[scikit-learn] random forests and multil-class probability
Sole Galli
solegalli at protonmail.com
Tue Jul 27 05:31:36 EDT 2021
Thank you!
I was confused because in the multiclass documentation it says that for those estimators that have multiclass support built in, like Decision trees and Random Forests, then we do not need to use the wrapper classes like the OnevsRest.
Thus I have the following question, if I want to determine the PR curves or the ROC curve, say with micro-average, do I need to wrap them with the 1 vs rest? Or it does not matter? The probability values do change slightly.
Thank you!
‐‐‐‐‐‐‐ Original Message ‐‐‐‐‐‐‐
On Tuesday, July 27th, 2021 at 11:22 AM, Guillaume Lemaître <g.lemaitre58 at gmail.com> wrote:
> > On 27 Jul 2021, at 11:08, Sole Galli via scikit-learn scikit-learn at python.org wrote:
> >
> > Hello community,
> >
> > Do I understand correctly that Random Forests are trained as a 1 vs rest when the target has more than 2 classes? Say the target takes values 0, 1 and 2, then the model would train 3 estimators 1 per class under the hood?.
>
> Each decision tree of the forest is natively supporting multi class.
>
> > The predict_proba output is an array with 3 columns, containing the probability of each class. If it is 1 vs rest. am I correct to assume that the sum of the probabilities for the 3 classes should not necessarily add up to 1? are they normalized? how is it done so that they do add up to 1?
>
> According to the above answer, the sum for each row of the array given by `predict_proba` will sum to 1.
>
> According to the documentation, the probabilities are computed as:
>
> The predicted class probabilities of an input sample are computed as the mean predicted class probabilities of the trees in the forest. The class probability of a single tree is the fraction of samples of the same class in a leaf.
>
> > Thank you
> >
> > Sole
> >
> > scikit-learn mailing list
> >
> > scikit-learn at python.org
> >
> > https://mail.python.org/mailman/listinfo/scikit-learn
More information about the scikit-learn
mailing list