ValueError: probabilities do not sum to 1

[zhao-zilong]

The bug comes from this line:

arfpy/arfpy/arf.py

Line 313 in b5bf5de

draws = np.random.choice(a=range(unique_bnds.shape[0]), p = unique_bnds['cvg'] / self.num_trees, size=n)

And I think why it happens it is because in this line of code: you intentionally set come cvg values to zero:

arfpy/arfpy/arf.py

Line 224 in b5bf5de

bnds.loc[bnds['cvg'] == 1/pred.shape[0],'cvg'] = 0

Do you have some insights on this?

[mnwright]

Can you give a reproducible example for the ValueError?

We set those to zero because we cannot estimate variances for single-observation leaves. In R, we now avoid that completely by using "class-wise min.bucket", which avoids nodes with less than 2 (or whatever the user sets for the parameter) real observations (see imbs-hl/ranger#721). But I think that is not possible with scikit-learn's RandomForestClassifier: There is min_samples_leaf but that can't be class-specific I think.