randomforest.py

# show that max_depth affects floating point precision of predict_proba in RandomForest

from collections import Counter
from sklearn.ensemble import RandomForestClassifier
from sklearn.datasets import make_classification

X, y = make_classification(n_samples=1000, n_features=20,
                                    n_informative=5, n_redundant=10,
                                    random_state=42)


print('with no max depth:')
clf = RandomForestClassifier(n_estimators=10)
print(Counter(clf.fit(X,y).predict_proba(X)[:,1]).most_common(10))


print('\nwith max depth=5:')
clf = RandomForestClassifier(n_estimators=10, max_depth=5)
print(Counter(clf.fit(X,y).predict_proba(X)[:,1]).most_common(10))

# output:
# with no max depth:
# [(1.0, 370), (0.0, 359), (0.9, 93), (0.1, 89), (0.2, 32), (0.8, 24), (0.3, 12), (0.7, 8), (0.6, 6), (0.4, 4)]
# 
# with max depth=5:
# [(0.035039564760949896, 48), (0.9961472577676984, 37), (0.01221724785227863, 32), (0.9961229150218367, 19), (0.013133914518945296, 18), (0.992163639003737, 17), (0.9980459919449135, 10), (0.025767534109608903, 10), (0.30845687643280967, 9), (0.01815332091163936, 8)]