Recidivism prediction: A machine learning approach for a more efficient justice system

Abstract

Predicting recidivism is crucial for public safety and improving the justice system. This study introduces ImAUC-PSVM, an improved support vector machine that better handles skewed data and hyperparameter sensitivity by directly optimizing the AUC, enhancing prediction accuracy. That change helps simplify things, and also means you don’t have to get lost in tweaking a ton of hyperparameters to make it work. This makes it especially useful when your dataset has those annoying, uneven class distributions. The theoretical side of it shows that the ImAUC-PSVM manages to keep all the important traits of the original PSVM but is much better at handling fast, incremental updates—something that's pretty essential when you want to keep recidivism predictions accurate over time. On top of that, they enhanced the artificial bee colony algorithm—which, I admit, sounds a bit quirky, but it's a smart bio-inspired optimization technique. Their version uses cooperative learning, where two “agents” share info to adjust positions of potential “food sources” based on which one performs better, guided by a shared learning factor. It's like bees working together more efficiently, you know? They tested the model using data from correctional facilities in China, and it scored a solid 90.26% in predictive accuracy. That's not just a good number; it underlines how well the model deals with those skewed datasets, which is often a big stumbling block in recidivism prediction. Overall, this research suggests that the ImAUC-PSVM could seriously boost how we model predictions in criminal justice, and it might be adaptable to the kinds of tricky, imbalanced problems you find in other fields, too. Seems like a neat step forward, even if some of the details are a bit dense. Code is publicly available at https://github.com/ShouXin-Guo/Recidivism/tree/main.