def train_detector(data, preprocessor, perc_outlier=5):
"""Train outliers detector."""
print("Initialize outlier detector.")
od = IForest(threshold=None, n_estimators=100)
print("Training on normal data.")
np.random.seed(0)
normal_batch = create_outlier_batch(data.data,
data.target,
n_samples=30000,
perc_outlier=0)
X_train = normal_batch.data.astype('float')
# y_train = normal_batch.target
od.fit(preprocessor.transform(X_train))
print("Train on threshold data.")
np.random.seed(0)
threshold_batch = create_outlier_batch(data.data,
data.target,
n_samples=1000,
perc_outlier=perc_outlier)
X_threshold = threshold_batch.data.astype('float')
# y_threshold = threshold_batch.target
od.infer_threshold(preprocessor.transform(X_threshold),
threshold_perc=100 - perc_outlier)
return od
def test_isolation_forest(iforest_params):
threshold, threshold_perc, return_instance_score = iforest_params
X, y = load_iris(return_X_y=True)
iforest = IForest(threshold)
assert iforest.threshold == threshold
assert iforest.meta == {
'name': 'IForest',
'detector_type': 'offline',
'data_type': 'tabular'
}
iforest.fit(X)
iforest.infer_threshold(X, threshold_perc=threshold_perc)
iscore = iforest.score(X)
perc_score = 100 * (iscore <
iforest.threshold).astype(int).sum() / iscore.shape[0]
assert threshold_perc + 5 > perc_score > threshold_perc - 5
od_preds = iforest.predict(X, return_instance_score=return_instance_score)
assert od_preds['meta'] == iforest.meta
assert od_preds['data']['is_outlier'].sum() == (
iscore > iforest.threshold).astype(int).sum()
if not return_instance_score:
assert od_preds['data']['instance_score'] is None