Deep learning on graphs for prediction of urban crimes

Abstract

The main goal of our PhD project is to develop a new crime forecasting methodology based in graph signal processing and deep learning by modeling São Paulo city in a natural street-level manner, based not only in past historical crime data, but also in other social and urban information that might enhance predictions. It is undeniable that crimes in Brazil have not only grown in recent years, but also become more violent and modernized. In contrast, agencies in charge of law and order control - police and criminal justice system - did not follow these trends and has aged. The gap between the dynamics of crime and violence and the state's ability to contain them within the rule of law has widened. Therefore, introducing modern instruments for the management of public order and crime containment is imperative to make public security policies more efficient. In this context, several works has been developed to map, analyse, identify, visualize and forecast crime data, being the last one our scope. Many works try to predict crime, however most of them rely on an inconvenient way of modeling a city, which we call regular grid approach. Despite simple to implement, this method has several limitations, mainly related to fine grain analysis and pattern identification. At the same time, there was a recent growth in the last years of researches about Graph Neural Network, which are models highly related to Graph Signal Processing theory and that rely on processing data embedded in a graph structure. Hence, these new methods broadens the possibilities to model a city more naturally in a street-level manner. (AU)