Recidivism Prediction: A Novel Machine Learning-based Imbalanced Learning Method Combined with the Differential Equation Algorithm

We propose a new method to improve recidivism prediction by addressing hyperparameter tuning and class imbalance. Our model, the Imbalanced AUC-Proximal Support Vector Machine (ImAUC-PSVM), incorporates AUC optimization to enhance standard PSVM performance while reducing the need for extensive parameter tuning. The model opts for direct AUC optimization rather than sticking to the usual loss functions like hinge loss or cross-entropy. This choice kind of gives it a leg up in ranking performance and makes it more resilient when classes are imbalanced—a common pain point, really. While adapting to changing prediction settings, it retains core PSVM elements and uses the Differential Equation (DE) approach, which aids hyperparameter tuning in complex environments. The DE algorithm kind of fine-tunes performance more effectively than the usual grid or random search methods. It explores a wider range of possibilities and manages to steer clear of getting stuck in local optima. Because of that, you can be a bit more confident that it’ll actually find well-tuned hyperparameters in the end. We assess the model using data from Chinese correctional centers and North Carolina prison records. On Chinese and American samples, the model attained predictive scores of 88.125 and 90.809. Compared to the strongest baseline, the proposed method improved AUC by 7.1% on the Chinese and 6.7% on the American datasets. The results show strong performance on imbalanced data and confirm the method’s reliability for real-world recidivism prediction. It improves criminal justice modeling and provides a flexible framework for complex classification tasks in other domains.