Adversarial Debiasing for Model Fairness

This project implements adversarial debiasing to help increase model fairness in the DiCE (Diverse Counterfactual Explanations) framework.

Overview

Adversarial debiasing is a technique where a model is trained to be both accurate in its predictions and fair with respect to protected attributes (e.g., gender, race) by using an adversarial approach. The implementation includes:

1. A standard model (baseline)
2. An adversarially debiased model
3. Evaluation of fairness metrics
4. Analysis of model performance
5. Assessment of recourse feasibility

Files

• adversarial_debiasing.py: Complete implementation of adversarial debiasing with all analyses
• train_adversarial_debiasing.py: Script to train a model with adversarial debiasing
• evaluate_fairness.py: Script to evaluate fairness metrics
• compare_performance.py: Script to compare model performance
• analyze_recourse.py: Script to analyze recourse feasibility
• run_all.py: Script to run all analyses in sequence

Requirements

• Python 3.7+
• TensorFlow 2.x
• NumPy
• Pandas
• Matplotlib
• Seaborn
• Scikit-learn
• DiCE

Usage

You can run the complete analysis with:

python adversarial_debiasing.py

Or run individual scripts:

python train_adversarial_debiasing.py
python evaluate_fairness.py
python compare_performance.py
python analyze_recourse.py

Or run all scripts in sequence:

python run_all.py

Metrics Evaluated

Fairness Metrics

• Demographic Parity Difference
• Equal Opportunity Difference
• Disparate Impact Ratio
• Equalized Odds Difference

Performance Metrics

• Accuracy
• Precision
• Recall
• F1 Score
• ROC AUC

Recourse Feasibility Metrics

• Success Rate
• Average Number of Changes
• Average Distance

Results

The analysis generates several plots in the plots directory:

• fairness_metrics_comparison.png: Comparison of fairness metrics
• performance_metrics_comparison.png: Comparison of model performance
• accuracy_by_gender.png: Analysis of accuracy by gender
• recourse_metrics_comparison.png: Comparison of recourse feasibility metrics

Datasets

The implementation uses the UCI Adult Income dataset, which predicts whether income exceeds $50K/yr based on census data, with gender as the protected attribute.