Fairness in Classification: A Look at Bias in Recidivism Prediction Instruments
Alexandra Chouldechova – Carnegie Mellon University
Recidivism prediction instruments (RPI’s) provide decision makers with an assessment of the likelihood that a criminal defendant will reoffend at a future point in time. While such instruments are already widely used across the country, their use is also attracting tremendous controversy. Much of the controversy concerns potential discriminatory bias in the risk assessments that are produced. In this talk I will discuss several fairness criteria that have been applied to assess the fairness of recidivism prediction instruments. I will present a simple impossibility result showing that these criteria cannot all be simultaneously satisfied when recidivism prevalence differs across groups. In the latter part of the talk I will discuss how disparate impact might naturally arise if certain fairness criteria are not satisfied. To conclude, I will illustrate some limitations of looking at aggregate-level fairness metrics.