've looked into it, I will just make splitrule a tuning parameter but have it default to gini/variance. There seems to be a bug in ranger so it won't happen right now.

[hadjipantelis]

No problem; thank you for following this up.

[topepo]

I'll try to put this together for the imminent CRAN release since the CRAN ranger is at 0.8.0.

[topepo]

I decided to incorporate the splitting rule into the basic ranger model. With regular grid search, it will try to usual splitting rule and the randomized trees. That is also included in random search too.

[topepo]

Give it a spin and let me know if there are any issues.

[hadjipantelis]

Thank you for this! I agree, having gini/extratrees as a choice is neater.

Unfortunately, no spin out of the box because of the is.tbl call. I just changed the call to dplyr::is.tbl and is fine. Thank you! Having extra-tree was really missing from caret (and in R in general) It should work out of the box with PR #711.

[hadjipantelis]

Feel free to close this issue. It seems sorted.