This repository contains code for the paper "Adaptive Bernstein Change Detector for High-Dimensional Data Streams" (https://arxiv.org/abs/2306.12974).
Change detection is of fundamental importance when analyzing data streams. Detecting changes both quickly and accurately enables monitoring and prediction systems to react, e.g., by issuing an alarm or by updating a learning algorithm. However, detecting changes is challenging when observations are high-dimensional.
In high-dimensional data, change detectors should not only be able to identify when changes happen, but also in which subspace they occur. Ideally, one should also quantify how severe they are. Our approach, ABCD, has these properties. ABCD learns an encoder-decoder model and monitors its accuracy over a window of adaptive size. ABCD derives a change score based on Bernstein’s inequality to detect deviations in terms of accuracy, which indicate changes.
Our experiments demonstrate that ABCD outperforms its best competitor by at least 5% and up to 20% in F1-score on average. It can also accurately estimate changes’ subspace, together with a severity measure that correlates with the ground truth.
We have exported our conda environments into a requirements.yml file (used on our linux-server) and a requirements-windows.yml file (used on windows PCs). To recreate the environments, run
conda env create -f requirements.ymlorconda env create -f requirements-windows.yml.
If you encounter any problems during installation please have a look at the packages listed in the file and install them manually.
Please add the following repositories to this project prior to executing any experiments:
- https://github.com/heymarco/StreamGenerators
- Add to this repository via
python -m pip install -e git+https://github.com/heymarco/StreamDatasets.git#egg=stream-datasets --upgrade
- https://github.com/heymarco/ChangeDetectors
- Add to this repository via
python -m pip install -e git+https://github.com/heymarco/ChangeDetectors.git#egg=change-detectors --upgrade
conda activate abcd- To recreate the experiments in the paper, run any of the files starting with
E_(for experiments)- The code will run in parallel by default on all - 1 available cores.
- To recreate the plots run any of the
mining/M_files (for experiment mining).