Regression-Based Machine Learning for Predicting Prior Convictions from Administrative Criminal Justice Records

Abstract

Predictive analytics is emerging in current police and court proceedings where data is used to analyse crime trends and court results; predictive, regression-based machine learning models use these data to predict dependent variables. This study explores the predictive use of regression-based machine learning models to predict counts of prior convictions based on structured data within a criminal justice dataset. The analysis is done on the Criminal Justice Dataset that is a publicly available data set with roughly 200,000 court cases data with demographic characteristics, criminal history factors, offense parameters and court decisions variables. Using descriptive statistical analysis estimates, correlation univariate tests and multiple regression predictive models the relationships are explored between the independent predictors and the dependent counts of prior convictions. Three model builds are used; the first is an OLS model to provide a comparison baseline followed by the use of a Ridge regression model with L2 regularization penalty and an ElasticNet model with L1 and L2 as a way of deploying multiple robust models to produce the best fit and prediction. We evaluate the performance of the model using standard regression metrics: coefficient of determination (R²), Mean Absolute Error (MAE), and Root Mean Squared Error (RMSE).