Fairness and bias of the COMPAS algorithm compared to human assessments
The Data Mining Group of Aalto University is collaborating with researchers from the Indian Institute of Science, Polish Academy of Sciences and the University of Turku to explore algorithmic and human fairness and bias in decision making. In particular, they focus on racial bias in the predictions of the Correctional Offender Management Profiling for Alternative Sanctions (COMPAS) algorithm, a criminal risk assessment tool used in sentencing in a number of states in the U.S. The project, titled “Algorithms, fairness, and race: Comparing human recidivism risk assessment with the COMPAS algorithm”, was initiated at the Helsinki Summer Institute in Computational Social Science 2018.
The project consists of two parts. In the first part the team explored a wide set of fairness metrics and demonstrated how data preprocessing performed in previous (published) studies affects the fairness assessment of algorithmic recommendations. In the second part they collected their own data to evaluate human fairness and bias, which they then compared with the COMPAS assessment. In doing so, they drew from sociological concepts of in-group bias, social status, and stereotyping to formulate hypotheses about the patterns of associations between the respondent’s race, the race of the evaluated defendant, and the recidivism predictions.
The second part was based on a vignette survey run on TurkPrime. The respondents were presented with short descriptions of defendants and asked to predict recidivism risk. The survey was designed to have a balanced composition of white and black respondents, who were presented with vignettes of white and black defendants. Defendant descriptions came from real data on pre-trial defendants, which makes it possible to compare their COMPAS scores, real recidivism data, and risk assessment by survey respondents.
The first results were presented as a poster at the European Symposium Series on Societal Challenges in Computational Social Science held in Cologne, Germany on December 5th-7th, 2018. Preliminary analyses show that respondents are more lenient towards the offenders of their own race, but the differences are not large. However, if we exclude the defendants with medium-risk COMPAS scores and cases with high disagreement among the respondents (the majority is supported by less than 3/4 of the respondents of the same race), then race does not play a role and prediction rates agree. This suggests that decisions in ambiguous medium-risk cases are most susceptible to bias. Currently the team is working on finalizing the analysis and writing up a paper.