Review: Anomaly detection using scikit-learn algorithms

[detlefarend]

Description/Motivation
The package MLPro-Int-scikit-learn integrates anomaly detectors from the scikit-learn project into MLPro-OA-Streams. Version 0.3.1 generalizes the wrapper provided to a broader range of algorithms and extends the list of howtos for three selected algorithms IF, OCSVM, LOF.

The problem is, that the current howtos partially show odd results. It is currently unclear whether this can be solved by tuning the parameters of the used algorithms.

Check list

• 1. Recap of current howtos
  • 1.1. Generalization/consolidation of current howtos
    All howtos related to a particular algorithm shall be replaced by one
• 2. Recap of current state of outlier/novelty detection in scikit-learn
  • 2.1. Question: What is the difference between novelty and outlier detection in scikit-learn, and how does this affect the models in MLPro-OA-Streams?
  • 2.2. Recap: suitable algorithms and their parameters
    • IF - Isolation Forest
    • OCSVM - One Class State Vector Machine
    • LOF - Local Outlier Factors
    • Further ones?
  • 2.3. Univariate and multivariate outlier/novelty detection
    • Do available algorithms support multivariate anomaly detection?
• 3. Extensions of the wrapper for anomaly detection
  • 3.1. Support of renormalization
• 4. RTD - Updates
  • 4.1. Updates/extensions of the howto's descriptions according to the changes made in 1.1. In particular, animated GIFs shall showcase the integration of scikit-learn algorithms into MLPro
  • 4.2. Extension of the Homepage by a nice animated GIF from one of the howtos (appetizer)

Cross references
Online docu scikit-learn, 2.7. Novelty and Outlier Detection