R: interpret package features

[ClaudiuPapasteri]

First of all, thank you for the brilliant ML technique you developed. I read some of the Python tutorials and decided to replicate some of them in R.

These are just a few of my early observations from using the package:

• Model fitting is reasonably fast for small and medium samples
• Very easy to use
• The documentation is very lacking

For binary classification

• no formula syntax (I get that it is based on Python, but in R formula class is very practical; it will be useful when considering implementing user-specified interactions)
• target needs to be numeric 0/1, factors seem unsupported
• predict function only for "prob", not for "class" (adding a type arg to ebm_predict() with values "class" and "prob" is fairly consistent in R)
• no pairwise interactions
• the ebm_show method for single features is informative, although it is only the global explainer and the local explainer is not yet implemented

No regression algorithm (I saw in source code that it's on TODO list)

I am eager to use interpret in my analyses so I have to ask:

1. When are you planning to implement the regression fitting and prediction functions?
2. Are you considering aligning the package with the tidymodels framework? I think it would fit right in.

Thanks again.

[interpret-ml]

Hi @ClaudiuPapasteri,

Thank you for the detailed notes, they are very helpful! Aligning with tidymodels in particular is a fantastic idea.

We've currently been focusing on making improvements to our Python package and shared C++ core layer, which we hope to eventually port to R. Unfortunately it's hard to put a timeline on when we'll be able to revisit the R package, so we'd recommend using the Python package if possible for the time being. As I think you've seen, most of the features you've requested (outside of the formula syntax) are present in our python package. We will update this issue if we have any updates on the R package side in the future!

-InterpretML Team