Open cspanbauer opened 1 year ago
Unfortunately, you are right. Currently, having more than one BART RV per model is not impossible.
Hi, are there any plans to having multiple BART trees in the same model to work into more complex types of interactions between the covariates. For example, what we see In this example in the PyMC documentation. Where we could replace some of the linear regressions with BART models.
The top priority is speeding up the current implementation and fixing some issues with parallel sampling. But allowing multiple BART RVs is something we should add at some point. Notice that you can have separated trees, under certain circumstances, see https://github.com/pymc-devs/pymc-bart/pull/98
Is it possible to define two BART RVs with different X and Y inputs in the same pymc-bart model? For example, using BART to estimate the nonlinear relationship for each link of a network? I seem to run into an error when trying a simple independent example for this that suggests this may not be possible so I wanted to be sure.
Assume we have Y1 = f1(X1) + e1 and Y2 = f2(X2) + e2 where e1 and e2 are independent errors for simplicity and X1 and X2 are distinct. I understand you can use 'size=2' when defining a BART RV, as in the heteroskedasticity example, but in this case we have two different inputs for X and Y. I'm guessing Y is needed to set the BART priors and X is used for the splitting rules. Is there a way to specify this model, or is it not currently possible? My use-case will eventually involve correlated errors, but that should be easy enough to incorporate with an LKJ prior and MvNormal as the likelihood specification.
Here's a simple example. Thanks in advance for any help you can provide.
Output: