Improvement-Focused Causal Recourse: Experiments

Installation

We used python version 3.9.7. We recommend to set up a fresh virtual environment, to install the dependenciees and then the package.

Dependencies

We recommend to install torch, scikit-learn, matplotlib, seaborn, networkx, pyro, deap as well as jaxand numpyro seperately. For torch, follow the instruction on their website. For scikit-learn check out their website. For pyro, follow the instructions on their website.

The remaining dependencies can be found in the requirements.txt file. You can install them using pip install -r requirements.txt.

Installation with pip

In order to install the package and its functionality, run pip install -e icr in the superfolder of the one that contains this README file.

The Package

The package allows to apply ICR, CR and CE to problems with specified causal knowledge. An SCM can be specified and a series of experiments can be run.

How to Reproduce Results

In order to reproduce our results, run the following script for [confidence] 0.75, 0.85, 0.95 and 0.9, for [savepath] being the path were you would like to store the experiment results, and for [nr_runs] being the number of runs computed for the given configuration.

python scripts/run_experiments.py 3var-noncausal 4000 200 [confidence] 300 [savepath]/3var-nc/ [nr_runs] --NGEN 600 --POP_SIZE 300 --n_digits 1 --nr_refits 5 --predict_individualized True

python scripts/run_experiments.py 3var-causal 4000 200 [confidence] 300 [savepath]/3var-c/ [nr_runs] --NGEN 600 --POP_SIZE 300 --n_digits 1 --nr_refits 5 --predict_individualized True

python scripts/run_experiments.py 5var-skill 4000 200 [confidence] 300 [savepath]/5var-skill/ [nr_runs] --NGEN 1000 --POP_SIZE 500 --n_digits 0 --nr_refits 5 --predict_individualized True --model_type rf

python scripts/run_experiments.py 7var-covid 20000 200 [confidence] 2999 [savepath]/7var-covid/ [nr_runs] --NGEN 700 --POP_SIZE 300 --n_digits 1 --nr_refits 5 --predict_individualized True --model_type rf

The experiments can be compiled and combined into a plot using

python scripts/plots.py --savepath [savepath]

where the [savepath] is the same path used above when running the experiments. If you use folder names different from those specified above for the results, you may have to go into the plots.py script to adapt them (by adjusting the scm_names dictionary). In the [savepath] folder, for each scm you can then find two files called

• resultss.csv: summary statistics for all experiment folders in the specified result folder. mean and standard deviation for
  • eta_mean: specified desired acceptance rate
  • gamma_mean: specified desired improvement rate
  • perc_recomm_found: percent of recourse-seeking individuals for which a recommendation could be made
  • eta_obs_mean: average acceptance rate (observed)
  • gamma_obs_mean: average improvement rate (observed)
  • eta_obs_individualized_mean: average acceptance rate for the individualized post-recourse predictor (observed)
  • eta_obs_refits_batch0_mean_mean: average acceptance rate mean over all (model multiplicity) refits on batch 1 evaluated over batch 2.
  • intv-cost_mean: averge cost of the suggested interventions
  • [...]_std: the respective standard deviations
• invs_resultss.csv: overview of interventions performed for each of the variables as well as aggregated for causes and non-causal variables

Furthermore, you find summary plots and text files used to produce the latex tables in the paper's Appendix.