Learning Causal Models with Non-Commutative Polynomial Optimization

Welcome to the repository for causal-discovery deliverable 3.2! This codebase encompasses notebooks and test results associated with each figure in the report.

Causal Learning Process

1. Generate Data: Use the IIDSimulation function in Generate Data.py to create artificial true causal graphs and observation data.
2. Learn Structure: Uncover the causal structure beneath the observation data.
3. Visualize Comparison: Generate heat maps to compare estimated and true graphs.
4. Calculate Metrics: Assess the performance metrics.
5. Demonstrate in Heatmap: Illustrate results in a heatmap for multiple datasets.

Getting Started

1. Synthetic Data: Utilize the Ancpop_Synthetic class to generate custom datasets.
2. Real-world Data: Employ the Ancpop_Real class in Ancpop_Real.py to test real data from the ./Real data folder. Alternatively, use generated data in ./Synthetic data.
3. Experiments: Experiment scripts for proposed methods and baselines are available in Jupyter notebooks named "Methods_withdevice.ipynb".
4. Results: Access all results and figures mentioned in the paper from the ./result folder. Repeat the results using "ANCPOP_test.ipynb".

Running Notebooks

• ANCPOP_test.ipynb: Execute this notebook for all experiments on synthetic and real-world data using the ANM_NCPOP approach. Upload datasets and the notebook to Colab for seamless execution.

Installation

To run experiments locally, ensure you have the following dependencies installed:

• Python (>= 3.6, <= 3.9)
• tqdm (>= 4.48.2)
• NumPy (>= 1.19.1)
• Pandas (>= 0.22.0)
• SciPy (>= 1.7.3)
• scikit-learn (>= 0.21.1)
• Matplotlib (>= 2.1.2)
• NetworkX (>= 2.5)
• PyTorch (>= 1.9.0)
• ncpol2sdpa 1.12.2 Documentation
• MOSEK (>= 9.3) MOSEK
• gcastle (>= 1.0.3)

PIP Installation

# Execute the following commands to run the notebook directly in Colab. Ensure your MOSEK license file is in one of these locations:
#
# /content/mosek.lic   or   /root/mosek/mosek.lic
#
# inside this notebook's internal filesystem.
# Install MOSEK and ncpol2sdpa if not already installed
pip install mosek 
pip install ncpol2sdpa
pip install gcastle==1.0.3

Provenance and Licensing

The prototype is based on the code by Quan Zhou at ICL (https://github.com/Quan-Zhou/Proper-Learning-of-LDS) and Serene He at CTU (https://github.com/sereneHe/Causal-Models-Learning/) and is licensed under Apache License 2.0 (https://github.com/codiet-eu/d32/blob/main/LICENSE).