Multimodal prediction of aggressive behavior occurrence using a decision-level approach

Abstract

Traditionally, aggressive behavior incidents have not been considered as a serious crime, but in some contexts such as Bogotá city, this type of behavior caused 70% of the reported personal injuries and homicides in 2017–2018. This phenomenon is a concern for modern cities decision-makers who require predictive models to mitigate aggressive behavior occurrence. There are different source data that can be used to model and predict aggressive behavior, for instance, legal complaints, police penalties and emergency call datasets. In this paper, we propose a decision-level data fusion to combine the prediction of the different aggressive behavior sensors and improve the model predictive capacity. Results suggest that decision-level data fusion using average and max operators improves hotspots hit rates but leads to higher mean squared errors between predicted and real events maps. A texture feature analysis over the predicted maps also revealed that maps generated using the decision-level approach have relatively high entropy, and lower energy and homogeneity values.