🎓 CS-673 Project (Intro to Deep Generative Models)

📜 Overview

The task of Causal Discovery is to uncover the true DAG $\mathcal{G}$ given a dataset $D$ such that $p_D ∼ \mathcal{G}$. Since the work of [ZARX18], which transformed the task of causal discovery to a continuous optimization program with acyclicity constraints, significant attention has shifted to deep learning-based causal discovery algorithms.

In this project, we select DECI (Deep End-to-end Causal Inference), a SOTA causal discovery algorithm for iid observational data by [GAF+22], a variational inference model modeling exogenous noise as a normalizing flow.

Since the literature on causal discovery algorithms for time-series data using deep learning techniques is quite limited and virginal, we opt to implement DECI on time-series data using lagged cross-sectional data. Finally, we evaluate our performance against non-deep-learning-inspired algorithms (PCMCI, PCMCI(+) etc.) on time-series synthetic data with known ground truth causal graph [LKSS20].

📚 References

[GAF+22] Tomas Geffner, Javier Antoran, Adam Foster, Wenbo Gong, Chao Ma, Emre Kiciman, Amit Sharma, Angus Lamb, Martin Kukla, Nick Pawlowski, et al. Deep end-to-end causal inference. arXiv preprint arXiv:2202.02195, 2022.
[LKSS20] Andrew R. Lawrence, Marcus Kaiser, Rui Sampaio, and Maksim Sipos. Data generating process to evaluate causal discovery techniques for time series data. Causal Discovery Causality-Inspired Machine Learning Workshop at Neural Information Processing Systems, 2020.
[VCB22] Matthew J Vowels, Necati Cihan Camgoz, and Richard Bowden. D’ya like DAGs? a survey on structure learning and causal discovery. ACM Computing Surveys, 55(4):1–36, 2022.
[ZARX18] Xun Zheng, Bryon Aragam, Pradeep K Ravikumar, and Eric P Xing. Dags with NO-TEARS: Continuous optimization for structure learning. Advances in Neural Information Processing Systems, 31, 2018.

🛠️ Setup Instructions

There are two separate environments that need to be configured to reproduce this project: CDML and DECI (causica).

🐍 Creating Virtual Environments

You may create the virtual environments with their respective requirements using the provided .yml files, using for example your Anaconda installation, on your shell as

For CDML:

 conda env create -f environment-cdml.yml

For causica:

conda env create -f environment-causica.yml

📌 Environment Details

The first environment runs on Python 3.8.19.
The second environment runs on Python 3.10.1, 🔥 PyTorch 1.13.0 and PyTorch lightning 2.2.2.

🥰 Reproducing the experiments

Run generate_dataset.ipynb to generate a CDML configuration, plot the causal graph and generate the corresponding time-lagged dataset.
Run experiments.ipynb to run a DECI model on a CDML configuration, compute the metrics and compare to the ground truth graph, as well as PCMCI.
Run RunAll.ipynb to evaluate all pre-trained DECI models on each datasetavailable at the datasets folder and compare with PCMCI.

Enjoy! 🚀