This might be related to #27993 , although it is more so related to multi-task random forests

Hi @adam2392 ,

If you could help me, I have some questions about this Issue.

1. Is this Issue still open, and is it worth working on it? I followed the history associated, and it seems that the first conversations about it started 10 years ago.

2. Do you think this is "easy enough" to be a good "second issue?" Because my first and only contribution was a few weeks ago. @glemaitre cc here just because your comment on a related Issue. Also, I would like a "second opinion" on this question.

Hi @vitorpohlenz !

Thank you for your interest in contributing!

1. Yes it is still open and it is worth working on it.
2. Is it easy? That depends on your experience with Cython and/or C++ and also understanding of decision trees. It also depends on how deep you want to dive into the code. The main thing to add is the same tree logic handling for missing values for sparse array inputs.

Lmk what you think and if you have a proposed idea of how to proceed.

Hi @vitorpohlenz !

Thank you for your interest in contributing!

1. Yes it is still open and it is worth working on it.
2. Is it easy? That depends on your experience with Cython and/or C++ and also understanding of decision trees. It also depends on how deep you want to dive into the code. The main thing to add is the same tree logic handling for missing values for sparse array inputs.

Lmk what you think and if you have a proposed idea of how to proceed.

Thanks for replying @adam2392 !

About [2.] I have no previous experience with Cython, and just a little experience with C++ (I have worked a little with C, but it was 10 years ago). But I understand the theory of decision trees. I know about the CART algorithm and the ways to measure the impurity using metrics like entropy, gini, etc...

So, I would like to ask your opinion if you think it is a good idea for me to work on this issue(based on your expertise and this piece of information), to get more acquainted with the tree-based code in sklearn.

Also, if I could ask for some guidance/advice to not struggle too much, if necessary

Hi @vitorpohlenz

So, I would like to ask your opinion if you think it is a good idea for me to work on this issue(based on your expertise and this piece of information), to get more acquainted with the tree-based code in sklearn.

Cython / C++ knowledge is definitely needed within the scikit-learn dev/reviewer community, so anyone interested is definitely welcome! There will be a learning curve, and require you to explore how Cython works, and what doesn't work. I think it's a net-positive to your skillset.

With that being said, yes you can ask for guidance / advice, but note my time is a bit limited these days, so will definitely require a lot of self-learning.

As a first step, I would look at this PR: #27966, review how Cython works. Then what's needed is for one to:

1. understand how sparse arrays work
2. understand how missing-values are split in the sklearn RF algo
3. apply that understanding to the sparse splitter logic

Feel free to propose/iterate on this plan accordingly!

Hi @adam2392, thanks a lot for replying!

With that being said, yes you can ask for guidance / advice, but note my time is a bit limited these days, so will definitely require a lot of self-learning.

Sure, I understand that this task will be more self-driven than based on external aid. But this kind of guidance, like you did, will help me a lot.

So I will /take it and will start by the steps you suggested, and will update the progress here until I have a clear proposal of enhancement.

Update about the progress/steps that I'm doing:

• Understanding how sparse arrays work
• Understanding how missing values are split in the sklearn Random Forest algorithm
• Define how to apply the modifications to the sparse splitter logic