def test_mae_tree(training_data_1d):
X, y = training_data_1d
tree = DecisionTree(node_class=MedianNode, min_samples_leaf=2)
tree.fit(X, y, splitter=splitters.MAESplitter())
preds = tree.predict(X)
if isinstance(X, pd.DataFrame):
assert list(preds.index) == list(X.index)
if isinstance(y, pd.DataFrame):
assert list(preds.columns) == list(y.columns)
preds = preds[0].tolist()
assert preds == pytest.approx(
[15.1, 15.1, 15.1, 25.2, 25.2, 25.2, 9.8, 9.8, 9.8, 17.0, 17.0, 17.0],
abs=1e-1)
# 4 leaves are formed:
node_ids = tree.apply(X)
assert len(set(node_ids)) == 4
for group_id in range(4):
begin_idx = 3 * group_id
end_idx = 3 * group_id + 2
assert len(set(node_ids[begin_idx:end_idx])) == 1
#TODO: write a test where training X is a dataframe, and the predicting X has a different column order.
def test_min_samples_leaf_affects_mae_split(training_data_1d):
X, Y = training_data_1d
training_data = TrainingData(X, Y)
mae_splitter = splitters.MAESplitter()
coeffs, cutpoint, cost = mae_splitter.select_feature_to_cut(
training_data.X, training_data.Y,
len(training_data.X) / 2)
index = np.argwhere(coeffs != 0.0)[0][0]
assert training_data.X_names[index] in {'x', 0}
assert cutpoint == 5.5
def test_mae_splitter(training_data_1d):
X, Y = training_data_1d
training_data = TrainingData(X, Y)
mae_splitter = splitters.MAESplitter()
coeffs, cutpoint, cost = mae_splitter.select_feature_to_cut(
training_data.X, training_data.Y, 2)
index = np.argwhere(coeffs != 0.0)[0][0]
assert training_data.X_names[index] in {'x', 0}
# Brute force calculation reveals 8.5 results in best MAE reduction:
assert cutpoint == 8.5
import pytest
import pandas as pd
from pyboretum import (
MeanNode,
MedianNode,
splitters, # MAE and MSE splitters
DecisionTree,
)
@pytest.mark.parametrize('splitter, node_class', [
(splitters.MSESplitter(), MeanNode),
(splitters.MAESplitter(), MedianNode),
])
def test_with_one_sample_nodes(splitter, node_class, training_data_1d):
X, y = training_data_1d
tree = DecisionTree(node_class=node_class, min_samples_leaf=1)
tree.fit(X, y, splitter=splitter)
# All y should form their own leaves:
preds = tree.predict(X)
if isinstance(X, pd.DataFrame):
assert list(preds.index) == list(X.index)
if isinstance(y, pd.DataFrame):
assert list(preds.columns) == list(y.columns)
preds = preds[0].tolist()
assert preds == pytest.approx(y.tolist(), abs=1e-1)
import pytest
from pyboretum import (
splitters,
TrainingData,
)
@pytest.mark.parametrize('splitter', [
splitters.MSESplitter(),
splitters.MAESplitter(),
])
def test_min_samples_leaf_can_stop_splitting(splitter, training_data_1d):
X, Y = training_data_1d
training_data = TrainingData(X, Y)
feature, cutpoint, cost = splitter.select_feature_to_cut(
training_data.X, training_data.Y, len(training_data.X))
assert feature == None
assert cutpoint == None
assert cost == float('inf')