Weighting samples by passing sample_weight to SurvivalTree.fit() is not considered.
This is essential for RandomSurvivalForest to work correctly, because bootstrap samples for each tree in the ensemble are created by passing sample_weight to SurvivalTree.fit(). For instance, sample_weight=[1, 0, 2, 1] would represent a bootstrap dataset where the first and last sample appear once, the second sample is not part of the sample, and the third sample appears twice.
Code Sample to Reproduce the Bug
```python
from sksurv.datasets import load_whas500
from sksurv.preprocessing import OneHotEncoder
from sksurv.tree import SurvivalTree
X, y = load_whas500()
Xt = OneHotEncoder().fit_transform(X)
n_samples = Xt.shape[0]
weights = np.ones(n_samples, dtype=int)
weights[:11] = np.arange(11, dtype=int)
y_array = np.empty((Xt.shape[0], 2), dtype=float)
y_array[:, 0] = y["lenfol"]
y_array[:, 1] = y["fstat"].astype(float)
X_repeat = np.repeat(Xt, weights, axis=0)
y_repeat = np.repeat(y_array, weights, axis=0)
# fit on the full data to create unique_times_ and is_event_time_
t0 = SurvivalTree(random_state=2).fit(Xt, y)
# fit on dataset where samples have been copiied
t1 = SurvivalTree(random_state=2)._fit(
X_repeat, (y_repeat, t0.unique_times_, t0.is_event_time_),
check_input=False
)
# fit on dataset where sample_weight is used
t2 = SurvivalTree(random_state=2)._fit(
Xt.values, (y_array, t0.unique_times_, t0.is_event_time_),
sample_weight=weights.astype(float),
check_input=False
)
value_1 = t1.tree_.value
value_2 = t2.tree_.value
# check that both trees are identical
assert np.allclose(value_1, value_2)
```
The example fails, whereas using a DecisionTreeClassifier for illustration does result in identical trees.
```python
from sklearn.tree import DecisionTreeClassifier
t1 = DecisionTreeClassifier(random_state=2).fit(
X_repeat, y_repeat[:, 1]
)
t2 = DecisionTreeClassifier(random_state=2).fit(
Xt, y["fstat"], sample_weight=weights
)
value_1 = t1.tree_.value
value_2 = t2.tree_.value
print(np.allclose(value_1, value_2))
```
Expected Results
The trees t1 and t2 should be identical.
Actual Results
ValueError: operands could not be broadcast together with shapes (195,395,2) (187,395,2)
Describe the bug
Weighting samples by passing
sample_weighttoSurvivalTree.fit()is not considered.This is essential for
RandomSurvivalForestto work correctly, because bootstrap samples for each tree in the ensemble are created by passingsample_weighttoSurvivalTree.fit(). For instance,sample_weight=[1, 0, 2, 1]would represent a bootstrap dataset where the first and last sample appear once, the second sample is not part of the sample, and the third sample appears twice.Code Sample to Reproduce the Bug
Expected Results The trees
t1andt2should be identical.Actual Results
Versions Latest version from the master branch.