def train():
dataset = get_data(1000, 10, 100)
contamination = 0.01
with mlflow.start_run():
base_estimators = [
LOF(n_neighbors=5, contamination=contamination),
LOF(n_neighbors=15, contamination=contamination),
LOF(n_neighbors=25, contamination=contamination),
PCA(contamination=contamination),
KNN(n_neighbors=5, contamination=contamination),
KNN(n_neighbors=15, contamination=contamination),
KNN(n_neighbors=25, contamination=contamination)]
model = SUOD(base_estimators=base_estimators, n_jobs=6,
rp_flag_global=True,
bps_flag=True,
approx_flag_global=False,
contamination=contamination)
model.fit(dataset)
model.approximate(dataset)
predicted_labels = model.predict(dataset)
voted_labels = vote(predicted_labels)
true_labels = [0]*1000 + [1]*10
auc_score = roc_auc_score(voted_labels, true_labels)
print("The resulted area under the ROC curve score is {}".format(auc_score))
mlflow.log_metric("auc_score", auc_score)
mlflow.sklearn.log_model(model, "anomaly_model", conda_env="conda.yaml")
class ensemble(abstract_occ_model):
"""
"""
def __init__(self, nu=0.1):
self.base_estimators = [
#OCSVM(contamination=nu),
KNN(n_neighbors=100, contamination=nu),
KNN(n_neighbors=25, contamination=nu),
KNN(n_neighbors=5, contamination=nu),
IForest(contamination=nu)
]
self.model = SUOD(base_estimators=self.base_estimators,
rp_flag_global=False,
bps_flag=True,
approx_flag_global=False)
self.scores = None
def fit(self, X):
self.model.fit(X)
self.model.approximate(X)
def predict(self, X):
self.scores = self.compute_score(X)
return np.where(self.scores >= 0.5, 1,
np.where(self.scores < 0.5, -1, self.scores))
def score_samples(self, X):
if type(self.scores) != np.ndarray:
self.scores = self.compute_score(X)
return self.scores
else:
return self.scores
def compute_score(self, X):
mean_prob = np.mean(self.model.predict_proba(X), axis=1)
return mean_prob
##########################################################################
model = SUOD(base_estimators=base_estimators,
rp_flag_global=True,
approx_clf=approx_clf,
n_jobs=n_jobs,
bps_flag=True,
contamination=contamination,
approx_flag_global=True)
start = time.time()
model.fit(X_train) # fit all models with X
print('Fit time:', time.time() - start)
print()
start = time.time()
model.approximate(X_train) # conduct model approximation if it is enabled
print('Approximation time:', time.time() - start)
print()
start = time.time()
predicted_labels = model.predict(X_test) # predict labels
print('Predict time:', time.time() - start)
print()
start = time.time()
predicted_scores = model.decision_function(X_test) # predict scores
print('Decision Function time:', time.time() - start)
print()
##########################################################################
# compare with no projection, no bps, and no approximation
X = np.append(arr=X, values=features_pca, axis=0)
X_num = X.shape[0]
base_estimators = [LOF(), IForest(), OCSVM(kernel="rbf", gamma=0.001)]
model = SUOD(
base_estimators=base_estimators,
n_jobs=2, # number of workers(if -1 it use full core)
rp_flag_global=True, # global flag for random projection
bps_flag=True, # global flag for balanced parallel scheduling
approx_flag_global=False, # global flag for model approximation
contamination=0.2)
# X_train, X_test = train_test_split(X, test_size=0, random_state=123)\
model.fit(X)
model.approximate(X)
predicted_labels = model.predict(X)
sum_labels = np.sum(predicted_labels, axis=1) / 3
sum_labels = np.where(sum_labels >= 0.5, -1,
1) # -1 abnormal, 1 normal
result_label = np.average(sum_labels)
result_label = result_label.tolist()
# Add outliers
fig = plt.figure()
colors = np.array(['r', 'b'])
ax = fig.add_subplot(111, projection='3d')
ax.scatter(X[:, 0],
X[:, 1],
X[:, 2],
class TestBASE(unittest.TestCase):
def setUp(self):
self.n_train = 1000
self.n_test = 500
