Returning the Favour: When Regression Benefits from Probabilistic Causal Knowledge

Getting started

Running code for Simulation Example

• Run from root directory

$ python run_mvn_experiment.py --cfg=config/runs/mvn_experiment.yaml --o=path/to/output/directory

Running code for Aerosol Radiative Forcing Example

• Generate dataset (takes some time)

$ python generate_FaIR_data.py --cfg=config/generation/generate_FaIR.yaml --o=data/FaIR/ --val

• Then run from root directory

$ python run_FaIR_experiment.py --cfg=config/runs/FaIR_experiment.yaml --o=path/to/output/directory

Reproducing paper results

Simulation Example

• Run experiment with multiple initialisation seeds

$ source ./repro/repro_mvn_experiment_multi_seeds.sh

• Run ablation study on number of training samples

$ source ./repro/repro_mvn_experiment_ntrain.sh

• Run ablation study on number semi-supervised samples

$ source ./repro/repro_mvn_experiment_semiprop.sh

• Run ablation study on number of dimensionality of X2

$ source ./repro/repro_mvn_experiment_d_X2.sh

• Run experiment for random forest model

Go to notebooks/mvn-random-forest-models.ipynb

• Visualise scores and generate plots

Go to notebooks/mvn-experiments-score-analysis.ipynb

Aerosol Radiative Forcing Example

• Run experiment with multiple initialisation seeds

$ source ./repro/repro_FaIR_experiment_multi_seeds.sh

• Run experiment for random forest model

Go to notebooks/FaIR-random-forest-models.ipynb

• Visualise scores and generate table

Go to notebooks/FaIR-experiments-score-analysis.ipynb

Installation

Code implemented in Python 3.8.0

Setting up environment

Create and activate environment (with pyenv here)

$ pyenv virtualenv 3.8.0 venv
$ pyenv activate venv
$ (venv)

Install dependencies

$ (venv) pip install -r requirements.txt

References

@inproceedings{BouFawSej2023,
  title={{Returning The Favour: When Regression Benefits From Probabilistic Causal Knowledge}},
  author={Bouabid, Shahine and Fawkes, Jake and Sejdinovic, Dino},
  year={2023},
  journal={International Conference on Machine Learning (ICML)}
}