BART: add partial dependence plots and individual conditional expectation plots #5091
Conversation
Codecov Report
@@ Coverage Diff @@
## main #5091 +/- ##
==========================================
+ Coverage 75.10% 77.77% +2.67%
==========================================
Files 87 88 +1
Lines 14142 14080 -62
==========================================
+ Hits 10621 10951 +330
+ Misses 3521 3129 -392
|
|
Can you post an example plot of what it looks like? Also, have you considered spawning this out into a separate package |
|
I have not considered a separated package. But not opposed to that. Most of the logic for BART is now on its own module. The only exception is the code related to the automatic assignment of PGBART sampler to BART variables. |
|
Maybe we can harbor these more specialized modules until they grow stable/popular and then spin them off to their own packages? |
|
Maybe that's a good strategy. One spin off library could be a more machine-learning/Bayesian non-paramateric kind of library and not just a BART library. Additionally one interesting feature of our BART implementation is precisely that is PPL friendly. There is still a lot of room for improvement and exploration here. Both at the implementation level and at the level of the types of model we can build and their uses. |
|
@twiecki @ricardoV94 can we merge this? |
|
@aloctavodia do you want to add a release note about the new features? |
This add functions to compute partial dependence plots and individual conditional expectation plots. It also move the in-sample and out-of-sample function predictions to the function
predict, also part ofbart.utils. After this PR is merged I will add an example of how to use these function and BART in general.This is an example of a partial dependence plot, where BART is used to predict the number of rented bikes given 4 covariates.