Feature importances

[fukatani]

Are there any method to calculate feature_importances from RGF?

[StrikerRUS]

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.

[fukatani]

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.

[StrikerRUS]

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…

[fukatani]

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.

[StrikerRUS]

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.

[fukatani]

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.

[StrikerRUS]

Updates:

Feature importance for RGF is already implemented in #161.

[fukatani]

FastRGF is not supported yet. PR (to FastRGF and here) is welcome.