Transparent and bias-resilient AI framework for recidivism prediction using deep learning and clustering techniques in criminal justice

Abstract

This paper presents the Recidivism Clustering Network (RCN), an effective approach for predicting repeat offenses using deep learning (DL), clustering, and explainable AI (XAI). The RCN improves offender profiling for more accurate and interpretable recidivism predictions, aligning with key legal principles like fair sentencing, transparency, and non-discrimination. The RCN employs machine learning (ML) models optimized with a Keras tuner, using the Synthetic Minority Over-sampling Technique (SMOTE) to handle class imbalance. With about 75% accuracy, the model shows strong recall, identifying 10,661 recidivists but producing 4,038 false positives—indicating a trade-off between sensitivity and specificity. Beyond predictions, RCN integrates clustering methods, including k-means, principal component analysis (PCA), and t-distributed Stochastic Neighbor Embedding (t-SNE), to identify hidden patterns within offender data. Visualizations reveal distinct clusters, linking characteristics, such as age, to recidivism behaviors. SHapley Additive exPlanations (SHAP) values enhance interpretability, showing that factors like time since the last conviction and age significantly impact predictions. The RCN approach offers substantial potential for criminal justice applications by combining predictive power with actionable insights, supporting a more ethical and accountable use of ML in offender profiling and aiding in fairer recidivism prevention strategies. The code and data are publicly available on GitHub at https://github.com/cavusmuhammed68/Recidivism-Clustering-Network-RCN-.