COMPAS Recidivism — Fairness Audit Dashboard

RandomForest model trained on ProPublica COMPAS data · N=7,214 defendants · Seed=42 · Test=20%

Accuracy
63.5%
Overall correctness
AUC-ROC
0.668
Discrimination ability
African-American FPR
42.2%
False positive rate
Caucasian FPR
27.4%
False positive rate
⚠ Fairness Alert — Racial Disparity Detected
African-American defendants have a 14.8pp higher false positive rate than Caucasian defendants (42.2% vs 27.4%). The model predicts higher recidivism risk for Black defendants who did NOT reoffend, at a substantially higher rate than for white defendants. This is the core finding from the ProPublica 2016 analysis.
Target Distribution
Decile Score Distribution by Race
False Positive Rate by Race (Key Fairness Metric)
Classification Report
Feature Importance (Random Forest)
Data: ProPublica COMPAS Dataset · Model: RandomForestClassifier (n_estimators=100, seed=42)
FPR = proportion of non-recidivists incorrectly predicted as recidivist · N=1,443 test set