Data-Driven Evaluation of Community-Based Youth Development Programs for Reducing Substance Use, Violence Risk, and Social Disengagement in Vulnerable Populations

Community-based youth development programs are widely used to address substance use, exposure to violence, antisocial behaviour, and social disengagement among vulnerable young populations, yet conventional program evaluations often rely on isolated outcome indicators and statistical models that inadequately capture the interaction among individual, family, peer, program, and community-level risk factors. This study develops a data-driven evaluation framework centred on a novel Youth Development Risk and Engagement Fusion (YD-REF) algorithm for assessing program effectiveness and predicting multidimensional youth outcomes. The proposed model integrates demographic attributes, program participation intensity, attendance consistency, mentoring exposure, substance-use indicators, violence-related risk factors, school or employment engagement, peer-network characteristics, family support, and community vulnerability measures. YD-REF combines multi-task learning, gradient-boosted feature extraction, temporal attention, and graph-based relational learning to simultaneously estimate substance-use risk, violence risk, social-disengagement probability, and overall program-response scores. Its performance is compared with Logistic Regression, Support Vector Machine, Random Forest, XGBoost, and a conventional Artificial Neural Network using stratified cross-validation and evaluation measures including accuracy, precision, recall, F1-score, area under the receiver operating characteristic curve, area under the precision-recall curve, Brier score, and calibration error. Comparative performance is examined using ROC curves, precision-recall curves, calibration plots, feature-importance graphs, subgroup performance charts, confusion matrices, and longitudinal risk-transition graphs. The framework further incorporates SHAP-based explainability to identify the program characteristics and social determinants most strongly associated with positive youth outcomes. The proposed approach is designed to provide superior predictive discrimination, better calibration, improved identification of high-risk participants, and more interpretable evidence of intervention effectiveness than conventional single-outcome models. The study establishes a technical foundation for using machine learning and explainable analytics to support evidence-based resource allocation, targeted intervention design, early-risk detection, and continuous evaluation of community-based youth development programs.