self.contamination = 0.1
self.roc_floor = 0.6
self.random_state = 42
self.X_train, self.y_train, self.X_test, self.y_test = generate_data(
n_train=self.n_train, n_test=self.n_test,
contamination=self.contamination, random_state=self.random_state)
self.base_estimators = [
LOF(n_neighbors=5, contamination=self.contamination),
LOF(n_neighbors=15, contamination=self.contamination),
LOF(n_neighbors=25, contamination=self.contamination),
LOF(n_neighbors=35, contamination=self.contamination),
LOF(n_neighbors=45, contamination=self.contamination),
HBOS(contamination=self.contamination),
PCA(contamination=self.contamination),
LSCP(detector_list=[
LOF(n_neighbors=5, contamination=self.contamination),
LOF(n_neighbors=15, contamination=self.contamination)],
random_state=self.random_state)
]
this_directory = os.path.abspath(os.path.dirname(__file__))
self.cost_forecast_loc_fit_ = os.path.join(this_directory,
'bps_train.joblib')
self.cost_forecast_loc_pred_ = os.path.join(this_directory,
'bps_prediction.joblib')
self.model = SUOD(base_estimators=self.base_estimators, n_jobs=2,
rp_flag_global=True, bps_flag=True,
contamination=self.contamination,
approx_flag_global=True,
cost_forecast_loc_fit=self.cost_forecast_loc_fit_,
cost_forecast_loc_pred=self.cost_forecast_loc_pred_)
def test_initialization(self):
self.model.get_params()
self.model.set_params(**{'n_jobs': 4})
def test_fit(self):
"""
Test base class initialization
:return:
"""
self.model.fit(self.X_train)
def test_approximate(self):
self.model.fit(self.X_train)
self.model.approximate(self.X_train)
def test_predict(self):
self.model.fit(self.X_train)
self.model.approximate(self.X_train)
self.model.predict(self.X_test)
def test_decision_function(self):
self.model.fit(self.X_train)
self.model.approximate(self.X_train)
self.model.decision_function(self.X_test)
class SUOD(BaseDetector):
# noinspection PyPep8
"""SUOD (Scalable Unsupervised Outlier Detection) is an acceleration
framework for large scale unsupervised outlier detector training and
prediction. See :cite:`zhao2021suod` for details.
Parameters
----------
base_estimators : list, length must be greater than 1
A list of base estimators. Certain methods must be present, e.g.,
`fit` and `predict`.
combination : str, optional (default='average')
Decide how to aggregate the results from multiple models:
- "average" : average the results from all base detectors
- "maximization" : output the max value across all base detectors
contamination : float in (0., 0.5), optional (default=0.1)
The amount of contamination of the data set,
i.e. the proportion of outliers in the data set. Used when fitting to
define the threshold on the decision function.
n_jobs : optional (default=1)
The number of jobs to run in parallel for both `fit` and
`predict`. If -1, then the number of jobs is set to the
the number of jobs that can actually run in parallel.
rp_clf_list : list, optional (default=None)
The list of outlier detection models to use random projection. The
detector name should be consistent with PyOD.
rp_ng_clf_list : list, optional (default=None)
The list of outlier detection models NOT to use random projection. The
detector name should be consistent with PyOD.
rp_flag_global : bool, optional (default=True)
If set to False, random projection is turned off for all base models.
target_dim_frac : float in (0., 1), optional (default=0.5)
The target compression ratio.
jl_method : string, optional (default = 'basic')
The JL projection method:
- "basic": each component of the transformation matrix is taken at
random in N(0,1).
- "discrete", each component of the transformation matrix is taken at
random in {-1,1}.
- "circulant": the first row of the transformation matrix is taken at
random in N(0,1), and each row is obtained from the previous one
by a one-left shift.
