Nompil/Bias-Audit-Report-SA

Bias audit for employment prediction in South Africa, simulating IBM AI Fairness 360. # Bias Audit Report: Employment Prediction in South Africa ## Project Overview This repository, hosted by the Masikhule organization, presents a bias audit of a synthetic employment prediction dataset (income > R50,000/year), simulating IBM AI Fairness 360. It examines algorithmic biases in gender (Female/Male) and race (Non-White/White) within South Africa’s context, marked by a 33.2% unemployment rate [ 5 ], a Gini coefficient of 0.63 [ 0 ], and historical inequalities from apartheid. Aligned with the Employment Equity Act and BEE goals [ 6 ], the audit evaluates fairness metrics (Disparate Impact, Equal Opportunity Difference, Equalized Odds) and applies mitigations (preprocessing, reweighing) to promote equitable AI-driven hiring. ## Team Members - Nompilo Mchunu - Siyabonga Mdaweni - Godlove Nyathi - Neliswa Mbele - Kabelo Mushi - Adivhaho Ndou ## Objectives - Identify biases in gender and race using statistical tests (chi-squared, p 12 , 13 ]. ## Methodology Using a synthetic dataset (80% non-White [ 0 ]), we trained a logistic regression model to predict employment outcomes. Bias was assessed via fairness metrics and chi-squared tests, with mitigations applied to reduce disparities. Visualizations (e.g., accuracy vs. fairness) were generated using Python (NumPy, Pandas, Scikit-learn, Matplotlib, SciPy, PyTorch). ## Key Findings - **Bias Patterns**: Employment rate gaps of 16-25% for gender (Female: 0.557, Male: 0.715) and race (Non-White: 0.594, White: 0.845), p 12 ]. - **Mitigations**: Improved Disparate Impact (0.696-0.788 to 0.989-1.043), with accuracy trade-offs (67.75% to 63.50-65.00%) [ 9 ]. - **SA Context**: Addresses unemployment (33.2% [ 5 ]), Gini (0.63 [ 0 ]), and apartheid legacies [ 6 ]. ## Deliverables - **View Notebook (Bias_Audit_Report_SA.ipynb)**: Jupyter notebook with code, analysis, and visualizations. View via nbviewer or run locally (see Usage). - **Download Notebook PDF (Bias_Audit_Report_SA.pdf)**: PDF export of the notebook. - …