Predicting Recidivism with a Transparent Ensemble Machine Learning Algorithm
JSHS · 2023
Overview
America cycles 9 million people through jails and 600,000 through prison annually. Within three years of release, 66% of released individuals are re-arrested. Judges and officers use statistical risk assessment tools to gauge a defendant’s odds of recidivating. However, such tools may contain bias. Two popular prediction algorithms, COMPAS and PATTERN, are criticized for over-predicting recidivism in Black and female defendants due to training with biased datasets. Misclassifications can produce widespread disparity. Defendants defined as medium/high risk are detained more often than low-risk defendants. Today, Black prisoners comprise 38% of incarcerated populations versus 13% of the general population, suggesting the concern is worth investigating. This study compares COMPAS and PATTERN algorithms to an ensemble machine learning model built from three supervised learning processes for binary classification: logistic regression, random forest, and neural networks. Through dimensionality reduction and sensitivity analyses, the ensemble was evaluated using 41 factors (e.g., adult/juvenile criminal history, criminal charges, age, gender, and race) from a 7000-person public registry dataset of past criminal offenses. Results suggest a 7-factor ensemble model predicts recidivism as accurately as a 137-factor COMPAS and 63-factor PATTERN model with more transparency and fairness. Prior adult criminal history was the biggest driver of recidivism predictions, and dynamic criminogenic factors had minimal impact. A lack of bias on race or gender suggests the ensemble could be a suitable replacement for COMPAS and PATTERN algorithms. The study concludes that recidivism risk tools improve the equity and efficacy of predictions with careful attention to accuracy, fairness, and transparency.
Competition history
- JSHS 2023
Resources
Related projects
ISEF · 2020
A Machine Learning Approach to Identify Socio-economic Factors Responsible for Patients Dropping Out of Substance Abuse Treatment
ISEF · 2017
A Novel Supervised Machine Learning Approach to Predicting Crime Information
ISEF · 2021
Generating Police Sketches Using a Generative Machine Learning Algorithm
ISEF · 2019
Predicting Opioid Use Disorder (OUD) Using Machine Learning
ISEF · 2025
Empathy Powered Pairing: A Trauma-Informed Pairing Algorithm for Traumatized Youth
ISEF · 2019
Predicting the Development of Secondary Central Nervous System Cancer through Ensemble Learning Methods
ISEF · 2023
Predicting Patients at Risk for Preventable Adverse Events Upon Discharge From the Hospital With an Acute Exacerbation of Congestive Heart Failure
ISEF · 2023
Predicting Large Wildfires Using Machine Learning Approach Towards Environmental Justice via Remote Sensing
Closest projects by meaning, across every fair and year in the corpus.