Predicting Crime By Exploiting Supervised Learning On Heterogeneous Data

Crime analysis supports law-enforcing agencies in preventing and resolving crimes faster and efficiently by providing methods and techniques to understand criminal behavior patterns. Strategies for crime reduction rely on preventive actions, e.g., where perform street lighting and police patrol. The evaluation of these actions is paramount to establish new security strategies and to ensure its effectiveness. In this article, we propose a supervised learning approach that exploits heterogeneous criminal data sources, aiming to understand criminal behavior patterns and predicting crimes. Thus, we extract crime features from these data to predict the tendency of increase or decrease, and the number of occurrences of crimes types by geographic regions. To predict crimes, we exploit four learning techniques, as k-NN, SVM, Random Forest, and XGBoost. Experimental results show that the proposed approach achieves up to 89% of accuracy and 98% of precision for crime tendency, and up to 70% of accuracy and 79% of precision for crime occurrence. The results show that Random Forest and XGBoost usually perform better when trained with a short time window, while k-NN and SVM perform better with a longer time window. Moreover, the use of heterogeneous sources of data can be effectively used by supervised techniques to improve forecast performance.