Aprendizado não supervisionado aplicado à análise de furtos de cabo de cobre em Belo Horizonte

Abstract

Traditional hotspot mapping methods, such as Kernel Density Estimation (KDE), fail to capture the temporal dynamics of criminal phenomena, thereby limiting the comprehension of their spatial evolution. This work proposes a dynamic spatio-temporal analysis methodology for the detection of cable theft hotspots in Belo Horizonte, for application at the Integrated Operations Center (COP-BH). The approach segments data into temporal windows, applying five clustering techniques (K-Means, K-Medoids, HAC, DBSCAN, and HDBSCAN) and evaluating their performance using internal metrics (Silhouette, Davies–Bouldin, Calinski–Harabasz, and Density-Based Clustering Validation) and the Predictive Accuracy Index (PAI). The results indicate that density-based methods, such as HDBSCAN, identify patterns of hotspot emergence, dissipation, and displacement more precisely than KDE. Furthermore, it was observed that the algorithms possess advantages and disadvantages relative to the requirements of COP-BH, highlighting the critical need for the appropriate application of each method. The proposed methodology contributes to the intelligent monitoring of critical infrastructure, providing support to COP-BH and establishing a solid foundation for future integrations with predictive and video analytics systems.