[WIP] GH-6769: multinomial dt yuliia

[syzonyuliia]

#6769

[wendycwong]

Please add to your CompressLeaf class accommodation for leave nodes that will contain an array of probabilities one for each of the multinomial class. Currently, it only takes one value for binary classification.

[wendycwong]

I run into NPE error with this dataset:
sdt_3EnumCols_10kRows_multinomial.csv

[wendycwong]

With this code:

import sys
sys.path.insert(1,"../../../")
import h2o
from tests import pyunit_utils
from h2o.estimators import H2ODecisionTreeEstimator

def test_dt_multinomial():
data = h2o.import_file(pyunit_utils.locate("smalldata/sdt/sdt_3EnumCols_10kRows_multinomial.csv"))
response_col = "response"
data[response_col] = data[response_col].asfactor()

predictors = ["C1", "C2", "C3"]

# train model
dt = H2ODecisionTreeEstimator(max_depth=3)
dt.train(x=predictors, y=response_col, training_frame=data)

dt.show()

if name == "main":
pyunit_utils.standalone_test(test_dt_multinomial)
else:
test_dt_multinomial()

[valenad1]

I would like to have an basic

• Java test for multinomial
• Python test for multinomial
• R test for multinomial
• Comparsion with DRF one tree multinomial

[valenad1]

Please finish those TODOs

[wendycwong]

@syzonyuliia

Thank you for your work. It looks great.

[valenad1]

@wendycwong what is the idea behind the distribution here? I see that we optimize entropy in splits.. the attribute is not used in code.