def test_construct_maj_int_min_when_wrong_strategy():
class_sizes = {
0: 5, # class 0 occurs 5 times in dataset
1: 6,
3: 7,
5: 10,
8: 2000
}
y = np.array([
class_label for class_label, class_size in class_sizes.items()
for _ in range(class_size)
])
np.random.shuffle(y)
with pytest.raises(ValueError):
construct_maj_int_min(y, strategy='WRONG_STRATEGY')
def _initialize_algorithm(self, X, y):
if self.maj_int_min is None:
self.maj_int_min = construct_maj_int_min(y)
self.majority_classes = self.maj_int_min['maj']
self.intermediate_classes = self.maj_int_min['int']
self.minority_classes = self.maj_int_min['min']
self.stds, self.means = [1] * X.shape[1], [0] * X.shape[1]
if self.cost is None:
self.cost = self._estimate_cost_matrix(y)
def test_construct_maj_int_min_when_correct_and_average_strategy():
class_sizes = {
0: 5, # class 0 occurs 5 times in dataset
1: 6,
3: 7,
5: 10,
8: 2000
}
y = np.array([
class_label for class_label, class_size in class_sizes.items()
for _ in range(class_size)
])
np.random.shuffle(y)
maj_int_dict = construct_maj_int_min(y, strategy='average')
assert len(maj_int_dict['int']) == 0
assert len(maj_int_dict['min']) == 4
assert all(i in maj_int_dict['min'] for i in [0, 1, 3, 5])
assert len(maj_int_dict['maj']) == 1
assert maj_int_dict['maj'][0] == 8
def test_construct_maj_int_min_when_correct_and_median_strategy():
class_sizes = {
0: 5, # class 0 occurs 5 times in dataset
1: 6,
3: 7, # median
5: 10,
8: 12
}
y = np.array([
class_label for class_label, class_size in class_sizes.items()
for _ in range(class_size)
])
np.random.shuffle(y)
maj_int_dict = construct_maj_int_min(y, strategy='median')
assert len(maj_int_dict['int']) == 1
assert maj_int_dict['int'][0] == 3
assert len(maj_int_dict['min']) == 2
assert all(i in maj_int_dict['min'] for i in [0, 1])
assert len(maj_int_dict['maj']) == 2
assert all(i in maj_int_dict['maj'] for i in [5, 8])
def _fit_resample(self, X, y):
"""
:param X:
two dimensional numpy array (number of samples x number of features) with float numbers
:param y:
one dimensional numpy array with labels for rows in X
:return:
resampled X, resampled y
"""
if self.class_balances is None:
self.class_balances = construct_maj_int_min(y)
self.knn.fit(X)
self.X, self.y = X, y
oversampled_X, oversampled_y = self.X.copy(), self.y.copy()
quantities = Counter(self.y)
goal_quantity = int(max(list(quantities.values())))
labels = list(set(self.y))
minority_classes = self.class_balances['min']
for class_label in labels:
if minority_classes is not None and class_label not in minority_classes:
continue
chosen_minor_class_samples_to_oversample, weights = self._choose_samples(
class_label)
if len(chosen_minor_class_samples_to_oversample) == 0:
continue
oversampling_rate = int(
(goal_quantity - quantities[class_label]) * self.prop)
if oversampling_rate > 0:
if len(chosen_minor_class_samples_to_oversample) == 1:
oversampled_set = np.repeat(
chosen_minor_class_samples_to_oversample,
oversampling_rate,
axis=0)
else:
chosen_samples_features_mean = np.mean(
chosen_minor_class_samples_to_oversample, axis=0)
zero_mean_samples = chosen_minor_class_samples_to_oversample - chosen_samples_features_mean
n_components = min(zero_mean_samples.shape)
pca = PCA(n_components=n_components).fit(zero_mean_samples)
uncorrelated_samples = pca.transform(zero_mean_samples)
variables_variance = np.diag(
np.cov(uncorrelated_samples, rowvar=False))
oversampled_set = self._MDO_oversampling(
uncorrelated_samples, variables_variance,
oversampling_rate, weights)
oversampled_set = pca.inverse_transform(
oversampled_set) + chosen_samples_features_mean
oversampled_X = np.vstack((oversampled_X, oversampled_set))
oversampled_y = np.hstack(
(oversampled_y,
np.array([class_label] * oversampling_rate)))
return oversampled_X, oversampled_y