A decision tree approximating the COMPAS recidivism prediction tool, with >70% accuracy. The input is information about the criminal record of a convict, and the output is "HIGH RISK" or "LOW RISK" (COMPAS's determination of risk of re-arrest within 2 years). Each node displays: - the test at that node - E.g., In the root node: priors_count < 3.5 checks if the person has 3 or fewer priors - Following the left arrow means that the test was true; the right means it was false. - Gini impurity: ignore - number of samples in that node - E.g., In the root node: 8830 people were in the whole sample used to learn this decision tree. - E.g., The left child of the root: 6224 of them had 3 or fewer priors - value: [# LOW RISK, # HIGH RISK] - E.g., The left child of the root: 4176 people in the sample were labeled LOW RISK by COMPAS. - the color: a visualization of how skewed the samples are towards HIGH RISK (blue) or LOW RISK (orange) Variable names: - c_charge_degree: Whether the most recent conviction is a felony (<1.5) or misdemeanor (>1.5) - priors_count: number of non-juvenile priors - juv_fel_count: number of juvenile felonies - juv_mid_count: number of juvenile misdemeanors - juv_other_count: I have no idea