Enhancing Accuracy and Explainability of Recidivism Prediction Models

Abstract

Predicting recidivism is a challenging task, but it helps support courts in their decision-making process. Automated prediction models suffer from low accuracy and are associated with criticism for biased and unexplainable decision-making. In this poster, we present different machine-learning models with just a few selected features that achieve accuracies as good as models that use larger sets of features. In addition, we investigate the influencing features that contribute to recidivism prediction, which can enhance the explainability of the learned models.