MultivariateAnomalies.jl

A julia package for detecting multivariate anomalies.

Keywords: Novelty detection, Anomaly Detection, Outlier Detection, Statistical Process Control, Process Monitoring

Please cite this package as: Flach, M., Gans, F., Brenning, A., Denzler, J., Reichstein, M., Rodner, E., Bathiany, S., Bodesheim, P., Guanche, Y., Sippel, S., and Mahecha, M. D. (2017): Multivariate anomaly detection for Earth observations: a comparison of algorithms and feature extraction techniques, Earth Syst. Dynam., 8, 677-696, doi:10.5194/esd-8-677-2017.

Requirements

Julia >= 1.0
Julia packages Distances, Combinatorics, LinearAlgebra, and LIBSVM.

Installation

add the package: ]add MultivariateAnomalies

Package Features

Detect anomalies in your data with easy to use high level functions or individual anomaly detection algorithms
Feature Extraction: Preprocess your data by extracting relevant features
Similarities and Dissimilarities: Compute distance matrices, kernel matrices and k-nearest neighbor objects.
Postprocessing: Postprocess your anomaly scores, by computing their quantiles or combinations of several algorithms (ensembles).
AUC: Compute the area under the curve as external evaluation metric of your scores.
Online Algorithms: Algorithms tuned for little memory allocation.

Using the Package

For a quick start it might be useful to start with the high level functions for detecting anomalies. They can be used in highly automized way.

Input Data

MultivariateAnomalies.jl assumes that observations/samples/time steps are stored along the first dimension of the data array (rows of a matrix) with the number of observations T = size(data, 1). Variables/attributes are stored along the last dimension N of the data array (along the columns of a matrix) with the number of variables VAR = size(data, N). The implemented anomaly detection algorithms return anomaly scores indicating which observation(s) of the data are anomalous.

Authors

<img align="right" src="img/MPG_Minerva.png" alt="Minerva" width="75"/> The package was implemented by Milan Flach and Fabian Gans, Max Planck Institute for Biogeochemistry, Department Biogeochemical Integration, Research group for Empirical Inference of the Earth System, Jena.