- "toeplitz": the first row and column of the transformation matrix
is taken at random in N(0,1), and each diagonal has a constant value
taken from these first vector.
bps_flag : bool, optional (default=True)
If set to False, balanced parallel scheduling is turned off.
approx_clf_list : list, optional (default=None)
The list of outlier detection models to use pseudo-supervised
approximation. The detector name should be consistent with PyOD.
approx_ng_clf_list : list, optional (default=None)
The list of outlier detection models NOT to use pseudo-supervised
approximation. The detector name should be consistent with PyOD.
approx_flag_global : bool, optional (default=True)
If set to False, pseudo-supervised approximation is turned off.
approx_clf : object, optional (default: sklearn RandomForestRegressor)
The supervised model used to approximate unsupervised models.
cost_forecast_loc_fit : str, optional
The location of the pretrained cost prediction forecast for training.
cost_forecast_loc_pred : str, optional
The location of the pretrained cost prediction forecast for prediction.
verbose : int, optional (default=0)
Controls the verbosity of the building process.
Attributes
----------
decision_scores_ : numpy array of shape (n_samples,)
The outlier scores of the training data.
The higher, the more abnormal. Outliers tend to have higher
scores. This value is available once the detector is
fitted.
threshold_ : float
The threshold is based on ``contamination``. It is the
``n_samples * contamination`` most abnormal samples in
``decision_scores_``. The threshold is calculated for generating
binary outlier labels.
labels_ : int, either 0 or 1
The binary labels of the training data. 0 stands for inliers
and 1 for outliers/anomalies. It is generated by applying
``threshold_`` on ``decision_scores_``.
"""
def __init__(self,
base_estimators=None,
contamination=0.1,
combination='average',
n_jobs=None,
rp_clf_list=None,
rp_ng_clf_list=None,
rp_flag_global=True,
target_dim_frac=0.5,
jl_method='basic',
bps_flag=True,
approx_clf_list=None,
approx_ng_clf_list=None,
approx_flag_global=True,
approx_clf=None,
cost_forecast_loc_fit=None,
cost_forecast_loc_pred=None,
verbose=False):
super(SUOD, self).__init__(contamination=contamination)
self.base_estimators = base_estimators
self.contamination = contamination
self.combination = combination
self.n_jobs = n_jobs
self.rp_clf_list = rp_clf_list
self.rp_ng_clf_list = rp_ng_clf_list
self.rp_flag_global = rp_flag_global
self.target_dim_frac = target_dim_frac
self.jl_method = jl_method
self.bps_flag = bps_flag
self.approx_clf_list = approx_clf_list
self.approx_ng_clf_list = approx_ng_clf_list
self.approx_flag_global = approx_flag_global
self.approx_clf = approx_clf
self.cost_forecast_loc_fit = cost_forecast_loc_fit
self.cost_forecast_loc_pred = cost_forecast_loc_pred
self.verbose = verbose
# by default we will provide a group of performing models
if self.base_estimators is None:
self.base_estimators = [
LOF(n_neighbors=15),
LOF(n_neighbors=20),
HBOS(n_bins=10),
HBOS(n_bins=20),
COPOD(),
IForest(n_estimators=50),
IForest(n_estimators=100),
IForest(n_estimators=150)
]
self.n_estimators = len(self.base_estimators)
# pass in the arguments for SUOD model
self.model_ = SUOD_model(
base_estimators=self.base_estimators,
contamination=self.contamination,
n_jobs=self.n_jobs,
rp_clf_list=self.rp_clf_list,
rp_ng_clf_list=self.rp_ng_clf_list,
rp_flag_global=self.rp_flag_global,
target_dim_frac=self.target_dim_frac,
jl_method=self.jl_method,
approx_clf_list=self.approx_clf_list,
approx_ng_clf_list=self.approx_ng_clf_list,
approx_flag_global=self.approx_flag_global,
approx_clf=self.approx_clf,
bps_flag=self.bps_flag,
cost_forecast_loc_fit=self.cost_forecast_loc_fit,
cost_forecast_loc_pred=self.cost_forecast_loc_pred,
verbose=self.verbose,
)
def fit(self, X, y=None):
"""Fit detector. y is ignored in unsupervised methods.
Parameters
----------
X : numpy array of shape (n_samples, n_features)
The input samples.
y : Ignored
Not used, present for API consistency by convention.
Returns
-------
self : object
Fitted estimator.
