MMD-SCD: MMD-based Sequential Change Detection with Reduced False Alarm Rate

Algorithms based on Maximum Mean Discrepancy to detect changes in time series.

Overview

This repository contains the implementation of the HKCUSUM algorithm and experiments described in the paper MMD-based Sequential Change Detection with Reduced False Alarm Rate, Thanh Lam Dang, Christophe Dousson, Sandrine Vaton, Thierry Chonavel. Citation incoming.

Purposes

• Through the package online_cd provide the implementation of different CUSUM-based algorithms for sequential change detection in time series.
• Present a performance comparison between the algorithm HKCUSUM proposed in the article and other CUSUM algorithms.
• Allow the simulation results presented in the paper to be reproducible for further research.

Instructions

Source Code Organisation

The project is organised as follows:

Directory	Contents
simulation/	Notebooks used to run experiments on synthetic data
src/main.py	Example that shows the usage of online-cd package on synthetic data
src/online_cd	Package for online change detection in time series
src/online_cd/detectors	Detectors that verify whether an observation is abnormal
src/online_cd/models	Models that compute anomaly scores for detectors
src/online_cd/processors	Data preprocessing before running detectors
src/online_cd/utils	Utils classes for different tasks

Dependencies

The online_cd package is written in Python3. The simulation notebooks are written in Julia. To build the online_cd package or reproduce experiments on real-world data, please install the Python modules required by running pip install -r requirements.txt inside the folder.

Usage

A complete detection process (data preparation, model initialisation, execution, analysis of results) can then be executed from a command line using the main.py executable. To reproduce the simulation results, specifically Table I, Figure 1 and Table II, run the notebooks in the simulation folder.