Stacking Sentinel: An Ai-Based Model for Child Abuse Prediction

Abstract

Despite advancements in crime prediction, accurately forecasting specific types of crime at a granular level remains a challenge. This study addresses this gap by developing a model for predicting various types of child abuse in Chicago. Through exploratory analysis methods, the dataset was examined to understand crime trends, followed by balancing using the SMOTEENN technique and feature selection with Recursive Feature Elimination (RFE). A stacking classifier was then employed to ensemble LGBM, Random Forest, and Extratrees as base classifiers, with Particle Swarm Optimized - Artificial Neural Network (PSO-ANN) served as the meta-classifier. This approach resulted in groundbreaking performance for predicting various child abuse types, achieving 97% accuracy and 96% precision, recall, and F1-score. Furthermore, the model exhibited significantly lower execution time (307.69 milliseconds) compared to the existing BS-SC Model (476.63 milliseconds). These findings showcase the effectiveness of machine learning in leveraging demographic and spatial-temporal data to predict child abuse subtypes with high accuracy and computational efficiency. The proposed Stacking Sentinel (SS) model holds tremendous potential for real-time prediction and prevention of child abuse, as well as contributing to enhanced child welfare initiatives.

Keywords: Child Sexual Abuse Prediction, Sexual Offending, ANN, Optimization, Stacking Classifier