dynagroup

Welcome to dynagroup. This is a Python repo for the under-review paper Discovering group dynamics in synchronous time series via hierarchical recurrent switching-state models.

Installation

See INSTALL.md to setup this project's micromamba environment.

You'll have an environment with Python 3.10 called dynagroup_env_310

Verifying install

Unit tests can be run from within the activated virtual environment using

python -m pytest

Experiment reproduction

Here we provide scripts for reproducing experiments on publicly available data (FigureEight, Basketball, and MarchingBand). Experiments from the paper can be reproduced by running the scripts/notebooks below. For exact reproducibility, hyperparameters and seeds described in the paper Appendix are required.

A. To train our model (HSRDM) as well as various ablations, use:

1. FigureEight: HSRDM, rAR-HMM
2. FigureEight: HSRDM, rAR-HMM Pool.
3. FigureEight: HSRDM, rAR-HMM Concat.
4. Basketball: HSRDM, rAR-HMM, no-recurrence ablation
5. MarchingBand: HSRDM, rAR-HMM, no-recurrence ablation

B. To train baseline forecasts, use:

1. FigureEight: DSARF Ind., Pool., Concat.
2. Basketball: AgentFormer
3. Basketball: GroupNet
4. MarchingBand: DSARF