Add possibility to run classification

[Vlad0922]

No description provided.

[RobbieShan]

+1 if what you mean is an ability to use MERFs for classification (especially multiclass)

[resdntalien]

@Vlad0922 @RobbieShan Thanks for the issue. The algo for classification is significantly more difficult (and slower) so it will take some time to put in. But from Prof. Denis Laroque:

"We have a generalization of this method for outer types of outcomes. The attached paper described GMERT which builds a single tree but it can be generalized easily to a forest. However, it is a lot slower that MERT. Personnaly, for a binary outcome, I would simply use MERF as it is. The prediction becomes a predicted probability and you just need a cut-of-point to assign a class (0 or 1). It does not really matter if you get predicted probabilities below 0 or above 1, as the final prediction is a class."

[richardneilbelcher]

Would be great to see this feature too. Is their a planned implementation on the horizon?

[resdntalien]

Richard, thanks for the issue. Not on the near horizon, but in the future. The convergence is really really slow, so may not be the best. I would recommend casting as regression and clipping for now. Probably similar results.

…

Sent from my iPhone

On Jul 24, 2019, at 8:57 AM, richardneilbelcher ***@***.***> wrote: Would be great to see this feature too. Is their a planned implementation on the horizon? — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.

[MarconiS]

Hi! Having that functionality could be very useful! I was wondering if substituting yi - f_hat_i to kl_divergence(yi, f_hat_i) would consistently break the math behind the EM. In theory, that should somehow measure the error between the observed class and predicted vector of probabilities?

[mariadelmarq]

Am I understanding correctly that currently MERF cannot be used for multi-class classification?

[gizemtanriver]

Hi, has anyone tried MERF with a binary outcome (using a cut off point for assigning a class)? I wonder how it performs on such task and if it is straightforward to adjust the code for it...Would be great if anyone could share their experience!

[simonprovost]

+1 interested in people having used RandomForestClassifier and not regressor, see how it goes :) ?

[Nicobruno92]

Hi everyone, I am interested in this issue. Has anybody tried it out? How was your implementation and success? Thx