Pragmatic Fairness: Developing Policies with Outcome Disparity Control

Steps to reproduce results from the submission.

0. Set up Environment (optional)
   • conda env create -f EqB_environment.yml (or ModBrk_envrionment.yml)
   • conda activate EqB_environment (or ModBrk_envrionment)
1. Paste experiment specific configs into training_configs.yml
   • NYCSchools_EqB_configs.yml (for EqB const)
   • NYCSchools_ModBrk_configs.yml (for ModBrk const)
2. Run python main.py

We will see results saved in out/ folder. We also include the results we obtained by running the code, and how we plot them in manuscript, in the plot/ folder.

Code to reproduce results from the submission.

data_preproc

• NYCSchools_dataset_EqB_clean.ipynb - contains data preprocessing needed for EqB constraint
• NYCSchools_dataset_ModBrk_clean.ipynb - contains data preprocessing needed for ModBrk constraint, NYC dataset
• IHDP_dataset_ModBrk_clean.ipynb - contains data preprocessing needed for EqB constraint, IHDP dataset

scripts

• main.py - includes main function, where the entire pipeline is run from.
• run_model.py - contains the main routine for the training of the policy models as well as record results
• configs
  • training_config.yml - should paste from experiment specific configs and run with python main.py
  • NYCSchools_EqB_configs.yml
  • NYCSchools_ModBrk_configs.yml
• data - contains processed data from according to data_preproc/ notebooks
• data_utils
  • NYCschools.py or IHDP.py - load preprocessed data to torch-ready formats
• utils
  • experiments_utils.py - consists most helper functions for model training
  • train_test_utils.py - contains the core train and test functions for a single epoch
• nets
  • NNs.py - contains class of NN that makes up our pretrained MLPs and policy models