prej_remover

Dataset Setup

Currently, for Federated Prejudice Removal we support following datasets:

• Adult
• Compas

Running Federated Prejudice Removal in FL

This example explains how to run Federated Prejudice Removal, a federated learning implementation of the Kamishima Algorithm. We use this algorithm in our paper on bias mitigation and federated learning; see it herehttps://arxiv.org/abs/2012.02447).

This example requires the gensim library. Run the following: pip install gensim

The following preprocessing was performed in AdultPRDataHandler on the original dataset:

• Drop following features: workclass, fnlwgt, education, marital-status, occupation, relationship, capital-gain, capital-loss, hours-per-week
• Map race, sex and class values to 0/1
• Split age and education columns into multiple columns based on value
• Map native-country values to an integer 1-7 based on continent

Further details in documentation of preprocess() in AdultPRDataHandler.

The following preprocessing was performed in CompasPRDataHandler on the original dataset:

• Map sex values to 0/1 based on underprivileged/privileged groups
• Filter out rows with values outside of specific ranges for days_b_screening_arrest, is_recid, c_charge_degree, score_text, race
• Quantify length_of_stay from c_jail_out and c_jail_in
• Quanitfy priors-count, length_age_cat, score_text, age_cat
• Drop following features: id, name, first, last, compas_screening_date, dob, age, juv_fel_count, decile_score, juv_misd_count, juv_other_count, priors_count, days_b_screening_arrest, c_jail_in, c_jail_out, c_case_number, c_offense_date, c_arrest_date, c_days_from_compas, c_charge_desc, is_recid, r_case_number, r_charge_degree, r_days_from_arrest, r_offense_date, r_charge_desc, r_jail_in, r_jail_out, violent_recid, is_violent_recid, vr_case_number, vr_charge_degree, vr_offense_date, vr_charge_desc, type_of_assessment, decile_score.1, screening_date, v_type_of_assessment, v_decile_score, v_score_text, v_screening_date, in_custody, out_custody, priors_count.1, start, end, event
• Split age-cat, priors_count and c_charge_degree columns into multiple columns based on value

Further details in documentation of preprocess() in CompasPRDataHandler.

No other preprocessing is performed.

• Set FL environment by running:

  ./setup_environment.sh <env_name>  # setup environment
  

• Activate a new FL environment by running:

  conda activate <env_name>          # activate environment
  

• Split data by running:

  python examples/generate_data.py -n <num_parties> -d <dataset_name> -pp <points_per_party>
  

• Generate config files by running:

  python examples/generate_configs.py -n <num_parties> -f prej_remover -d <dataset_name> -p <path>
  

• In a terminal running an activated FL environment, start the aggregator by running:

  python -m ibmfl.aggregator.aggregator <agg_config>
  

  Type START and press enter to start accepting connections

• In a terminal running an activated FL environment, start each party by running:

  python -m ibmfl.party.party <party_config>
  

  Type START and press enter to start accepting connections.

  Type REGISTER and press enter to register the party with the aggregator.

• Finally, start training by entering TRAIN in the aggregator terminal.