gbelouze/mva-time-series

Code for the project of the 2021 MVA course 'ML for time series'

Automated Time-Series Anomaly Detection

Code for the 2021 MVA course 'ML for time series'

Installation

To install, run

$ pip install -e .

To install in a venv (in particular if the default pip is too old):

python3.9 -m venv ./venv/
source ./venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -e .

Codebase description

The anomaly module is implemented in src/anomaly:

• base/ specifies abstract modules which will be implemented in the following modules
• tmm/ is the time-series modelisation module, which implements models (ARIMA, polynomial, naive...) to fit the data
• adm/ contains the anomaly detection module which computes the anomalies from the residuals of the data
  • Naive.py is a simple threshold over the residuals
  • Ksigma.py is an adaptive threshold
• utils/ contains utility functions
  • statsutils.py implements the functions to compute the features from the time-series
  • modelselect_utils.py automates the computation of the features and of the scores of a model on a benchmark dataset
• io.py provides convenience functions to read the benchmark datasets
• tests/ contains scripts to check if the code is working correctly

The data/ folder contains the datasets provided by the author (we give the datasets here rather than a link to download them, as downloading them requires approval from Yahoo which takes several days).

The notebooks/ folder containes a Jupyter Notebook to run the experiments. Computing the anomaly detection on the datasets is fairly long (~15min per dataset) so we have saved some results in notebooks/saved_data/ which can be used directly in the notebook.