def test_anytime(self):
# The feature weights should be as uniformly distributed as possible.
# We select old features with very little restraint -> we can mess up old distributions,
# furthermore, we do not correct that -> fulfil at least these loose constraints.
rf = RF(y, mtry=0.75, n_jobs=30, seed=2001)
rf.fit(random.randint(0, 10, (5, 20)))
weights = rf.fit(random.randint(0, 10, (5, 1)))
self.assertTrue(min(weights) > (mean(weights) - 3 * std(weights)))
self.assertTrue(max(weights) < (mean(weights) + 3 * std(weights)))
assert_almost_equal(
mean(weights[0:20]),
weights[20],
err_msg=
"The new feature should have the weight equivalent to the average weight of all the previous features"
)
weights = rf.fit(random.randint(0, 10, (5, 1)))
assert_almost_equal(
mean(weights[0:21]),
weights[21],
err_msg=
"The new feature should have the weight equivalent to the average weight of all the previous features"
)
def test_incremental_learning(self):
# Test that we can initialize the RF, add a feature, score, add features, score.
rf = RF(y, mtry=0.8, n_jobs=2, seed=2001)
rf.fit(X0)
prediction1 = rf.score(X0)
rf.fit(X1)
prediction2 = rf.score(column_stack((X0, X1)))
assert_almost_equal(
prediction1,
y,
err_msg="Feature X0 is a leaking feature - overfit on it!")
assert_almost_equal(
prediction2,
y,
err_msg="Feature X0 is a leaking feature - overfit on it!")
class RFWrapper:
"""
Online Random Forest.
"""
def __init__(self, y, X_t, y_t, n_jobs, mtry, random_state):
self.rf = RF(y, n_jobs=n_jobs, mtry=mtry, seed=random_state)
self.X_t = X_t
self.y_t = y_t
def fit(self, x):
self.rf.fit(x)
def get_auc(self):
prediction = self.rf.score(self.X_t)
fpr, tpr, thresholds = metrics.roc_curve(self.y_t,
prediction,
pos_label=1)
return metrics.auc(fpr, tpr)
def test_seed_diff(self):
## When the seeds are different, the random forest provides different results.
rf = RF(y, mtry=0.75, n_jobs=20, seed=2001)
rf.fit(X1)
weights1 = rf.fit(X2)
rf = RF(y, mtry=0.75, n_jobs=20, seed=2002)
rf.fit(X1)
weights2 = rf.fit(X2)
self.assertFalse(all(weights1 == weights2))
def test_seed(self):
# When the seeds are the same, the random forest provides identical results.
rf = RF(y, mtry=0.75, n_jobs=20, seed=2001)
rf.fit(X1)
weights1 = rf.fit(X2)
rf = RF(y, mtry=0.75, n_jobs=20, seed=2001)
rf.fit(X1)
weights2 = rf.fit(X2)
assert_almost_equal(weights1, weights2)
def test_vector(self):
rf = RF(y, mtry=0.8, n_jobs=2, seed=2001)
rf.fit(y)