Offense-History-Based Optimization of COMPAS

Abstract

Recent advances in AI bring societal benefits but can also lead to concerns regarding biased results and inaccurate predictions. This study focused on prediction error within COMPAS―a prominent U.S. recidivism risk assessment tool―and seeks to improve both its False Positive Rate (FPR) and False Negative Rate (FNR). Crime type (felony vs. misdemeanor) is introduced as a grouping variable to improve predictive accuracy and practical applicability. Experimental results revealed that, beyond COMPAS’s high overall error rate, error patterns differed distinctly between these subgroups. Compared to the reproduced COMPAS baseline, statistically significant improvements in FPR and FNR were achieved. Moreover, a single optimization model tended to yield remarkable reductions in specific error rates, while a separate optimization model demonstrated a balanced improvement. These findings offer flexible strategies for reducing recidivism prediction errors and provide future model enhancements.