[FEATURE] Implement `tsBART`

[jdtuck]

Is your feature request related to a problem? Please describe.
It would be good to have the extension to tsBART [1]. Have the option for univariate functions on the terminal nodes would be a useful feature

[1] Starling, J. E., Murray, J. S., Carvalho, C. M., Bukowski, R. K. & Scott, J. G. BART with targeted smoothing: An analysis of patient-specific stillbirth risk. Ann. Appl. Stat. 14, 28–50 (2020).

[andrewherren]

Hi there! Thanks for opening this feature request. I agree that tsBART is a super interesting and useful model. Two quick notes:

1. Implementing functional leaf models is a long term interest of ours, but it's unlikely to get done in the near term without external contributions. PRs are certainly welcome!
2. The current stochtree "leaf regression" interface (i.e. passing a multivariate leaf_basis_train to the bart() function in R or to the sample() method of BARTModel in Python) can get you most of the way there right now, if you basis-expand the variable that you want to GP in the leaf nodes

On (2), I pulled up some code that a collaborator wrote to do this awhile back and I have it as a TODO item to clean that code up and publish a vignette on it. (We've recently harmonized the R and Python vignettes so that they sit in the same multilingual documents, see here.) I will follow up when the vignette demonstrating this functionality is live.

In the meantime let me know if you have any questions or want to discuss what it would take to contribute this feature to stochtree!

[andrewherren]

Hi @jdtuck, just following up with a link to the vignette on approximating tsBART with a leaf regression on a basis expansion of the target variable: https://stochtree.ai/vignettes/targeted-smoothing.html

Let me know if you have any questions or are interested in contributing an implementation of the full leaf-GP model in stochtree at some point