This repo contains an implementation of the Robust Random Cut Forest anomaly detection model. This model attempts to find "anomalies" by seeking out points whose structure is not consistent with the rest of the data set. The random_cut_forest folder contains the RandomCutForest algorithm while the notebooks folder contains ipython notebooks showing examples leveraging the module.

If you want to contribute to this repo simply submit a pull request!

To install the package you can do any of the following:

• Run the command pip install ...

Using a RobustRandomCutForest to classify potential anomalies in your data is simple. Assuming you already have a vector of data stored in X you would run the following:

from robust_random_cut_forest import robust_random_cut_forest
forest = robust_random_cut_forest.RobustRandomCutForest()
forest = forest.fit(X)

From there you can choose to get the normalized depths of each point within the forest by calling average_depths or have the forest label potential anomalies by calling predict:

depths = forest.decision_function(X)
labels = forest.predict(X)

The function decision_function will return an array with numbers ranging from zero to one. The lower the number the more anomalous the point appears (this is how sklearn implements scoring). By default any points that are given a score of 0.3 are labelled as anomalies. To stream new points into your forest simply call the add_point method:

# Given an array of points....
for point in points:
    forest.add_point(point)
depths = forest.decision_function(points)
labels = forest.predict(points)

All tests are written using pytest. Simply pip install pytest to be able to run tests. All tests are located under the tests folder. Any new tests are always welcome!

• For more information on Robust Random Cut Forests, see Guha et al.'s 2016 paper which can be located here.
• The original isolation forest paper can be found here.
• Isolation Forests have been implemented in sklearn

[email protected]