Skip to content

Latest commit

 

History

History
54 lines (43 loc) · 3.97 KB

README.md

File metadata and controls

54 lines (43 loc) · 3.97 KB

Data Generating Process to Evaluate Causal Discovery Techniques for Time Series Data

This repository captures source code and data sets for our paper "Data Generating Process to Evaluate Causal Discovery Techniques for Time Series Data" at the Causal Discovery & Causality-Inspired Machine Learning Workshop at Neural Information Processing Systems (NeurIPS) 2020. For ease, the PDF of the paper is provided in this repository, but can also be found on the workshop website here.

If you utilise our data generating process, please cite:

@article{dgp_causal_discovery_time_series,
  title={Data Generating Process to Evaluate Causal Discovery Techniques for Time Series Data},
  author={Lawrence, Andrew R. and Kaiser, Marcus and Sampaio, Rui and Sipos, Maksim},
  journal={Causal Discovery & Causality-Inspired Machine Learning Workshop at Neural Information Processing Systems},
  year={2020}
}

Usage

Setup

The code was developed and tested using Python 3.7.8. The required Python packages, and their versions, are provided in requirements.txt. After cloning the repository locally, create a virtual environment by running the following commands:

python3 -m venv <env_name>
source <env_name>/bin/activate
pip install -r <path_to_local_repo>/requirements.txt
deactivate

Example Script

A simple example script is provided in the dgp folder. Running example_script.py will generate some synthetic data, plot it, and plot the causal diagram of the underlying Structural Causal Model (SCM) from which the data was generated. Looking at the unit tests in dgp/tests provide further insight into how the code works.

Experiment Data Sets

The data sets used for the experiments in our NeurIPS 2020 workshop paper are too large to maintain on GitHub. However, they can be downloaded from here.

All the data sets reside in a single zip file. When extracted, there is a pickle file for each experiment:

  1. Causal Sufficiency: causal_sufficiency_data.pickle
  2. Non-Linear Dependencies: linear_data.pickle
  3. Instantaneous Effects: instantaneous_effects_data.pickle
  4. IID Data: iid_data.pickle
  5. Non-Gaussian Noise: increasing_non_gaussian_noise_data.pickle

The following code snippet demonstrates how to load the causal sufficiency data:

import dill as pickle

filename = '<path_to_file>/causal_sufficiency_data.pickle'

with open(filename, 'rb') as f:
    data = pickle.load(f)

The data variable will be List[List[Tuple[pandas.DataFrame, networkx.DiGraph]]], where the outer list captures each unique configuration under test, which are described further below. Therefore, the length of the outer list is different for each experiment as a different number of configurations were used for each experiment. The inner list always has 200 elements as we tested the causal discovery methods on data produced from 200 randomly generated SCMs for each configuration. Please see the paper for more information.

  1. Causal Sufficiency: 13 configurations, where number of latent variables is [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 15, 20] and the number of feature variables are kept constant at 10.
  2. Non-Linear Dependencies: 11 configurations where likelihood of linear functions vs. other functional forms goes from 100% to 0% in 10% increments.
  3. Instantaneous Effects: 2 configurations where the minimum possible lag is [1, 0].
  4. IID Data: For IID data, min_lag = max_lag = 0. There are 6 configurations, where the number of samples is [50, 100, 200, 500, 1000, 2500].
  5. Non-Gaussian Noise: 7 configurations where likelihood of Gaussian noise vs. other possible noise distributions goes from 100% to 10% in 15% increments.