COMPAS Bias Audit — Full-Stack App

ASU Statistical Machine Learning Project

Quick Start

1. Backend (Flask API)

cd backend
pip install -r requirements.txt
python app.py
# Starts on http://localhost:5050

2. Frontend Open frontend/index.html in any browser.
Or serve it: python -m http.server 3000 then visit http://localhost:3000

API Endpoints

Endpoint	Description
GET /api/overview	KPIs — FPR gap, recidivism rate, totals
GET /api/race_metrics?threshold=5	Per-race FPR/FNR/TPR at any threshold
GET /api/score_distribution	Score 1-10 distribution per race
GET /api/threshold_sweep	FPR/FNR at all thresholds 1-10
GET /api/intersectional	Race × Gender subgroup analysis
GET /api/model	LR + DT accuracy/AUC + feature importances
GET /api/sensitivity?sens_thresh=70	Fair decision policy simulation
GET /api/eda	EDA — age, priors, gender breakdowns

Key Findings (from real data)

• AA FPR: 44.85% vs Caucasian FPR: 23.45% → 21.4pp gap
• Logistic Regression AUC: 73.79% vs COMPAS AUC: 70.22%
• Sensitivity policy (top 30% deferred): reduces FPR gap by ~5pp
• priors_count is #1 feature (56.9% importance in DT)