Poisson criterion in RandomForestRegressor

Describe the workflow you want to enable

I want to officially use the Poisson splitting criterion in RandomForestRegressor.

Describe your proposed solution

#17386 implemented the poisson splitting criterion for DecisionTreeRegressor and ExtraTreeRegressor. This also enabled—somewhat silently—to do:

import numpy as np
from sklearn.ensemble import RandomForestRegressor
y = [0, 1, 2]
X = np.arange(6).reshape(3, 2)
rf = RandomForestRegressor(criterion="poisson")
rf.fit(X, y)

Note: The same is true for ensemble.ExtraTreesRegressor.

Tasks:

• Add the poisson splitting criterion to the docstring of RandomForestRegressor.
• Add input validation (non-negative y) to RandomForestRegressor.
• Expand the tests for RandomForestRegressor.

[pk1130]

Hey @lorentzenchr! Apologies for the delay in getting around to this. Opened a pull request like you’d suggested. Done everything other than expanding the test suite. Let me know what you think and whether you’d be able to review. Thanks!

[lorentzenchr]

@pk1130 Welcome. I hope your motivation keeps on going. You’ll find general advice in the developer’s guide for contributing. The class RandomForestRegressor lives in sklearn/ensemble/_forest.py. Input validation should happen in def fit(..) which is inherited from class BaseForest. So you need to adapt the code there. I would start with adding the poisson criterion to the docstring of RandomForestRegressor. Let me know if you need more guidance.