Hybrid ensemble model for automated assessment of delinquency levels in adolescents via random subspace learning

Timely intervention and behavioral correction are contingent upon the early detection of juvenile delinquency. With the use of several decision-based classifiers and the random subspace technique, this study suggests a strong ensemble learning framework for automated assessment of adolescent delinquency. Based on a delinquency dataset, we assess the ensemble framework’s predictive performance in four experimental setups: (i) various base learners in the random subspace ensemble; (ii) Logistic Model Tree (LMT) under various train-test splits; (iii) conventional single classifiers; and (iv) comparison with other ensemble methods such as AdaBoost, Bagging, and Logit Boost. The results demonstrate that the random subspace technique significantly enhances classification performance when paired with more potent base learners. LMT continuously outperformed other classifiers, with a peak accuracy of 96.77% and an AUC of 1.0 under a 75−25 train-test split. Despite their higher performance, more traditional ensemble techniques like AdaBoost and bagging were outperformed by the proposed random subspace + LMT architecture for multi-class classification of delinquency levels in adolescents. While Random Forest also performed well, simple models like Decision Stump and Random Tree showed instability and poor generalization. The study generally supports the use of ensemble subspace approaches with improved base classifiers for successful multi-class risk categorization in adolescent delinquency behavior analysis.