@agramfort @glouppe Apologies for the ridiculous delay! The training now is done considering the missing values. Could you please take a look and tell me if my approach is correct?

I'll try to give you some feedback asap. Dont hesitate to ping if you dont see anything from me in the next days.

Also cc: @jmschrei @arjoly @pprett

@raghavrv I'm curious what the status of the smaller PRs are. Which of your branches those PRs correspond to?

I've been depending on this branch in my current work to handle missing values in my random forest training. It would be nice to see this fully merged into master, and I'd be willing to help out to get it there.

I've been depending on this branch in my current work to handle missing values in my random forest training. It would be nice to see this fully merged into master, and I'd be willing to help out to get it there

Oh really? Nice to hear! :) How did you find the performance and usability? Do you have any benchmarks comments or shortcomings to share. I'd be really interested... I could update it with current master if you want me to?

I got busy with another thing but the state of this PR is - Done and waiting for reviews. Except we have planned to avoid supporting all cases in the first go and instead add a small subset of what is here (say only for dense matrices, best splitter and depth first tree building) to help make the review a less daunting task...

I'm curious what the status of the smaller PRs are. Which of your branches those PRs correspond to?

Ah sorry that was a specific question... :) I've not yet made those smaller branches yet... I'll try to make them soon. Next week? (It's always motivating if someone finds your work useful ;))

I've only been using this on a basic level by seeing missing_values=np.nan when I create a RandomForestClassifier. I have a special hand-crafted dense feature vector that I need to perform probabilistic 3-state classification on (determine if two images show the same individual animal: yes, no, or not-enough-information). Some of the measurements like time and GPS on the photos are unreliable, and when building my feature vector some dimensions will be nan, these dimensions will be MAR. In some cases the fact that a dimension is nan might even have significance (MNAR), which is what led me to this PR.

I would link to an example of my code, but unfortunately we've had to close source the repo for political reasons until the work is published. Instead I can link a gist shows a toy example that I made when testing out this PR: https://gist.github.com/Erotemic/c9532be23e21a44a63d0cc77ed2e65fd

As far as rebasing the branch on master, that would be extremely helpful. I've had to write a script to do this and recover from conflicts because I need to integrate this functionality into sklearn whenever I install a python environment on a new machine.

@raghavrv latest rebase lives here https://github.com/unravelin/scikit-learn/tree/rf_missing. I need this branch to have a better control of handling missing values at the company I work in.

@Erotemic and others - @glemaitre and myself are working on a rewrite of the tree code to make it parallel (and if possible, faster even for single-threaded mode)... I'm unable to spend some bandwidth to keep this branch rebased :/

@afiodorov Thanks for sharing the rebase! Much appreciated... If you could keep it rebased and synced with latest master for another month or two, it would be awesome! I'll take it from then... Possibly by introducing the missing value functionality into the new rewrite...

@raghavrv I completely understand the scarcity of bandwidth. Its much more important to have the rewrite of the branch (and I'm very much hoping it includes missing value support).

Will this PR eventually support a random forest based missing data imputer in the vein of missForest? A quick read through the discussion suggests that it will not and will be mostly focused with fitting decision trees / RFs in the presence of NaN rather than imputing them. Is that correct or did I miss something? I ask because I was thinking about working on a missForest type Imputer but just wanted to make sure I was not duplicating any work. Would much appreciate your feedback. Thanks!

@ashimb9 this branch will allow for feature vectors passed to the fit method of the random forest to contain nan values.

I'm not familiar with what a missForest is, but I can tell you that this PR does not define any new Inputers.

@Erotemic Yeah, I did not think an Imputer was part of things but just wanted to make sure. Anyway, thank you for responding.

PS: In case you were wondering, missForest is an R package for imputing missing values using random forests.

Hey @raghavrv and others, do you know when this PR will be available in the production? It's definitely helpful. Thanks!

What is the status of this? Like many people I would be interested in the surrogate splitting method existing in the trees of sklearn.

@DrEhrfurchtgebietend perhaps see #5870. But note that raghavrv is no longer with us

Thank you for having explored this support.

In the meantime, the codebase evolved and #23595 is now more relevant than this PR, which I think can be closed.