def test_non_transformer(dataset_with_single_grouping):
X, y, _, _, grouper = dataset_with_single_grouping
# This is not a transformer
trf = LinearRegression()
transformer = GroupedTransformer(trf, groups=grouper)
with pytest.raises(ValueError):
transformer.fit(X, y)
def test_exception_in_group(multiple_obs_fitter):
X = np.array([
[1, 2],
[1, 0],
[2, 1],
])
# Only works on groups greater than 1, so will raise an error in group 2
transformer = GroupedTransformer(multiple_obs_fitter,
groups=0,
use_global_model=False)
with pytest.raises(ValueError) as e:
transformer.fit(X)
assert "group 2" in str(e)
def test_missing_groups_transform_noglobal(dataset_with_single_grouping,
scaling_range):
X, y, groups, X_with_groups, grouper = dataset_with_single_grouping
trf = MinMaxScaler(scaling_range)
transformer = GroupedTransformer(trf,
groups=grouper,
use_global_model=False)
transformer.fit(X_with_groups, y)
# Array with 2 rows, first column a new group. Remaining top are out of range so should be the range
X_test = np.concatenate([
np.array([[3], [3]]),
np.stack([X.min(axis=0) - 1, X.max(axis=0) + 1], axis=0)
],
axis=1)
with pytest.raises(ValueError):
transformer.transform(X_test)
def test_missing_groups_transform_global(dataset_with_single_grouping,
scaling_range):
X, y, groups, X_with_groups, grouper = dataset_with_single_grouping
trf = MinMaxScaler(scaling_range)
transformer = GroupedTransformer(trf, groups=grouper)
transformer.fit(X_with_groups, y)
# Array with 2 rows, first column a new group. Remaining top are out of range so should be the range
X_test = np.concatenate([
np.array([[3], [3]]),
np.stack([X.min(axis=0), X.max(axis=0)], axis=0)
],
axis=1)
transformed = transformer.transform(X_test)
# Top row should all be equal to the small value of the range, bottom the other
assert np.allclose(transformed[0, :], scaling_range[0])
assert np.allclose(transformed[1, :], scaling_range[1])
def test_all_groups_scaled(dataset_with_single_grouping, scaling_range):
X, y, groups, X_with_groups, grouper = dataset_with_single_grouping
trf = MinMaxScaler(scaling_range)
transformer = GroupedTransformer(trf, groups=grouper)
transformed = transformer.fit(X_with_groups, y).transform(X_with_groups)
df_with_groups = pd.concat(
[pd.Series(groups.flatten(), name="G"),
pd.DataFrame(transformed)],
axis=1)
assert np.allclose(df_with_groups.groupby("G").min(), scaling_range[0])
assert np.allclose(df_with_groups.groupby("G").max(), scaling_range[1])
def test_group_correlation_minmaxscaler(dataset_with_single_grouping,
scaling_range):
X, y, groups, X_with_groups, grouper = dataset_with_single_grouping
trf = MinMaxScaler(scaling_range)
transformer = GroupedTransformer(trf, groups=grouper)
transformed = transformer.fit(X_with_groups, y).transform(X_with_groups)
# For each column, check that all grouped correlations are 1 (because MinMaxScaler scales linear)
for col in range(X.shape[1]):
assert (pd.concat([
pd.Series(groups.flatten(), name="group"),
pd.Series(X[:, col], name="original"),
pd.Series(transformed[:, col], name="transformed"),
],
axis=1).groupby("group").corr().pipe(np.allclose, 1))
def test_multiple_grouping_columns(dataset_with_multiple_grouping,
scaling_range):
X, y, groups, X_with_groups, grouper = dataset_with_multiple_grouping
trf = MinMaxScaler(scaling_range)
transformer = GroupedTransformer(trf, groups=grouper)
transformed = transformer.fit(X_with_groups, y).transform(X_with_groups)
df_with_groups = pd.concat(
[pd.DataFrame(groups, columns=["A", "B"]),
pd.DataFrame(transformed)],
axis=1)
assert np.allclose(
df_with_groups.groupby(["A", "B"]).min(), scaling_range[0])
# If a group has a single element, it defaults to min, so check wether all maxes are one of the bounds
maxes = df_with_groups.groupby(["A", "B"]).max()
assert np.all(
np.isclose(maxes, scaling_range[1])
| np.isclose(maxes, scaling_range[0])
# We have at least some groups larger than 1, so there we should find the max of the range
) and np.any(np.isclose(maxes, scaling_range[1]))