Various spatio-temporal clustering methods have been proposed to detect groups of jointly moving objects in space and time. However, such spatio-temporal clustering methods are rarely compared against each other to evaluate their performance in discovering moving clusters. Hence, in this work, we present a spatio-temporal clustering benchmark for the field of collective animal behavior. Our reproducible benchmark proposes synthetic datasets with ground truth and scalable implementations of spatio-temporal clustering methods. The benchmark reveals that temporal extensions of standard clustering algorithms are inherently useful for the scalable detection of moving clusters in collective animal behavior.
This repository implements various spatio-temporal clustering algorithms and showcases their ability to operate on large amounts of data:
- Built on top of
sklearn - Hyperparameter optimization by means of grid search
- Execution time control
- clustering.ipynb: Performs hyperparameter optimization for clustering algorithms and logs clustering results
- st_clustering.py: Implements ST DBSCAN, ST Agglomerative clustering, ST KMeans, ST OPTICS, ST Spectral Clustering, ST Affinity Propagation, ST BIRCH, ST HDBSCAN
- test_files: Contains small example datasets for testing purposes
The synthetic datasets with ground truth are avaiable here: [Datasets]
- Install Python requirements
pip install -r requirements.txt-
Open and run the
clustering.ipynbnotebook -
Access the results in the
cluster_results.loglog file
Released under GNU General Public License v3.0. See the LICENSE file for details. The benchmark was developed at the Data Analysis and Visualization Group at the University Konstanz funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy – EXC 2117 – 422037984.