Comparative Analysis of Neural Networks with Traditional ML Algorithms for Crime Prediction

Around the world, the number of crimes has sharply increased over time. The use of systems that can recognize and anticipate patterns of crime has become essential due to the rise in crime rates worldwide. This would help the authorities to better allocate resources and prevent such incidents, thereby helping to drop crime rates. This research aims to perform a comparison of a few traditional machine learning (ML) models with neural networks (NN) in crime prediction. This is done by taking a relevant set of data as input and predicting whether or not an arrest will be made, which allows one to assess the severity of the crime. The dataset, “Chicago Crime data,” was selected, a number of feature selection and pre-processing approaches are used in order to get the best dataset for training. K-Nearest Neighbor (KNN), Adaptive Boosting (AdaBoost), Extreme Gradient Boosting (XGBoost), and Neural Networks are the four models selected for comparison in the proposed work. The obtained results include the training time needed for each model, along with its accuracy, precision, recall, and F1 score. As recall is the ratio of the number of arrest cases correctly predicted by the model to the total number of arrest cases, it is given the top consideration in the proposed work. The F1 scores of the models are analyzed to identify the best model.