Robust Optimization for Fairness with Noisy Protected Groups

Code for experiments in paper Robust Optimization for Fairness with Noisy Protected Groups, NeurIPS 2020.

Prerequisites

Python 3, tensorflow 1.14.0, numpy, pandas

Data preprocessing

Adult dataset

The Adult dataset is public and available here. The function preprocess_data_adult indata.py is used to preprocess the dataset. The preprocessed Adult dataset adult_processed.csv is included.

Taiwan Credit dataset

The Taiwan Credit dataset is public and available here. The function preprocess_data_credit indata.py is used to preprocess the dataset. The function load_dataset_credit loads the dataset and binarizes the features. The preprocessed binarized Credit dataset credit_default_processed.csv is included.

Running the experiments

We provide general procedure to load data, import supplementary .py files and set global variables. Then we give instructions on running each individual algorithm.

Load data

Import: data.py, losses.py, optimization.py, model.py, utils.py, tensorflow, numpy

Run: df = data.load_dataset_adult() or df = data.load_dataset_credit()

Set variables:

LABEL_COLUMN = "label" (for Adult); "default" (for Credit)
FEATURE_NAMES = list(df.keys())
FEATURE_NAMES.remove(LABEL_COLUMN)
PROTECTED_COLUMNS = ['race_White', 'race_Black', 'race_Other_combined'] (for Adult); ['EDUCATION_grad', 'EDUCATION_uni', 'EDUCATION_hs_other'] (for Credit)

To set the variables for the protected groups and proxy groups:

For Oracle baseline algirithm without noise:

PROXY_COLUMNS = PROTECTED_COLUMNS 

For Adult/Credit dataset with a noise parameter:

PROXY_COLUMNS = data.get_proxy_column_names(PROTECTED_COLUMNS, noise_parameter)

Naive algorithm

Additional import:naive_training.py

Run the algrithm:

naive_training.get_results_for_learning_rates(df, FEATURE_NAMES, PROTECTED_COLUMNS, PROXY_COLUMNS, LABEL_COLUMN,constraint='tpr') with a list of learning rates. Please use constraint='tpr_and_fpr' for experiments on Credit dataset.

DRO algorithm

Additional import:dro_training.py

Run the algrithm:

dro_training.get_results_for_learning_rates(df, FEATURE_NAMES, PROTECTED_COLUMNS, PROXY_COLUMNS, LABEL_COLUMN,constraint='tpr') with a list of learning rates.

Softweights algorithm

Additional import:softweights_training.py

Run the algrithm:

softweights_training.get_results_for_learning_rates(df, FEATURE_NAMES, PROTECTED_COLUMNS, PROXY_COLUMNS, LABEL_COLUMN,constraint='tpr') with a list of learning rates.