The COMPAS system, used in the US justice system to predict recidivism, represents one of the most controversial applications of machine learning. In 2016, a ProPublica article sparked a debate regarding its alleged racial bias. Due to the lack of a universal definition of bias and conflicting fairness criteria, there is still no consensus today on whether the system is biased.
The thesis combines a theoretical synthesis of the technical, legal, and social aspects of bias. I meticulously documented the attributes and analyzed the dataset published by the ProPublica authors, which has been widely used to study bias since 2016. In the empirical section, I experimentally confirmed the findings of previous studies that more complex models (decision trees, logistic regressino, decision rules). I also reproduced the error pattern where Black defendants have a higher false positive rate (FPR), while White defendants have a higher false negative rate (FNR). I identified two key mechanisms contributing to these disparities: 1) the distribution of training data (primarily the difference in the base recidivism rate, as well as the distribution of age and prior offenses between the groups), and 2) the binarization process of probability scores, which further amplifies the differences between the groups.
The attempt to predict race based on the remaining attributes shows that models do not reliably reconstruct race from the remaining data, which limits the argument about proxies. Furthermore, the final assessment of model bias depends on the selected fairness criterion. According to the equalized odds criterion, all analyzed models are biased, whereas under the predictive parity criterion, they are unbiased. The findings explain the contradictory conclusions of past research and emphasize the necessity of considering probability scores and the original distribution of training data when evaluating machine learning models.
|
