Model Details
Overview
COMPAS (Correctional Offender Management Profiling for Alternative Sanctions) is a public dataset, which contains approximately 18,000 criminal cases from Broward County, Florida between January, 2013 and December, 2014. The data contains information about 11,000 unique defendants, including criminal history demographics, and a risk score intended to represent the defendant’s likelihood of reoffending (recidivism). A machine learning model trained on this data has been used by judges and parole officers to determine whether or not to set bail and whether or not to grant parole.
In 2016, an article published in ProPublica found that the COMPAS model was incorrectly predicting that African-American defendants would recidivate at much higher rates than their white counterparts while Caucasian would not recidivate at a much higher rate. For Caucasian defendants, the model made mistakes in the opposite direction, making incorrect predictions that they wouldn’t commit another crime. The authors went on to show that these biases were likely due to an uneven distribution in the data between African-Americans and Caucasian defendants. Specifically, the ground truth label of a negative example (a defendant would not commit another crime) and a positive example (defendant would commit another crime) were disproportionate between the two races. Since 2016, the COMPAS dataset has appeared frequently in the ML fairness literature 1, 2, 3, with researchers using it to demonstrate techniques for identifying and remediating fairness concerns. This tutorial from the FAT* 2018 conference illustrates how COMPAS can dramatically impact a defendant’s prospects in the real world.
It is important to note that developing a machine learning model to predict pre-trial detention has a number of important ethical considerations. You can learn more about these issues in the Partnership on AI “Report on Algorithmic Risk Assessment Tools in the U.S. Criminal Justice System. The Partnership on AI is a multi-stakeholder organization -- of which Google is a member -- that creates guidelines around AI.
Model Performace
Owners
Intel XAI Team, xai@intel.com
References
- Wadsworth, C., Vera, F., Piech, C. (2017). Achieving Fairness Through Adversarial Learning: an Application to Recidivism Prediction. https://arxiv.org/abs/1807.00199.
- Chouldechova, A., G'Sell, M., (2017). Fairer and more accurate, but for whom? https://arxiv.org/abs/1707.00046.
- Berk et al., (2017), Fairness in Criminal Justice Risk Assessments: The State of the Art, https://arxiv.org/abs/1703.09207.