Fork of: SUOD: Accelerating Large-scare Unsupervised Heterogeneous Outlier Detection

NEW: Now available in PyPi!

pip install sel-suod

Please refer to the original package for more information about the base functionalities. This fork forces SUOD to use pre-selected axis-parallel subspaces, such as those obtained after Feature Bagging or Feature Selection. These subspaces must be declared as a np.array, and can take any structure such that the operation X[:, subspace] yields the desired projected dataset. It uses the same class declaration as base SUOD, only adding a new variable: subspaces, and changing the class name to sel_SUOD. This fork additionally contains a number of QOL additions, like:

• During initialization, if base_estimators is an array of length 1, it will sklearn.clone() the estimator once per each subspace.
• During initialization, it will automatically check whether the number of detectors and estimators coincide.
• It will, by default, not run approximation on any method unless the global flag for approximation is manually turned to true.

There should be no conflict between SUOD and sel_SUOD. Take a look at the following code for a practical example:

base_estimators = [LOF()] #The class sel_SUOD automatically initizializes itself with subspaces.shape[0] clones of this array if len < 2.

#Creating exemplary subspaces
subspaces = [True]*20
subspaces.append(False)
subspaces = np.array([subspaces, subspaces])
subspaces[1][4] = False

model = sel_SUOD(base_estimators=base_estimators, subspaces=subspaces,
                 n_jobs=6, bps_flag=True,
                 contamination=contamination, approx_flag_global=True)
model.fit(X_train)  # fit all models with X
predicted_scores = model.decision_function(X_test)  # predict scores