Open fukatani opened 6 years ago
There is no built-in method to do it. I suppose the only method to do it is to parse .model
file, like https://github.com/Far0n/xgbfi for XGBoost. But in my opinion it is out of scope of rgf_python
project.
I'm not stuck with to support it in this repository.
However, both RGF
and FastRGF
seem inactive, we should implement here or make new repository.
Maybe... But I have a conversation with Tong Zhang this week and that's what he said about FastRGF:
There are many ways to improve the current code base. While I am busy now, I still hope to find time at some point to upgrade the code either myself or with the help of someone else…
I see.
I think that there are people who would like to help. I am as well. (But unfortunately, I don't have expertise on decision trees.) I want him to upload his plan as issue. Also, FastRGF may need collaborater who can merge PR.
Also, FastRGF may need collaborater who can merge PR.
I think it won't be easy because fast_rgf
is baidu's repo, not his personal one.
Unfortunately, he haven't answered to my last letter:
It's great to hear that you won't abandon the project. Unfortunately, I'm not very much familiar with C++, so could not help to improve the code. But the guy with whom we are working on Python wrapper told me that he found some bug with discretization - I'll tell him to create a PR in the FastRGF repo. Also I believe that our wrapper will help RGF to gain more popularity and bring more contributors on GitHub. From my side I'll create a PR with a fix of common issue while compiling with MinGW.
and, as you can see, he haven't merged the PR yet.
Well, let's see if users of FastRGF will increase. If it will become popular, it may be worth to fork by us or anyone.
Updates:
Feature importance for RGF is already implemented in #161.
FastRGF is not supported yet. PR (to FastRGF and here) is welcome.
Are there any method to calculate feature_importances from RGF?