Crime Forecasting with Optimized Ensemble Machine Learning

Over the past few years, an increase in urban crime has been a major challenge for law enforcement agencies across the globe. Crime pattern prediction and analysis have an important role to play in preventing crime and the proper allocation of resources. This research work discusses a machine learning methodology for crime pattern identification using a combination of classifiers, such as XGBoost, Random Forest, Gradient Boosting, and HistGradientBoosting. The model is trained and validated on the publicly released Chicago Crime dataset, with rigorous preprocessing and feature engineering steps done in an attempt to enhance prediction accuracy. Hyperparameter tuning was carried out with the Optuna framework in order to attain the best performance for each model independently. The outputs show that the ensemble model performs better than the singular classifiers, with a final accuracy of 84%, as an illustration of the power of soft voting processes in ensemble learning. The study not only points out the strength in using an ensemble of different algorithms but also offers a scalable approach towards real-world crime forecasting systems. (as supported by studies such as [6], [15], and [16]).