Official implementation of
Causal Temporal Regime Structure Learning (AISTATS 2025)
Abdellah Rahmani, Pascal Frossard

CASTOR learns number of regimes, their indices and their corresponding causal graphs from multivariate time series (MTS) without any prior knowledge. It maximizes data log-likelihood via an Expectation–Maximization (EM) procedure, alternating between:

• E-step: regime learning
• M-step: learn regime-specific causal structures

Under Gaussian SEMs with equal error variances, both regime indices and their DAGs are identifiable up to label permutation.

• CASTOR.py — core EM algorithm for regime discovery + causal graphs
• CASTOR_nonlinear.py — class that builds the NNs with training procedure used in our nonlinear settings
• causal_structure.py — base class for structure models, which are an extension of networkx.DiGraph
• data_generation.py — synthetic SEM generators
• dynotears.py — baseline implementation used in our linear setting
• linear_utils.py — helpers for linear SEMs
• MLP_lag.py — MLP-based lagged predictor (nonlinear), builds the Locally Connected layers
• time_varying_weights.py — utilities for time-varying regimes
• utils.py — shared helpers
• CASTOR_tutorial.ipynb — end-to-end tutorial

# Clone
git clone https://github.com/arahmani/CASTOR.git
cd CASTOR

# (Recommended) create a fresh environment
python -m venv .venv && source .venv/bin/activate   # Windows: .venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

If you find this code useful, please cite the following paper:

@inproceedings{rahmanicausal,
  title={Causal Temporal Regime Structure Learning},
  author={Rahmani, Abdellah and Frossard, Pascal},
  booktitle={The 28th International Conference on Artificial Intelligence and Statistics}
}