"""
# validate inputs X and y (optional)
X = check_array(X)
n_samples, n_features = X.shape[0], X.shape[1]
self._set_n_classes(y)
# fit the model and then approximate it
self.model_.fit(X)
self.model_.approximate(X)
# get the decision scores from each base estimators
decision_score_mat = np.zeros([n_samples, self.n_estimators])
for i in range(self.n_estimators):
decision_score_mat[:, i] = self.model_.base_estimators[
i].decision_scores_
# the scores must be standardized before combination
decision_score_mat, self.score_scalar_ = standardizer(
decision_score_mat, keep_scalar=True)
# todo: may support other combination
if self.combination == 'average':
decision_score = average(decision_score_mat)
else:
decision_score = maximization(decision_score_mat)
assert (len(decision_score) == n_samples)
self.decision_scores_ = decision_score.ravel()
self._process_decision_scores()
return self
def decision_function(self, X):
"""Predict raw anomaly score of X using the fitted detectors.
The anomaly score of an input sample is computed based on different
detector algorithms. For consistency, outliers are assigned with
larger anomaly scores.
Parameters
----------
X : numpy array of shape (n_samples, n_features)
The training input samples. Sparse matrices are accepted only
if they are supported by the base estimator.
Returns
-------
anomaly_scores : numpy array of shape (n_samples,)
The anomaly score of the input samples.
"""
check_is_fitted(
self, ['model_', 'decision_scores_', 'threshold_', 'labels_'])
X = check_array(X)
# initialize the output score
predicted_scores = self.model_.decision_function(X)
# standardize the score and combine
predicted_scores = self.score_scalar_.transform(predicted_scores)
# todo: may support other combination
if self.combination == 'average':
decision_score = average(predicted_scores)
else:
decision_score = maximization(predicted_scores)
assert (len(decision_score) == X.shape[0])
return decision_score.ravel()
class TestModelSaveLoad(unittest.TestCase):
def setUp(self):
self.n_train = 1000
self.n_test = 500
self.contamination = 0.1
self.roc_floor = 0.6
self.random_state = 42
self.X_train, self.y_train, self.X_test, self.y_test = generate_data(
n_train=self.n_train, n_test=self.n_test,
contamination=self.contamination, random_state=self.random_state)
self.base_estimators = [
LOF(n_neighbors=5, contamination=self.contamination),
LOF(n_neighbors=15, contamination=self.contamination),
LOF(n_neighbors=25, contamination=self.contamination),
LOF(n_neighbors=35, contamination=self.contamination),
LOF(n_neighbors=45, contamination=self.contamination),
HBOS(contamination=self.contamination),
PCA(contamination=self.contamination),
LSCP(detector_list=[
LOF(n_neighbors=5, contamination=self.contamination),
LOF(n_neighbors=15, contamination=self.contamination)],
random_state=self.random_state)
]
this_directory = os.path.abspath(os.path.dirname(__file__))
self.cost_forecast_loc_fit_ = os.path.join(this_directory,
'bps_train.joblib')
self.cost_forecast_loc_pred_ = os.path.join(this_directory,
'bps_prediction.joblib')
self.model = SUOD(base_estimators=self.base_estimators, n_jobs=2,
rp_flag_global=True, bps_flag=True,
contamination=self.contamination,
approx_flag_global=True,
cost_forecast_loc_fit=self.cost_forecast_loc_fit_,
cost_forecast_loc_pred=self.cost_forecast_loc_pred_,
verbose=True)
def test_save(self):
self.model.fit(self.X_train) # fit all models with X
self.model.approximate(
self.X_train) # conduct model approximation if it is enabled
# save the model
dump(self.model, 'model.joblib')
assert (os.path.exists('model.joblib'))
os.remove('model.joblib')
def test_load(self):
self.model.fit(self.X_train) # fit all models with X
self.model.approximate(
self.X_train) # conduct model approximation if it is enabled
# save the model
dump(self.model, 'model.joblib')
model = load('model.joblib')
predicted_labels = model.predict(self.X_test) # predict labels
predicted_scores = model.decision_function(self.X_test) # predict scores
predicted_probs = model.predict_proba(self.X_test) # predict scores
assert (len(predicted_labels) != 0)
# assert (predicted_scores)
# assert (predicted_probs)
def tearDown(self):
if os.path.exists('model.joblib'):
os.remove('model.joblib')